doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.pgen.1007211 | Th-POK regulates mammary gland lactation through mTOR-SREBP pathway | The Th-inducing POK (Th-POK, also known as ZBTB7B or cKrox) transcription factor is a key regulator of lineage commitment of immature T cell precursors. It is yet unclear the physiological functions of Th-POK besides helper T cell differentiation. Here we show that Th-POK is restrictedly expressed in the luminal epithelial cells in the mammary glands that is upregulated at late pregnancy and lactation. Lineage restrictedly expressed Th-POK exerts distinct biological functions in the mammary epithelial cells and T cells in a tissue-specific manner. Th-POK is not required for mammary epithelial cell fate determination. Mammary gland morphogenesis in puberty and alveologenesis in pregnancy are phenotypically normal in the Th-POK-deficient mice. However, Th-POK-deficient mice are defective in triggering the onset of lactation upon parturition with large cellular lipid droplets retained within alveolar epithelial cells. As a result, Th-POK knockout mice are unable to efficiently secret milk lipid and to nurse the offspring. Such defect is mainly attributed to the malfunctioned mammary epithelial cells, but not the tissue microenvironment in the Th-POK deficient mice. Th-POK directly regulates expression of insulin receptor substrate-1 (IRS-1) and insulin-induced Akt-mTOR-SREBP signaling. Th-POK deficiency compromises IRS-1 expression and Akt-mTOR-SREBP signaling in the lactating mammary glands. Conversely, insulin induces Th-POK expression. Thus, Th-POK functions as an important feed-forward regulator of insulin signaling in mammary gland lactation.
| Th-POK, aka cKrox or ZBTB7B, is a zinc finger transcription factor that specifies the cell fate of immature T cell precursors towards the CD4 lineage. It is yet largely unknown if Th-POK participates in the regulation of other biological processes. It is also not known if Th-POK functions beyond cell fate determination. In this study, we found that Th-POK is expressed in the luminal, but not the basal epithelial cells in the mammary glands. Despite the lineage restricted expression in the mammary glands, Th-POK is dispensable for mammary epithelial cell lineage specification and mammary gland development. Rather, Th-POK regulates the onset of lactation and milk lipid production via mammary epithelial cell autonomous mechanisms, suggesting that Th-POK exerts distinct functions in a tissue specific manner. Th-POK regulates the expression of insulin receptor sunstrate-1 and insulin-induced mTOR-SREBP pathway activation and lipid biosynthesis in the mammary alveolar cells at lactation. Thus, Th-POK functions as an important metabolic regulator in the lactating mammary glands, in addition to its well documented cell fate specification functions in T cell development.
| Transcription factor Th-POK (Th-inducing POK, also known as ZBTB7B or cKrox) is a key regulator of lineage commitment of immature T cell precursors [1–3]. Loss of Th-POK expression or impaired Th-POK function disrupts the development of CD4 T cells [4], whereas enforced expression of Th-POK results in lineage conversion to the CD4 fate [4–6]. Th-POK is positively regulated by GATA-3, a transcription factor essential for CD4 fate determination [6, 7], and negatively regulated by CD8 lineage-specific transcription factor Runx3 [8]. Th-POK expression is also regulated post-transcriptionally by p300-mediated acetylation [9]. Th-POK acts as both transcriptional repressor and activator to regulate gene expression [1, 10, 11]. Cross-antagonism between Th-POK and Runx3 are determinative to CD4 versus CD8 cell fate decision. It is yet largely unknown the physiological functions of Th-POK besides helper T cell differentiation.
Mammary glands develop from a rudimentary tree to a branched epithelial network of ducts during puberty that further undergo alveologenesis to create a lactation-competent gland upon pregnancy [12–14]. Morphogenesis and differentiation of the mammary gland to a functional milk-producing organ are precisely regulated [12–14]. GATA-3, the critical transcription factor in T cell CD4 fate determination upstream of Th-POK, is the most highly enriched transcription factor in the mammary epithelium of pubertal mice [15]. GATA-3 is critical to the maintenance of proper luminal progenitor pool at puberty and alveolar differentiation during pregnancy [15, 16].
Upon parturition, milk protein and lipid biosynthesis capability in the alveolar epithelial cells are sharply increased [12, 14, 17]. The key event in the onset of milk secretion is the release of cytosolic lipid droplets (CLDs) by the alveolar epithelial cells into the luminal space [12, 14, 17]. Reduced milk lipid production and secretion in lactating mammary glands have been linked to poor newborn survival [18–25]. Interaction between xanthine oxidoreductase (XOR) and butyrophilin (BTN) is essential to the secretion of CLDs [21, 23, 26]. Proteins regulating XOR expression, e.g. Cidea [24], are important regulators of the CLD secretion. Lipid biosynthesis are tightly regulated at transition from pregnancy to lactation [17, 27]. Insulin signaling through its receptor (IR) is a potent regulator of cellular metabolism. Deletion of IR in the mammary epithelial cells at pregnancy resulted in reduced milk protein and lipid production [28]. Akt1 downstream of IR is required for the upregulated lipid synthesis in the lactating mammary glands. Akt1 deficiency resulted in impaired lipid biosynthesis in the lactating mammary gland [19]. Ectopic expression of constitutively active Akt in the mammary glands led to excess lipid synthesis during pregnancy and lactation [29]. Prolactin is demonstrated to regulate a subset of mammary epithelial cell specific lipogenic gene expression in lactating mammary glands [30]. Mice deficient of Src, a regulator of the prolactin signaling, or its binding protein actin filament-associated protein 1 (AFAP1) exhibited retention of large CLDs [20, 25]. Sterol regulatory element binding proteins (SREBPs), the key transcription factors regulating fatty acid and cholesterol biosynthesis [31–33], are proposed as key regulators in efficient lipid synthesis at the onset of milk secretion [17, 20, 27, 34]. Akt-mTOR signaling is required for insulin-induced SREBP1 expression and processing [35–37]. Akt regulates lipid biosynthesis via SREBP1 [37]. AFAP1 deficiency reduced SREBP1 levels in the lactating mammary glands [20].
In this study, we identified Th-POK as a luminal-specific expressing transcription factor in the mammary glands. Knockout of Th-POK did not affect pubertal growth or alveologenesis of mammary gland. However, the transition from pregnancy to lactation upon parturition was compromised in the Th-POK knockout mice. Th-POK deficiency impaired insulin-induced activation of mTOR-SREBP pathway and lipid biosynthesis. As a result, Th-POK knockout mice were unable to efficiently secret milk and to nurse the offspring. Thus, Th-POK functions as a critical regulator in mammary gland lactation.
In an attempt to breed with Th-POK knockout (KO) mice, wild-type (WT) females were mated with KO males and KO females were mated with WT males. All (30/30) KO males were competent in producing healthy offspring (Fig 1A). However, significantly decreased survival of pups born to KO females was observed (Fig 1A). This was not due to the pup genotype, as all the pups were heterozygous for Th-POK. To investigate if the observed defect in pup viability born to KO mice was dependent on the genotype of the mother, pups born to WT females were fostered to KO mice. ~50% of the pups died during the course of study (Fig 1B). The surviving pups exhibited reduced growth, compared to the pups nursed by WT mothers (Fig 1C). Thus, Th-POK is required for lactating mice to support the survival and growth of their litters.
GATA-3, a transcription factor upstream of Th-POK in T cell development, is the most highly enriched transcription factor in the mammary epithelium of pubertal mice and a critical regulator of luminal differentiation [15, 16]. The inability of KO mice to properly nurse their pups promoted us to study if Th-POK is expressed in the mammary gland and plays a role in mammary gland development and function. Immunohistochemical staining on mammary gland sections showed that Th-POK was expressed in mammary epithelial cells of virgin mice (Fig 1D). Western blot analysis further confirmed that Th-POK protein was expressed in the mammary epithelial cells isolated from the mammary glands of virgin mice (Fig 1E). The mammary gland is composed of basal layer myoepithelial cells and inner layer luminal cells [13, 38, 39]. Th-POK colocalized with luminal marker cytokeratin 8 (K8), but not basal marker α-smooth muscle actin (αSMA) (Fig 1F). Th-POK mRNA levels were significantly higher in the K8-positive luminal cells than in the K14-positive basal cells (Fig 1G). Thus, Th-POK is expressed restrictedly in the luminal lineage. At lactation, Th-POK was expressed in the luminal epithelial cells of alveoli (Fig 1H–1J). Analysis of Th-POK expression at different mammary developmental stages revealed that its expression levels were upregulated at late pregnancy (day 17.5) and remained high at the lactation stage (Fig 1K and S1 Fig). Analyses of Th-POK expression in the isolated mammary epithelial cells further revealed increased Th-POK mRNA and protein levels at late pregnancy and lactation (Fig 1L and 1M).
As Th-POK is specifically expressed in luminal epithelial cells, we next examined if Th-POK deficiency would affect mammary gland development in a manner similar to GATA-3. As shown by whole-mount analyses, comparable ductal outgrowth and numbers of terminal end buds were noticed in WT and KO mammary glands from virgin mice (S2A–S2C Fig). Histology inspection showed indistinguishable architecture of terminal end buds at 5 weeks and mammary ducts at 7 and 10 weeks between WT and KO mice (S2D Fig). Mammary epithelial compartment undergoes lobuloalveolar development in pregnancy to form the alveolar secretory units [13, 38, 39]. Mammary glands from the KO mice at early-, mid- to late-term pregnant mice (postcoital days 5.5, 12.5 and 17.5, respectively) showed normal organization of lobuloalveolar structures (S2E and S2F Fig, Fig 2A and 2B). Expansion of the mammary epithelium is achieved by active cell proliferation. BrdU incorporation analyses revealed comparable rates of cell proliferation in the mammary glands from WT and KO mice (S2G and S2H Fig). The mammary epithelial cells differentiate into functional alveolar cells during pregnancy. Such secretory differentiation process is characterized by upregulation of milk protein expression [12]. mRNA levels of milk proteins were indistinguishable between WT and KO mammary glands (S3A Fig). Thus, Th-POK deficiency does not affect mammary epithelial cell fate determination and mammary gland development at puberty and pregnancy.
The increased mortality and growth retardation in pups nursed by the KO mice suggested that the KO mice may be defective in milk production at lactation stage. Indeed, lactating KO mice had 2.2-fold reduction in oxytocin-stimulated milk secretion compared to WT mice (Fig 2C). Protein and lipid are the major components of milk [12, 14]. Milk protein concentration and composition from the lactating KO mice were largely similar to that from the WT mice (S3B and S3C Fig). This promoted us to investigate whether lipid production and secretion were compromised in the lactating KO mice. The concentration of triacylglycerol (TAG), the main component of the milk lipid [12, 17], was markedly lower in milk taken from lactating KO mice compared with WT mice (Fig 2D). Milk lipids are secreted as milk fat globules (MFGs), with lipid droplets wrapped by a plasma membrane bilayer. Although the numbers of MFGs were similar in the milk from lactating WT and KO mice (Fig 2E and 2F), MFGs from the lactating KO mice were substantially smaller in size compared to those from lactating WT mice (Fig 2E and 2G). At parturition, release of CLDs from the alveolar epithelial cells is a critical event in producing sufficient milk lipid [12, 14, 17]. At late-pregnancy (P17.5), large CLDs accumulated in the alveolar epithelial cells in both WT and KO mammary glands (Fig 2B). Following parturition, large CLDs were replaced by small lipid droplets in alveolar epithelial cells in the WT mammary glands at lactation day 2 (L2) (Fig 2B). In contrast to that in the WT mammary glands, luminal space was much less expanded in the KO mammary glands (Fig 2B and 2H). Large CLDs were retained within alveolar epithelial cells in the KO mammary glands after parturition (Fig 2B), suggesting impaired milk lipid secretion in the KO mice. Cytosolic TAG concentration was significantly lower, whereas nonesterified fatty acid (NEFA) concentration was significantly higher in KO mammary epithelial cells than in WT mammary epithelial cells at L2 (Fig 2I). Expression of genes invloved in lipolysis and TAG synthesis was substantially altered in the KO mammary epithelial cells at L2 (Fig 2J). To further confirm the retention of large CLDs in the KO alveolar epithelial cells and failure in lipid secretion, sections of mammary glands were stained for Perilipin2 (Plin2, also known as adipophilin), a lipid droplet surface protein present in the CLDs in the alveolar epithelial cells [40] (Fig 2K). At late-pregnancy (P17.5), large CLDs were present in alveolar epithelial cells in both WT and KO mammary glands with similar size (Fig 2K and 2L). On lactation day 2, only small size CLDs were present at the apical surface of the WT alveolar epithelial cells, whereas large amount of large size CLDs were apparent in the KO alveolar epithelial cells (Fig 2K and 2L).
Upon weaning of the pups, the mammary gland undergoes involution, a process characterized by the apoptosis of the mammary epithelial cells and remodeling of the mammary gland back to a virgin-like morphology. Precocious involution was frequently observed in mouse models that display defects in secretory activation [20, 21, 23, 25]. At lactating day 9, mammary glands from 50% of the KO mice displayed morphology similar to virgin mice, suggesting precocious involution had occurred (Fig 3A). Genes involved in cell apoptosis [41] were enriched in the WT mammary gland at involution stage (Fig 3B). In the mammary glands of lactating KO mice, the apoptosis signature was marginally enriched (Fig 3C). Cleaved caspase-3 signal was detected in the KO alveoli, but not in the WT alveoli at lactation (Fig 3D and 3E). IL-6-triggered activation of JAK-Stat3 pathway is the key modulator of apoptosis during mammary gland involution [42, 43]. IL-6-JAK-STAT3 pathway [41] was enriched in the mammary gland at involution stage (Fig 3F), and in the lactating KO mammary glands (Fig 3G). Expression of Cebpd and Socs3, genes that are specifically expressed at the involution stage [42–44], was significantly upregulated in the lactating KO mammary glands (Fig 3H). Expression of Th-POK in HC11 mammary epithelial cells did not affect expression of Cebpd and Socs3 (Fig 3I), suggesting Th-POK did not proactively regulate Cebpd and Socs3 expression and the premature involution program. Indeed, cleaved caspase-3 signal was not detected in the KO alveoli at late pregnancy (Fig 3D and 3E).
Th-POK deficiency disrupts the development of CD4 T cells [4]. To investigate if the failure in milk lipid secretion is due to altered immune environment, mammary epithelial cells isolated from wild-type or Th-POK-deficient mammary glands were implanted into the cleared fat pads of wild-type or Th-POK KO mice (Fig 4). Both wild-type and Th-POK-deficient mammary epithelial cells developed into properly structured mammary glands, further reinforcing the notion that Th-POK is not required for mammary gland development (Fig 4A). Mammary glands developed from the wild-type mammary epithelial cells successfully released lipid into the luminal space with small lipid droplets retained in the mammary epithelial cells on lactation day 2, regardless the genotype of the recipient mice (Fig 4). However, the mammary glands developed from the Th-POK-deficient mammary epithelial cells retained large size lipid droplets within the mammary epithelial cells upon parturition in both wild-type and Th-POK-deficient recipient mice (Fig 4). These data collectively suggested that the defective lipid secretion phenotype observed in the Th-POK-deficient mice was mainly due to the defects in the mammary epithelial cells.
Cell polarity is prerequisite for milk secretion [39]. The apical marker Ezrin and basolacteral marker E-cadherin were properly expressed and localized in the KO alveolar epithelial cells (S4A Fig). Mice deficient of BTN [21] or heterozygous for XOR [23] had defective lipid droplet secretion and accumulated large lipid droplets in the cytoplasm of mammary epithelial cells after parturition. Expression of BTN and XOR or the XOR reductase activity was indistinguishable between WT and KO mammary glands (S4B–S4D Fig). Expression of lipid droplet surface protein Plin2 or Cidea was also unchanged between WT and KO mammary glands (S4B Fig) [24, 40]. Mice deficient of Src or its binding protein AFAP1 exhibit retention of large CLDs [20, 25]. However, there was no significant difference in Src signature [45] (NES = -1.09, P = 0.285) between WT and KO mammary glands (S4E Fig). Western blot analyses also revealed that neither Src expression nor its phosphorylation was disturbed in the lactating KO mammary glands (S4C Fig).
Expression of the genes regulating lipid biosynthesis increases at the onset of lactation [17, 27, 34]. SREBPs are the key transcription factors regulating fatty acid and cholesterol biosynthesis [31–33]. SREBP-1 was proposed as a critical regulator of milk lipid synthesis and secretion at lactation [17, 27, 34]. Genes regulating fatty acid synthesis (S5A and S5D Fig) [27] and SREBP target genes (S5C and S5F Fig) [46] were enriched, whereas genes regulating fatty acid oxidation (S5B and S5E Fig) [27] were negatively enriched in the lactating mammary glands compared to that at late pregnancy. The SREBP signature correlated to Th-POK expression levels during pregnancy and lactation (S5G and S5H Fig). Consistent to the defects in milk lipid production in the KO mammary glands, the SREBP signature and genes regulating fatty acid synthesis were enriched in the lactating WT mammary glands, whereas genes regulating fatty acid oxidation were enriched in the lactating KO mammary glands (Fig 5A–5C). Higher levels of nuclear SREBP1 were detected in the lactating wild-type mammary glands, compared to the KO mammary glands (Fig 5D and 5E). RT-qPCR analyses revealed that genes in the SREBP pathway were significantly downregulated in the lactating KO mammary glands and isolated mammary epithelial cells (Fig 5F and 5G). Th-POK is upregulated at late pregnancy and lactation (Fig 1K–1M and S1 Fig). Lactogenic factors insulin, but not prolactin or dexamethasone, induced Th-POK expression in HC11 mammary epithelial cells (Fig 5H and 5I). Expression of Th-POK in HC11 cells potentiated SREBP target gene expression (Fig 5J), whereas Th-POK deficiency compromised SREBP target gene expression in primary mammary epithelial cells upon insulin stimulation (Fig 5K), suggesting Th-POK regulates insulin-induced SREBP pathway by cell autonomous mechanism in mammary epithelial cells.
Akt-Mammalian target of rapamycin (mTOR) signaling is a central regulator of growth and metabolism [19, 37, 47–50]. SREBP and MTORC1 signatures correlated to each other during pregnancy and lactation (S5I and S5J Fig), and in the WT and KO mammary glands at lactation day 1 (Fig 6A), supporting previous findings that mTOR promotes lipid synthesis by activating SREBP-1 [19, 35–37, 50–53]. MTORC1 signature [41] and genes upregulated by mTOR signaling [54] were significantly enriched in the lactating mammary glands (S6A–S6D Fig). mTOR signaling was active at late pregnancy that was further activated upon parturition (S6E Fig). These promoted us to study if Th-POK functions through mTOR in the mammary glands. The MTORC1 signature [41] correlated to Th-POK expression levels during pregnancy and lactation (S6F and S6G Fig), and were negatively enriched in the KO mammary glands, compared to the WT mammary glands (Fig 6B and 6C). mTOR signaling was less activated in the alveolar epithelial cells in the KO mice at late pregnancy that was minimally further activated upon parturition (S6E Fig). Phosphorylation levels of Akt, mTOR, S6 kinase and S6 were significantly decreased in the lactating KO mammary glands (Fig 6D–6F and S6E Fig). Overexpression of Th-POK in HC11 cells significantly enhanced insulin-induced activation of Akt-mTOR signaling (Fig 6G). Mcl-1 is a key pro-survival factor during lactation, which is induced by mTORC1 signaling [49]. In the lactating Th-POK KO mammary glands, Mcl-1 protein levels were reduced to ~50% of that in the WT tissues (Fig 6D). Th-POK significantly induced Mcl-1 expression in HC11 cells (Fig 6G).
Th-POK deficiency downregulated insulin receptor substrate-1 (IRS-1), Akt1 and mTOR expression in lactating mammary glands (Fig 7A–7E). Overexpression of Th-POK in HC11 cells significantly upregulated IRS-1 expression (Fig 7F and 7G). IRS-1 is the major adaptor protein transducing insulin signaling. Mammary IRS-1 expression levels were increased upon parturition [55], and were significantly correlated to that of Th-POK and insulin receptor during pregnancy and lactation (S7 Fig). IRS-1-deficient mice displayed a reduction in insulin-dependent Akt activation in the mammary glands and reduced lactation capacity [55]. Potential Th-POK-binding site in the Irs1 5'-UTR, but not in the proximal promoter, effectively bound Th-POK (Fig 7H), suggesting Th-POK directly regulated IRS-1 expression. IRS-1 knockdown in HC11 cells significantly decreased Akt and S6 phosphorylation elicited by Th-POK overexpression (Fig 7I).
Transition from pregnancy to lactation upon parturition is a critical event in producing sufficient milk and nutrients to the infants [12, 14, 17]. Such a process, however, is not well understood. In this report, we show that Th-POK, a key determinant in T cell development, is an important regulator of mammary gland lactation onset upon parturition. Th-POK is restrictedly expressed in the luminal epithelial cells in mammary glands that activates mTOR-SREBP signaling and lipid biosynthesis at the transition from pregnancy to lactation.
Th-POK is expressed in the mammary glands in a manner similar to GATA-3 (Fig 1), a transcription factor most highly expressed in the mammary glands [15, 16]. GATA-3-deficient mice exhibited severe defects in mammary gland development during puberty and defective alveolar differentiation during pregnancy [15, 16]. Although Th-POK is the master regulator in T cell fate determination downstream of GATA-3 [6, 7], Th-POK is not involved in regulating mammary epithelial cell fate determination (S2 Fig). Th-POK-deficient mice were largely normal in mammary gland development during puberty and pregnancy. GATA-3 regulates T cell differentiation via both Th-POK-dependent and -independent mechanisms [6]. GATA-3 may regulate mammary luminal cell differentiation independent of Th-POK. Effectors, e.g. FOXA1 [15] but not Th-POK, may account for the luminal cell fate determination in the mammary glands.
Th-POK expression is upregulated at late pregnancy and is maintained at high levels at lactation (Fig 1). Such expression kinetics indicates that Th-POK may function primarily at the lactation stage. Expression of milk proteins, e.g. caseins and WAP, is turned on at early- to mid-pregnancy [17]. In contrast, milk lipid synthesis mainly occurs at late pregnancy, and is sharply upregulated upon parturition [17]. The upregulated milk protein expression and secretion of lipid into the lumen are indicative of successful functional differentiation [12] and lactation onset [17], respectively. Deficiency of Th-POK did not affect milk protein production, but substantially decreased the amount of milk lipid (Fig 2). Such phenotypic abnormality further indicates that Th-POK is not involved in the regulation of mammary gland differentiation, but responsible for regulating efficient transition from pregnancy to lactation.
Interaction between XOR and BTN and XOR oxidoreductase activity are essential to the secretion of CLDs [21, 23, 26]. Th-POK, however, did not regulate XOR expression or activity. Rather, Th-POK regulates mTOR-SREBP pathway. Mammary epithelial cells coordinately upregulate multiple biosynthetic pathways in order to fulfill vast demand for milk lipid and protein production during lactation. Akt-mTOR signaling is a central regulator of metabolism [19, 37, 47–49]. Blockade of mTOR signaling results in lactational insufficiency and decrease in pup weight [49]. Akt-mTOR signaling promotes lipid synthesis by activating SREBP-1 [19, 35–37, 50–53], the latter was proposed as a critical regulator of transition from pregnancy to lactation [17, 20, 27, 34]. mTOR and SREBP-1 pathways were coordinately activated in mammary glands upon parturition. Th-POK regulates insulin-induced mTOR signaling and SREBP-1 activity by binding to Irs1 locus and regulating IRS-1 expression (Fig 7). Deficiency of Th-POK compromised Akt-mTOR signaling and downstream SREBP activity. Conversely, insulin induced Th-POK expression in mammary epithelial cells (Fig 5). Therefore, Th-POK may function as a feed forward regulator of insulin signaling in the mammary glands. Interestingly, cytosolic NEFA level was increased in Th-POK deficient mammary epithelial cells (Fig 2). Expression of genes responsible for TAG synthesis was substantially decreased in Th-POK deficient mammary epithelial cells (Fig 2). mTOR signaling not only promotes lipid biogenesis, but also inhibits lipolysis [56]. Lipa expression was upregualted in Th-POK deficient mammary epithelial cells (Fig 2). High NEFA levels in Th-POK deficient mammary epithelial cells may be due to impaired TAG synthesis and enhanced TAG breakdown, as a result of reduced mTOR signaling. In addition to reduced mTORC1-mediated SREBP activation, high level cytosolic NEFA may further exert feedback inhibition on SREBP.
Besides IRS-1, expression of multiple components in the insulin-Akt-mTOR pathway were altered in the Th-POK deficient lactating mammary glands, e.g. Akt1 and mTOR. Akt1 were more significantly downregulated in the Th-POK deficient lactating mammary glands at the protein level than the mRNA level (Fig 7). Although Akt1 mRNA level was not changed, its protein level was significantly higher in the Th-POK expressing HC11 cells (Fig 7). In addition to the transcriptional regulation, Th-POK may regulate Akt-mTOR pathway at the protein level. Th-POK may also regulate the response of mammary epithelial cells to other lactogenic stimuli, e.g. prolactin, which coordinates the expression of milk protein and a subset of mammary epithelial cell specific lipogenic gene expression [30].
Defective transition from pregnancy to lactation in the Th-POK-deficient lactating mice is mainly attributed to the defects in the mammary epithelial cells in an autonomous manner. Transplanted Th-POK-deficient, but not the wild-type, mammary epithelial cells show defective transition upon parturition, independent of the Th-POK expression status in the mammary gland microenvironment (Fig 4). The possibility exists that Th-POK in the microenvironment may as well affect milk lipid biosynthesis and secretion in the alveolar epithelial cells. Although the difference was not statistically significant, size of the CLDs in the Th-POK deficient recipient mice was slightly larger than the ones in the wild-type recipient mice (Fig 4). Th-POK deficiency impairs CD4 T cell differentiation [1–3]. Infiltration of lymphocytes into the mammary gland during lactation is critical for the passive immunity to the newborn [57, 58]. It warrants further investigation whether Th-POK in the infiltrated immune cells regulates milk secretion process. Th-POK is recently reported to be expressed in the brown adipose and regulates brown fat development [59], providing another line of evidence that Th-POK regulates metabolic processes. Th-POK deficient mice lose mammary adipose tissues at lactation, despite mammary adipose tissues are largely normal at puberty and pregnancy (Fig 2 and S2 Fig). Fatty acids for milk lipid synthesis originate either from de novo lipogenic pathways within the alveolar epithelial cells, or from dietary fat and adipose tissues [60, 61]. De nove lipogenic pathways are the major source of fatty acids for milk triacylglycerol synthesis for mice on chow diet [60–62]. Th-POK deficient dams may mobilize more fatty acids from the mammary adipose tissues, due to misregulated lipogenesis in the Th-POK deficient alveolar epithelial cells.
Premature involution is frequently observed in mouse models with defects in secretory activation/initiation, including in the Th-POK deficient mice [20, 21, 23, 25]. Defective secretory activation/initiation triggers alveolar cell apoptosis and premature involution. Alternatively, existing premature involution program may impair milk production and secretion. Although apoptosis was detected in a small population of Th-POK deficient alveolar epithelial cells at lactation, no apoptotic cell was evident at late pregnancy (Fig 3). Expression of Cebpd and Socs3 was upregulated in the lactating Th-POK deficient mammary glands (Fig 3). However, Th-POK does not affect the expression of Cebpd and Socs3 in HC11 mammary epithelial cells (Fig 3), indicating that Th-POK does not proactively regulate Cebpd and Socs3 expression and the premature involution program. mTOR pathway activation was impaired in the Th-POK deficient mammary glands. The difference in mTOR activity is already evident at late pregnancy (S6 Fig). Unlike in the wild-type mammary glands, mTOR signaling is not further activated in the Th-POK deficient mammary glands upon parturition (S6 Fig). Therefore, it is more likely that mTOR signaling is compromised in the Th-POK deficient alveolar epithelial cells and the resultant inability to initiate adequate milk secretion triggers involution program, i.e. Cebpd and Socs3 expression and alveolar epithelial cell apoptosis. Mcl-1 was recently reported as a key pro-survival factor during lactation [49]. Th-POK induces Mcl-1 expression (Fig 6). It warrants further investigation whether lower level Mcl-1, as a result of impaired mTOR signaling, is responsible for the induction of the premature involution in the lactating Th-POK KO mammary glands.
In summary, we discovered a novel physiological function of Th-POK in mammary glands. Th-POK is expressed in the mammary luminal epithelial cells. Unlike its role in T cell fate determination, Th-POK is not required for mammary epithelial cell fate determination and mammary gland development during puberty and pregnancy. Upregulated Th-POK expression in mammary glands at late pregnancy and lactation is necessary for mTOR-SREBP pathway-mediated lipid synthesis in the mammary epithelial cells. Ablation of Th-POK impairs secretory activation and efficient milk production upon parturition. These data demonstrate Th-POK as a critical regulator of mammary gland functions in lactation. Lineage-restrictedly expressed Th-POK exerts distinct and independent biological functions in T cells and mammary epithelial cells in a tissue-specific manner.
The mice were housed in a specific pathogen-free environment at the Shanghai Institute of Biochemistry and Cell Biology (SIBCB) and treated in strict accordance with protocols approved by the Institutional Animal Care and Use Committee of Shanghai Institute of Biochemistry and Cell Biology (Approval number: SIBCB-NAF-15-003-S325-006).
Th-POK knockout mice in C57/Bl background were gift from Dr. R. Bosselut (National Institutes of Health, Bethesda, MD). The mice were fed with standard rodent chow diet. Age matched female mice were used.
Female mice were mated with males between age 10–12 weeks. Pup survival was calculated as the percentage of surviving pups in each litter at 48 hours postpartum [20]. For cross feeding experiment, 7 heterozygote pups born to wild-type female mice were fostered to each WT or KO mother at the time of parturition. The pups were weighed daily for 8 days and average pup weight was calculated. Survival of the pups was recorded to perform the Kaplan-Meier survival analysis.
Pups were removed for 3 hours at day 2 of lactation and dams were injected with 10 units of oxytocin to induce milk letdown. Milk was manually removed from the fourth mammary glands. For MFG detection, milk was diluted in PBS and stained with Bodipy @493/503 (Life Technologies). The captured images were analyzed by ImagePro Plus (v6.0, Media Cybernetics). Triacylglycerol analyses were performed following manufacturer's protocols (Triglycerides kit, Shanghai KEHUA bio-engineering CO.).
The carmine alum-stained whole mounts were prepared as described [63]. To measure ductal growth, the whole mounts of mammary glands stained with Carmine alum were imaged with a microscope. The captured images were imported to ImagePro Plus (v6.0, Media Cybernetics) and analyzed as described [64].
Mouse mammary glands were isolated and fixed in 4% PFA followed by embedding in paraffin. Paraffin-embedded tissues were sectioned and stained with hematoxylin and eosin (H&E). The immunofluorescent and immunohistochemical staining were performed as previously described [65]. Sections were incubated at 4°C overnight with primary antibodies. Antibodies used are listed in S1 Table. H-score was used for semi-quantitative analysis of the immunohistochemical staining. Mammary epithelial cells (three mice and five randomly selected fields of each sample) were counted, and the H-score was calculated by adding the percentage of strongly stained (×3), moderately stained (×2), and weakly stained cells (×1) as described [66, 67]. H-score has a possible range of 0–3. Fore regions of the mammary glands were compared in the histological and immunohistochemical analyses.
Two hours before sacrifice, mice were injected intraperitoneally with 5 μl BrdU/g of body weight (Roche). Mammary glands were prepared as described above. For detection of incorporated BrdU, 7-μm sections were processed according to the manufacturer’s instructions (Roche). The BrdU-positive and total mammary epithelial cells were counted. Three mice and five randomly selected fields of each sample were analyzed.
Mammary fat pad transplantation experiment was performed as described [68]. Mammary epithelial cells isolated from wild-type or Th-POK-deficient mice were injected into the cleared fat pads of 3-week-old wild-type or Th-POK-deficient female mice. 8–10 weeks after surgery, the mice were mated with wild-type male mice. Reconstituted mammary glands were harvested after parturition at lactation day 2, and subjected to paraffin embedding, histology and immuno-staining.
Mammary glands from 8- to 12-wk-old virgin female mice or mice at lactation day 2 were isolated. The minced tissue was digested in RPMI 1640 with 25 mM HEPES, 5% FBS, 1% PSQ, 300U/mL Collagenase III (Worthington) for 2 hours at 37°C. After lysis of the red blood cells, single cell suspension was obtained by sequential incubation with 0.25% Trypsin-EDTA for 5 min and 0.1 mg/mL DNase I (Sigma) for 5 min at 37°C with gentle pipetting, followed by filtration through 40-μm cell strainers. The antibodies used for flow cytometry were: PE-conjugated CD31, CD45, TER119 (BD PharMingen), CD24-PE/cy7 and CD29-APC (Biolegend). All sorting was performed using FACSAria or FCASJazz (Becton Dickinson). The purity of sorted population was routinely checked and ensured to be >95%. Cells were harvested for quantitative RT-PCR experiments. For insulin treatment experiment, primary mammary epithelial cells were seeded in DMEM (Invitrogen) supplemented with 10% FBS. Adhered cells were treated with 10μg/mL insulin for 24 hours and were harvested for quantitative RT-PCR experiments. Cytosolic triacylglycerol and nonesterified fatty acid in primary mammary epithelial cells were quantified following manufacturer's protocols (Triglycerides kit, Shanghai KEHUA bio-engineering CO., and LabAssay NEFA kit, Wako).
HC11 cells (generously provided by Dr. Peng Li, Tsinghua University) were cultured in 1640 (Invitrogen) supplemented with 10% fetal bovine serum (FBS, Biochrom), 100 units/ml penicillin, 100 μg/ml streptomycin (Invitrogen), 5 μg/ml insulin (Sigma) and 20 ng/ml EGF (Abcam) in 5% CO2 at 37°C. The cell lines were routinely tested for mycoplasma contamination. For insulin treatment experiments, HC11 cells were cultured in the absence of insulin and EGF overnight before 5μg/mL insulin treatment for 24 hours. Cells were harvested for western blotting and quantitative RT-PCR experiments.
FLAG-tagged Th-POK was subcloned into pCDH-puro lentiviral vector [11]. The shRNA sequences were cloned into the pLKO.1-puro lentiviral vector. The shRNA target sequences were 5’-GCCTGGAGTATTATGAGAACG-3’ and 5’-GCGATTTCCGAAGTTCCTTCC-3’ (mouse IRS-1). Virus packaging and infection were performed as described [65].
Western blot analysis was performed as previously described [65]. Antibodies used are listed in S1 Table.
Total RNA was prepared from mouse tissues or HC11 cells using Trizol reagents (Invitrogen). Equal amounts of RNA from mammary gland were subjected to quantitative RT-PCR using SYBR green with the BIO-RAD Q-PCR Systems according to the manufacturer's protocol. Relative expression levels were calculated using the comparative CT method. Gene expression levels were normalized to Actin. The primers used are listed in S2 Table.
Real-time PCR-based ChIP analysis was performed as described previously [69]. Cells were incubated with medium containing 0.9% formaldehyde for 10 min at room temperature. Sonicated chromatin fragments averaged ~300 to 500 bp. Soluble chromatin was incubated with the M2 Anti-FLAG Agarose (Sigma). DNA was quantitated by real-time PCR. The amounts of products were determined relative to input chromatin. The primers used are listed in S2 Table.
Gene expression profiles of mammary gland tissues from wild-type and Th-POK knockout mice on lactation day 1 were analyzed using Agilent Mouse 4 × 44 K Gene Expression Arrays, following the manufacturer's instructions. Four independent sets of biological replicates of mammary gland samples were used. Data were normalized by Quantile algorithm, Gene Spring Software 11.0 (Agilent technologies, USA). Gene expression data are available at the Gene Expression Omnibus (GEO) under accession number GSE97566. Gene expression data of mammary glands during pregnancy, lactation, and involution were downloaded from GEO (GDS2843 and GDS1805) [17, 27]. Gene set enrichment analysis (GSEA) were performed on gene signatures obtained from the MSigDB database v5.0 (March 2015 release) [27, 41, 46, 54]. Statistical significance was assessed by comparing the enrichment score to enrichment results generated from 1,000 random permutations of the gene set to obtain P values (nominal P value) and false discovery rate (FDR).
Sample size for each figure is denoted in the figure legends. Statistical significance between conditions was assessed by two-tailed Student’s t-tests. For multiple group comparison, two-way Anova analysis was performed followed by Bonferroni's multiple comparison test. All error bars represent SEM, and significance is denoted as *P < 0.05, **P < 0.01 and ***P < 0.001. n.s. denotes not significant.
|
10.1371/journal.pcbi.1003439 | Consequences of Converting Graded to Action Potentials upon Neural Information Coding and Energy Efficiency | Information is encoded in neural circuits using both graded and action potentials, converting between them within single neurons and successive processing layers. This conversion is accompanied by information loss and a drop in energy efficiency. We investigate the biophysical causes of this loss of information and efficiency by comparing spiking neuron models, containing stochastic voltage-gated Na+ and K+ channels, with generator potential and graded potential models lacking voltage-gated Na+ channels. We identify three causes of information loss in the generator potential that are the by-product of action potential generation: (1) the voltage-gated Na+ channels necessary for action potential generation increase intrinsic noise and (2) introduce non-linearities, and (3) the finite duration of the action potential creates a ‘footprint’ in the generator potential that obscures incoming signals. These three processes reduce information rates by ∼50% in generator potentials, to ∼3 times that of spike trains. Both generator potentials and graded potentials consume almost an order of magnitude less energy per second than spike trains. Because of the lower information rates of generator potentials they are substantially less energy efficient than graded potentials. However, both are an order of magnitude more efficient than spike trains due to the higher energy costs and low information content of spikes, emphasizing that there is a two-fold cost of converting analogue to digital; information loss and cost inflation.
| As in electronics, many of the brain's neural circuits convert continuous time signals into a discrete-time binary code. Although some neurons use only graded voltage signals, most convert these signals into discrete-time action potentials. Yet the costs and benefits associated with such a switch in signalling mechanism are largely unexplored. We investigate why the conversion of graded potentials to action potentials is accompanied by substantial information loss and how this changes energy efficiency. Action potentials are generated by a large cohort of noisy Na+ channels. We show that this channel noise and the added non-linearity of Na+ channels destroy input information provided by graded generator potentials. Furthermore, action potentials themselves cause information loss due to their finite widths because the neuron is oblivious to the input that is arriving during an action potential. Consequently, neurons with high firing rates lose a large amount of the information in their inputs. The additional cost incurred by voltage-gated Na+ channels also means that action potentials can encode less information per unit energy, proving metabolically inefficient, and suggesting penalisation of high firing rates in the nervous system.
| Information is encoded, processed and transmitted in neural circuits both as graded potentials (continuous, analogue) and action potentials (pulsatile, digital). Although sensory and chemical synaptic inputs to neurons are graded [1], in most neurons these are converted into a train of action potentials. This conversion overcomes the attenuation of graded signals that occurs as they are propagated over long distances within the nervous system [2], and may prevent noise accumulation in neural networks because pulsatile signals are restored at each successive processing stage [3], [4]. However, because spike trains use discrete pulses of finite precision they have a lower dimensionality than analogue voltage signals, reducing their signal entropy [4]. Consequently, spike trains can encode fewer states within a given time period than analogue voltage signals. This is borne out by experimental measurements that show the conversion of the graded generator potential into a spike train reduces the information rate [5]–[7]. Thus, non-spiking neurons that encode information as graded potentials typically have much higher information rates than spiking neurons [5],[8],[9].
A drop in the energy efficiency of information coding has also been suggested to accompany the conversion of graded to action potentials [3], [10]. Neuronal energy consumption is dominated by the influx/efflux of ions, which must be pumped back across the cell membrane by the Na+/K+ ATPase consuming ATP [3], [11], [12]. These ion movements can incur substantial energy costs even in graded potential neurons [3], [13]. However, the large Na+ influx during action potentials requires additional cellular energy to extrude, though the precise energy cost will vary among neuron types [11], [14]–[16].
Our aim is to identify the causes of the loss of information and energy efficiency when graded potentials are converted to action potentials. Although some causes of information loss in spiking neurons have been studied previously, such as channel noise [17]–[19] or dimensionality reduction [6], [20], in most cases their effects on information rates have not been quantified. We quantified both information rates and energy efficiency using single compartment models. We compared the information rates, energy consumptions and energy efficiencies of spike trains with those of the generator potentials that triggered the spike trains, and of the graded response produced in the absence of voltage-gated Na+ channels. We find that three previously unreported effects reduce the information rate and efficiency of the generator potential by 50%; namely the finite durations of action potentials, and the noise and nonlinearity introduced by voltage-gated ion channels. The effect of channel noise on spike timing reduces the information rate and efficiency by <10%. We conclude that the conversion of graded signals to “digital” action potentials imposes two penalties; spikes increase energy costs and both spike coding mechanisms and the spike code reduce information rates. As a result energy efficiency falls by well over 90%.
We simulated the responses of a 100 µm2 single compartment model containing stochastic voltage-gated Na+ and K+ channels to a 300 Hz band-limited white-noise current stimulus to assess information coding in a spiking neuron model (see Methods) (Figure 1A,B). By altering the stimulus mean and standard deviation the model captured a wide range of neuronal activity patterns. Low mean, high standard deviation inputs produced voltage responses that resembled relay neurons, the activity of which is dominated by large post-synaptic potentials from relatively few pre-synaptic neurons, such as principle cells of the Medial Nucleus of the Trapezoid Body that receive synaptic inputs from the Calyx of Held [21]. High mean, low standard deviation inputs produced voltage responses that resembled those of integrator neurons, the activity of which is determined by a large number of small post-synaptic potentials, such as motor neurons [22].
By incorporating voltage-gated Na+ and K+ channels within the same compartment as a current input stimulus, we modelled the conversion of an analogue signal into a train of action potentials (APs or spikes), as would occur at the spike initiation zone of a neuron [23]. No extrinsic noise was added to the current stimulus in most of our simulations, consequently stochastic fluctuations of the voltage-gated ion channels were the only noise source. This stimulus produced small, sub-threshold graded fluctuations in membrane potential as well as action potentials approximately 100 mV in amplitude (Figure 1C). These transient 100 mV excursions to the peak voltage produced a skewed probability density function (PDF) of the membrane potential with a long tail (Figure 1D).
We compared the information encoded by the spiking neuron model with that encoded by an equivalent analogue model in response to the same white-noise current stimuli with varying mean amplitudes and standard deviations (Figure 1E,F). The analogue model lacked voltage-gated Na+ channels but was identical to the spiking neuron model in all other respects. In this model, current stimuli produced small, graded fluctuations in membrane potential with an approximately Gaussian PDF (Figure 1F). We extracted the power spectra of the signal and noise from these graded fluctuations and used them to calculate Shannon information rates [24], [25] (Methods). The rates at which spike trains coded information was calculated from the total entropy and noise entropy of the spikes using the direct method [26]. Both models, graded and spiking, encoded the most information when stimulated by low mean, high standard deviation currents and the least information with high mean, low standard deviation currents (Figure 2A,B). Thus, the information rate of both neuron models is critically dependent upon the statistics of the input stimulus.
The information encoded by the graded neuron model for each input stimulus was greater than that of the spiking neuron model (Figure 2A,B). The highest information rate attained by the spiking neuron model was 235 bits/s, whereas the graded neuron model attained information rates of 2240 bits/s. Thus, the graded neuron model encodes almost an order of magnitude more information per second than the spiking neuron model, reproducing experimentally observed differences between graded and spiking neurons [5]–[8].
Information coding in spiking neurons is dependent upon the rate and timing of the action potentials with which it samples the input stimulus [27]. We calculated the firing rate of the spiking neuron model in response to the same set of band-limited white noise current stimuli used previously to calculate information rates (see Methods) (Figure 3A). Increasing the stimulus mean or standard deviation increased the firing rate; low mean, high standard deviation or high mean, low standard deviation stimuli produced approximately 57 spikes/s whereas high mean, high standard deviation stimuli generated the highest spike rates of approximately 86 spikes/s (Figure 3A). Because these firing rates are lower than the maximum firing rates that the spiking neuron model can achieve, the information rates are not limited by the absolute refractory period.
The total entropy of a spike train reflects its total variability over time [26]. The highest total entropy occurred with high mean, high standard deviation stimuli that produced the highest spike rates, conversely, the lowest total entropy occurred with low mean, low standard deviation stimuli that produced the lowest spike rates (Figure 3B). However, noise prevents neurons from achieving the maximal information rates, as bounded by the total entropy [26]. We quantified the differences in action potential reliability by calculating the noise entropy among spike trains generated by many repetitions of an identical current stimulus (Figure 3C). Increasing the stimulus standard deviation increased the number of transients in the stimulus that cross the voltage threshold at high velocity. Consequently, high standard deviation stimuli generated spike trains that were both precise and reliable among trials with low noise entropy (Figure 3C; Figure S1A) [28]. Conversely, as the mean increased the variance in the interspike interval influenced spike timing, reducing the reliability of the spike trains and increasing the noise entropy (Figure 3C; Figure S1B).
The total and noise entropy together determine the information rate of the spiking neuron model for a particular input stimulus (Figure 2A). Low mean, high standard deviation stimuli generated spike trains that have only intermediate firing rates and total entropies but have the highest information rates due to their low noise entropy. High mean, high standard deviation stimuli generated spike trains with lower information rates despite their higher firing rates because noise entropy is higher. Consequently, the information per spike was highest (4 bits/spike) with low mean, high standard deviation stimuli that produced the highest information rates (235 bits/s) with only moderate firing rates (57 Hz), and lowest (0.2 bits/spike) with high mean, low standard deviation stimuli that produced the lowest information rates (10.2 bits/s), also with moderate firing rates (58 Hz) (Figure 3D).
To determine the effect of noise generated by the voltage-gated Na+ and K+ channels on the information rates of the spiking neuron model, we replaced either the stochastic Na+ or K+ channels with deterministic channels thereby eliminating this component of the channel noise. In comparison to the stochastic model, the deterministic Na+ channel model generated more reliable spike trains for a given stimulus (Figure S1A,S2A). Similarly, replacing the stochastic K+ channels in the spiking neuron model with deterministic channels also generated more reliable spike trains for a given stimulus in comparison to the original spiking neuron model (Figure S1A,S2B).
We quantified differences in the reliability between the original stochastic spiking neuron model, the modified model with deterministic Na+/stochastic K+ channels, and the modified model with stochastic Na+/deterministic K+ channels. We compared the total entropy, noise entropy, information rate and information per spike for spike trains generated by low mean, high standard deviation stimuli or high mean, low standard deviation stimuli (Figure 4). All three models produced 50–57 spikes/s in response to the stimuli (Figure 4A). In comparison to the original spiking neuron model with stochastic Na+ and K+ channels, the total entropy of the deterministic K+ channel model was lower by 1–7%, whereas the total entropy of the deterministic Na+ model was almost identical (Figure 4B). The deterministic K+ channel model also had the lowest noise entropy, making the APs more reliable (Figure 4C). Both models with deterministic ion channels had higher information rates than the original model because of their lower noise entropy, but the difference was just 7%, irrespective of the stimulus statistics (Figure 4D). This suggests that channel noise has relatively little impact on the information rate of the 100 µm2 single compartments we modelled. Thus, in addition to channel noise and dimensionality reduction, there must be other sources of information loss.
The information in the spike train comes from the generator potential (Figure 5A). However, the generator potential is not equivalent to the voltage signals produced by the graded potential model, which lacks voltage-gated Na+ channels. We constructed an approximation of the generator potential, the pseudo-generator potential, by removing the action potentials from spike trains and replacing them with a 6 ms linear interpolation of the membrane potential, corresponding to the maximum action potential width (Figure 5A). The pseudo-generator potential probability density function is distorted in comparison to the graded potential being narrower with a more pronounced peak because voltage excursions beyond threshold are truncated, the action potential being replaced with an interpolated response (Figure 5A,B). For a particular stimulus the information rate of the pseudo-generator potential was intermediate between that of the spike trains and that of the graded potential model (Figure 5C). The information rates of the pseudo-generator potential were highest (1094 bits/s) with low mean, high standard deviation stimuli, 860 bits/s (366%) higher than that of the corresponding spike trains but 1146 bits/s (51%) lower than that of the corresponding graded potential (Figure 5C). The information rates of the pseudo-generator potential were lowest (188 bits/s) with low mean, low standard deviation stimuli. This lowest value was 158 bits/s (531%) higher than that of the corresponding spike trains, but 352 bits/s (65%) lower than that of the corresponding graded potential (Figure 5C).
What reduces the information rate of the pseudo-generator potential relative to the graded potential? We identify three processes: the duration of the action potential and associated refractory period, and two effects caused by the presence of voltage-gated Na+ channels, noise and non-linearity. We will assess each of these processes, in turn.
The action potential and accompanying refractory period creates a ‘footprint’ on the generator potential during which information is lost (Figure 6A). To assess the impact of this ‘footprint’ on the information rate, we stimulated the graded model with a white noise stimulus (Figure 1A,B) to generate a set of graded responses from which we could estimate the signal, noise and information rate. These graded responses produced a high information rate (1427 bits/s). We then inserted 6 ms long sections of linear interpolation spaced at least 10 ms apart into the individual graded responses to mimic action potential footprints (Figure 6B). We added between 10 and 80 linear interpolations per second into each response to represent the spike footprints at different firing rates and re-calculated the Shannon information rate (Figure 6B) [25]. Interpolations were added at exactly the same positions in all responses, termed deterministic interpolation (Figure 6B), to represent the footprints of noise-free spikes and give an upper bound on signal entropy. The placement of the interpolations was then jittered by up to 4 ms (Figure 6B), termed jittered interpolation, to represent reliable spike trains with low noise entropy. Finally, interpolations were placed randomly in each response (Figure 6B), termed random interpolation, to resemble unreliable spike trains with high noise entropy.
The Shannon information rate [25] was unaffected by the deterministic or jittered interpolation, irrespective of the number of interpolations inserted (Figure 6C) because it depends only upon the signal-to-noise ratio (SNR) and the response bandwidth [25]. Thus, inserting increasing numbers of interpolations, even when jittered, does not affect the Shannon information rate because these interpolations are inserted in identical (deterministic) or similar (jittered) positions, leaving the regions between the interpolations unaffected. Conversely, increasing the number of random interpolations reduced the Shannon information rate from 1427 to 485 bits/s (Figure 6C) because these interpolations add noise to the responses, thereby reducing the SNR.
In addition to the Shannon information rate [25], we calculated coherence-based information rates to determine the effect of the footprint on information loss from the stimulus (see Methods). The coherence-based estimate of the information rate is a measure of linear dependence between the stimulus and the response, and describes different forms of signal corruption including non-linear distortion [29]. The coherence-based information rate decreased as the number of interpolations inserted increased for all three types of interpolation, deterministic, jittered and random (Figure 6D). The coherence-based information rate dropped from 1148 bits/s with no interpolations to 346 bits/s with 80 interpolations.
Although we inserted linear interpolations into the voltage responses, there is still a fluctuation at the corresponding position in the current stimulus. The mismatch between the interpolations and the stimulus may reduce the coherence-based information rate by inflating the non-linearity. To determine whether this is the case, we added linear interpolations at exactly the same positions to both the stimulus and the response, and recalculated the coherence-based information rate (Figure 6D). This difference between the coherence-based information rates calculated with or without interpolations added to the stimulus as well as the response is the information lost due to the action potential footprint. For the same number of interpolations, all three types of interpolation, deterministic, jittered and random, had higher information rates (between 177 and 516 bits/s) with interpolations added to the stimulus than without (Figure 6D). These coherence-based information rates were dependent upon the number of interpolations inserted. For example, inserting 10 interpolations reduced the information rate from 1148 bits/s to 1090 bits/s but inserting 80 interpolations reduced the information rate to 860 bits/s. Thus, the coherence-based method demonstrates that the action potential footprint blanks out information about the stimulus. This loss of information increases with spike rate from 5.3% at 10 Hz to 33.5% at 80 Hz.
Channel noise affects sub-threshold potentials as well as spike timing and reliability [30]. We measured the standard deviation of the voltage noise at sub-threshold membrane potentials for the spiking neuron model, the deterministic Na+/stochastic K+ channel model, the stochastic Na+/deterministic K+ channel model and the graded neuron model (Figure 7A). In the absence of an input stimulus, the voltage noise was generated entirely by the spontaneous opening and closing of the voltage-gated ion channels. The noise standard deviation of all the models was highest at the most depolarised potentials and dropped as the membrane potential was hyperpolarised towards the reversal potential of the K+ ions (Figure 7A). Between −74 to −70 mV the voltage noise standard deviation was highest for the spiking neuron model and lowest for the stochastic Na+/deterministic K+ channel model.
The voltage noise of the deterministic Na+/stochastic K+ was close to that of the spiking neuron model (Figure 7A). However, near the K+ reversal potential of −77 mV the voltage noise of all three models containing stochastic K+ channels dropped as the driving force on K+ ions approached zero. The drop was less pronounced in the spiking neuron model because stochastic Na+ channels continued to produce noise. Below the K+ reversal potential, the voltage noise of all three models containing stochastic K+ channels increased (Figure 7A), with the driving force on K+ ions.
The voltage noise of the deterministic K+ channel model dropped as the membrane potential was hyperpolarised, even below K+ reversal potential, because the probability of spontaneous Na+ channel opening, the only source of channel noise, drops at hyperpolarised potentials. Indeed, the deterministic K+ channel model had the lowest voltage noise at holding potentials more depolarised than ∼−74 mV and more hyperpolarised than ∼−80 mV (Figure 7A). Thus, although the noise generated by the spontaneous opening of both Na+ and K+ channels contributes to the voltage noise of the spiking neuron model, the K+ channel noise apparently makes the greater contribution at potentials between −74 to −70 mV. Note that the voltage noise standard deviation with both channel types together is less than the sum of the standard deviations of the individual channel types because their variances add.
We assessed the impact of the sub-threshold voltage noise on the Shannon information rate by stimulating each model with a white noise current with a zero mean and low standard deviation (μ = 0, σ = 1, τc = 3.3 ms). An additional tonic current was injected and adjusted to hold the mean membrane potential at either −77 or −70 mV. This tonic current prevented the models containing voltage-gated Na+ channels from reaching threshold, permitting a direct comparison of the effects of stochastic and deterministic channel combinations upon sub-threshold information coding.
We calculated the Shannon information rate [25] of each model at the two mean potentials, −77 and −70 mV (Figure 7B). The highest information rates of all the models occurred at the more hyperpolarised potential because the voltage noise was lower. Due to a distinct drop in voltage noise near the K+ reversal potential, the deterministic Na+/stochastic K+ channel model and the graded neuron model, attain the highest information rates of 3123 bits/s at −77 mV. These information rates were ∼30% greater than those of the sub-threshold spiking neuron model and the stochastic Na+/deterministic K+ channel model, which are lower because of voltage-gated Na+ channel noise. At −70 mV the increased voltage noise in all the models reduces their information rates (Figure 7B). The information rate of the sub-threshold spiking neuron model dropped 86% to 321 bits/s. The sub-threshold information rates of both models with stochastic K+ channels dropped 63% to 1142–1168 bits/s, whilst the stochastic Na+/deterministic K+ channel model has the lowest voltage noise and, consequently, the highest sub-threshold information rate of 1288 bits/s. The drop in the information rates of all the models at the more depolarised holding potential shows the substantial effect of channel noise upon the sub-threshold and graded potentials. The combination of both stochastic Na+ and stochastic K+ ion channels in the spiking neuron model reduce the information content of the sub-threshold potential relative to the graded neuron model by 24% at −77 mV to 73% at −70 mV.
Voltage-gated ion channels introduce non-linearities [31], [32] that could reduce the information content of the generator potential by distorting the voltage signal. We assessed the sub-threshold effect of non-linearity on each of the models, at −77 mV and −70 mV, using the coherence-based information rates we previously calculated to assess the impact of the action potential footprint (Figure 6D). Higher coherence-based information rates indicate better reconstruction of the original stimulus, based solely on linear decoding principles [29]. In the spiking neuron model the coherence-based information rates dropped by more than 63% as the holding potential becomes more depolarised i.e., from 1027 bits/s at −77 mV to 382 bits/s at −70 mV (Figure 7C). This fall indicates a decline in the quality of linear reconstruction. By comparison, the stochastic Na+/deterministic K+ model was the least affected by depolarisation, the coherence-based information rates dropping by just 1.5%. For the model with deterministic Na+/stochastic K+ and the model with only stochastic K+ channels, the coherence-based information rates drop ∼4.2–4.8% at the more depolarised potential (Figure 7C). Increasing the holding potential to −68 mV causes all three models containing voltage-gated Na+ channels to produce spikes, making them increasingly non-linear (data not shown).
In addition to coherence-based information rates, we used the normalised root mean squared error (nRMSE) between the original stimulus and the reconstructed stimulus to assess the effect of non-linearity. An nRMSE value that tends towards zero represents perfect reconstruction [29]. The nRMSE increased as the membrane potential increased indicating a drop in the quality of reconstruction (Figure 7D); the increase in nRMSE was largest for the sub-threshold spiking model (67%) but the nRMSE of the three other models also increased by 8–13%. This decline in reconstruction quality is due to an increase in the open channel probability with depolarization. For the models containing voltage-gated Na+ channels, the voltage threshold for eliciting an action potential is close to −68 mV. At −70 mV the increase in the numbers of open voltage-gated Na+ channels increases positive-feedback and, consequently, the magnitude of the non-linearity. A fluctuating input stimulus superimposed upon the holding current also reduces the distance from the voltage threshold, though the effect of this on reconstruction will depend on the magnitude and polarity of the fluctuations.
Using linear systems analysis (see Methods), we assessed how much of the input (current) can be predicted from the response (voltage) by reconstructing the input stimulus current. We find that when the graded voltage response was used for the reconstruction based on linear decoding the predicted input stimuli were most coherent, with the lowest nRMSE (Figure S3C,D) and the highest coherence-based information rates (Figure S3C,E). The reconstruction accuracy (nRMSE and coherence based information) of the pseudo-generator potentials was lower than that of the graded potentials (Figure S3B,D,E). The highest nRMSE and, consequently, the lowest coherence-based information rate was obtained from reconstructions based on action potentials (Figure S3A,D,E), although these were only marginally worse than reconstructions based on pseudo-generator potentials (Figure S3). Thus, voltage-gated Na+ channels distort both the subthreshold (pseudo-generator) and suprathreshold responses so that the incoming stimulus current cannot be accurately reconstructed using just a linear decoder.
Neuronal information rates are constrained by extrinsic noise in the input stimuli, as well as by intrinsic noise generated by ion channels [33], [34]. To investigate this constraint, we added broadband Gaussian noise to the white noise input stimulus. This enabled us to quantify and compare the effect of extrinsic noise upon the information rates of the spiking model, the pseudo-generator potentials from the spiking model and the graded model. In our simulations, although the presence of the extrinsic noise source facilitates a marginal increase in precision of the APs for inputs with low standard deviations, it does not alter the variability of the APs, consequently noise-aided enhancement of mutual information is absent (cf. McDonnell et al. [35]).
The amount of extrinsic noise was altered to produce an input stimulus with either a low or a high SNR input stimulus (Equations 2 and 3; SNR = 2 or 20). The SNR is defined as the ratio of the signal power to the noise power. In our simulations, we decreased the SNR by increasing the noise power (see Methods; Equation 3). For the spiking model, increasing the input noise produces a relatively small increase in total entropy (∼5%, SNR = 2; ∼2%, SNR = 20) (Figure S4A) but a relatively large increase in noise entropy (∼180%, SNR = 2; ∼50%, SNR = 20) (Figure S4B), and this produces a significant drop in the mutual information (∼40%, SNR = 2; ∼10%, SNR = 20) (Figure S4C,S5A).
The information rates of the pseudo-generator potentials also decrease with increased extrinsic noise (Figure S5B). The loss in relation to the noise-less stimulus is greater in the pseudo-generator potentials (∼69%, SNR = 2; ∼29%, SNR = 20) with higher standard deviation input signals. A 10-fold increase in the input SNR caused a 133% increase in information rate, from 335 bits/sec (SNR = 2) to 780 bits/sec (SNR = 20), compared to 1094 bit/sec in the absence of extrinsic noise. Likewise, the information rates of the graded model were reduced by up to 73% for low SNR input signals (SNR = 2) and by up to 36% for high SNR input signals (SNR = 20) (Figure S4C), the higher quality input signal (SNR = 20) causing the information rate to increase from 595 to 1422 bits/sec. Thus, the information rates of the spiking model were the least affected by the extrinsic noise whilst those of the graded model were the most affected (Figure S5A–C).
The energy consumption of each model was determined from the K+ ion fluxes across the membrane needed to generate the voltage signals, as the number of ATP molecules hydrolyzed by the Na+/K+ pump [12]. This pump maintains the ionic concentration gradients that generate electrical responses and operates stoichiometrically, pumping back 2 K+ ions for every ATP molecule that it consumes [36]. The energy consumption of the spiking neuron model is strongly correlated with its firing rate (Figure 8A) because the energy consumption of an action potential is high compared to the consumption between action potentials. Higher standard deviation stimuli evoke larger membrane potential fluctuations, eliciting more action potentials and, therefore, consuming more energy. Consequently, the high mean, high standard deviation stimuli that evoked the highest firing rates also incurred the highest energy consumption, 3.9*108 ATP molecules/s (Figure 8A). Low mean stimuli with high standard deviations consume 3.1 times more energy than stimuli with low standard deviations but for high mean stimuli it is just 1.4 times more (Figure 8A). This is because the standard deviation of signal fluctuations has less of an effect upon the average firing rate with high mean input stimuli.
Pseudo-generator membrane potentials consume less energy than the spiking neuron model. Indeed the maximum energy consumption of the pseudo-generator potentials is 6.4*107 ATP molecules/s, almost an order of magnitude less than the spiking neuron model (Figure 8B). Like the spiking model, when the pseudo-generator potential model is driven with a high mean stimulus, increasing the stimulus standard deviation increases energy consumption. But, unlike the spiking model, when the stimulus mean is low, increasing its standard deviation reduces energy consumption. Low mean, high standard deviation stimuli consume less energy because they hyperpolarise the membrane potential by 10 mV or more below the resting potential, and this reduces the number of open K+ channels (Figure S6A,B). Conversely, with high mean stimuli the maximum peak-to-peak voltage of the compartment is approximately the same, irrespective of the standard deviation (Figure S6A,B). The greater energy consumption of the high standard deviation is due to the 1.6-fold greater numbers of open K+ channels, which cause a doubling of the mean K+ current at equivalent membrane potentials, thereby inflating the energy consumption.
The energy consumption of the graded model showed the same trends as the pseudo-generator potentials (Figure 8C). Again, less energy is consumed in response to low mean high standard deviation stimuli than to low standard deviation stimuli, due to an 85% decrease in the number of open K+ channels (Figure S7A,B). In contrast, at high means, high standard deviation stimuli consumed 64% more energy than low standard deviation stimuli (Figure 8C) because high input standard deviations open greater numbers of K+ channels (Figure S7A,B).
We calculated the energy efficiency of information coding by dividing the information rates of the spiking neuron model, the pseudo-generator potentials and the graded neuron model by their corresponding energy consumptions. The energy efficiency of the spiking neuron model was highest (8.4*10−7 bits/ATP molecule) for low mean, high standard deviation stimuli and lowest (3.8*10−8 bits/ATP molecule) for high mean, low standard deviation stimuli (Figure 9A). This 22-fold difference in energy efficiency was accompanied by a 23-fold difference in information rate. Thus the coding of low mean, high standard deviation stimuli was most efficient because these stimuli generated the highest information rates with firing rates, and therefore energy costs, similar to high mean, low standard deviation stimuli (Figure 9A). In other words, energy efficiency rises with information per spike.
Indeed, in all models, spiking, pseudo-generator potential, and graded, increasing input stimulus mean reduced energy efficiency because it increased the mean level of response without introducing more information (Figure 9A,B). As expected, the energy efficiency of all three models improved when the information rate increased in response to an increase in stimulus standard deviation at a given stimulus mean (Figure 9A,B). For example at low means, the spiking model's information rate increased by 689% with a concomitant increase in efficiency of 151%. For the pseudo-generator potentials information increased by 482%, and efficiency increased by 889% and in the graded neuron efficiency increased by 363% and information increased by 315%.
Both pseudo-generator (8.0*10−5 bits/ATP) and graded potential (1.3*10−4 bits/ATP) models were 95–156 times as energy efficient as the spiking model (8.4*10−7 bits/ATP), when all models were compared with low mean, high standard deviation inputs. At higher information rates the energy efficiency of both the pseudo-generator and graded potentials improved substantially (Figure 9B). However, the graded potentials achieved higher information rates than the pseudo-generator potentials and in this regime they were as much as 1.6 times more energy efficient at 1.3*10−4 bits/ATP molecule.
The addition of extrinsic noise did not affect this general pattern of relationships between input stimuli, information rate and energy efficiency in the three models. However, by reducing the information rates of all three models the extrinsic noise reduced the energy efficiency for any given input stimulus (Figure 9A,B). For example, adding noise to the inputs reduced the efficiencies of the pseudo-generator potentials by 71% for a low quality input (SNR = 2) and by 26% for a high quality input (SNR = 20). Similarly, the efficiency of the graded potential model dropped by 74% at low SNR and by 36% for high input SNR. Given that extrinsic noise only marginally altered the energy consumption, it decreases efficiency by decreasing the amount of information that can be coded.
Analogue voltage signals in non-spiking neurons and generator potentials in spiking neurons typically have higher information rates than spike trains [5]–[9]. This information loss is a consequence of a change in coding strategy; non-spiking neurons and generator potentials encode information as a continuous analogue voltage signal whereas spiking neurons use discrete pulses of finite precision and width, limiting the number of states that can be coded within a given time period. However, there are also biophysical causes of this information loss, and these were the focus of our study. Spiking neurons can be lossless encoders of band-limited inputs if their spike rates exceed the Nyquist limit [37], both at the level of a single neuron or across a population of neurons [38], [39]. But below this limit information loss occurs and is affected by the factors we have examined.
Our simulations show that voltage-gated Na+ channels, which are necessary for action potential generation, are the primary biophysical cause of information loss in sub-threshold potentials because they increase intrinsic noise and introduce non-linearities. Indeed, this information loss in sub-threshold potentials is greater than the information loss in spike generation attributable to voltage-gated Na+ channels. Further information loss in the sub-threshold potential occurs because each action potential obscures the generator potential, reducing its information content. This suggests that the biophysical factors we identify have their major impact upon sub-threshold information processing. Comparing the energy efficiencies of our models, spike trains consume an order of magnitude more energy than graded or pseudo-generator potentials for a given stimulus. This result emphasizes the two-fold penalty of action potentials on coding efficiency; lower information rates and higher energy costs. Graded and generator potentials consume similar amounts of energy, the primary determinant of which is the input mean, but due to their lower information rates generator potentials are less energy efficient than graded potentials.
Our models contained voltage-gated ion channels with the same biophysical properties as those found in the squid giant axon because well-established kinetic models exist for them [40], [41]. Different channel kinetics will alter channel noise [42], affect the shape of the action potential [11] and alter the information rates of a spiking model [43]. However the main effects of channel noise in our models are on the graded and generator potentials. Previous modeling studies have used squid voltage-gated Na+ channels to show that they increase sub-threshold noise [30], but did not quantify their effect on sub-threshold information rates. We find that the noise from voltage-gated Na+ channels and voltage-gated K+ channel noise substantially decreases the information rate of the generator potential. This finding suggests that the high densities of voltage-gated Na+ channels at the spike initiation zone [44], as well as voltage-gated Na+ channels and Ca2+ channels in dendrites and dendritic spines [45], [46] could also reduce the information rate of sub-threshold signals, and this could have a deleterious effect on information processing.
Our models suggest that action potential duration (including the absolute refractory period) is an important source of information loss, imposing a lower limit on the interspike interval and preventing the spike initiation zone from integrating new information for a brief period. In vivo many neurons have considerably higher spike rates than our models, which had moderate spike rates below approximately 90 Hz. At these high spike rates, substantial portions of the information would be lost from the generator potential, promoting narrower action potentials and sparse codes that require relatively few action potentials [47]. However, many neurons use signals that are considerably longer than typical action potentials such as bursts and plateau potentials [48] that obscure far more of the generator potential and incur a greater information loss. This emphasizes the importance of these long-duration signals as indicators of high salience signals.
The non-linearity of all the models incorporating voltage-gated Na+ channels increases with the sub-threshold depolarization because the positive feedback generated by the Na+ channels increases as the threshold approaches. Thus, at sub-threshold levels the Na+ channels distort the voltage waveform. This distortion could reduce the information content of the sub-threshold potentials, though this depends upon whether the transformation of any synaptic metric (current, conductance, etc.) into the voltage waveform is linear in the sub-threshold regime. Linear as well as non-linear mapping may occur between the synaptic input and the resultant voltage waveform [20], [49]. Voltage-gated Na+ channels may constitute one such non-linearity, distorting the synaptic input [50]. In such cases, although a linear decoder cannot fully represent and recover the input information, a decoder relying on higher order features of the membrane potential may prevent any information loss (also see [39]).
Our use of current rather than conductance as the input stimulus ignores the energy cost associated with conductance inputs, which will reduce the energy efficiency of information coding of all the models. Conductance inputs close to the spike initiation zone will also alter the membrane time constant and affect action potential initiation [51], [52]. Consequently, conductance inputs will affect the bandwidth and temporal precision of all the models and the maximum spike rate of the spiking neuron model [53]. The synaptic channels needed to implement the conductance changes will also contribute noise to the models [54], reducing their information rates. By incorporating extrinsic noise, however, we have shown that the relationships we have found will remain qualitatively similar.
The squid giant axon action potentials that we modeled consume substantially more energy than other vertebrate and invertebrate action potentials [11], [14]–[16], inflating the energy consumption of the spiking neuron model and reducing its efficiency. Nevertheless, the efficiency drop that occurs when generator potentials are converted to action potentials is substantial and will remain, albeit with a smaller difference. The topological class of model (e.g. Type I or Type II) may also influence energy consumption through the dynamics and time course of the ionic and synaptic currents determining the threshold manifold [55]. Indeed, minimizing metabolic consumption in single compartment models [55] leads to the leak and the inward currents competing with each other even before reaching the spiking threshold, via a Hopf bifurcation (Type II). This causes an increase in energy consumption forcing the optimal action potentials to steer away from such bifurcations; gradient descent on metabolic consumption leads to saddle-node bifurcations as in Type I cortical neurons (unpublished observation – BS, personal communication – Martin Stemmler, [55]). The energy consumption of the graded potential neurons will also be affected by changes in the biophysical properties of voltage-gated ion channels, though this is unlikely to substantially affect the relationship between the input stimuli and the energy consumption.
Our models systematically explored combinations of the mean and standard deviation of a Gaussian input. Those spike trains with the lowest information rates and bits per spike were evoked by low standard deviation stimuli, whereas high standard deviation stimuli evoked consistently higher information rates for a given mean stimulus. Consequently, across all our models there was no systematic relationship between the mean spike rate and the information rate, total entropy, noise entropy or coding efficiency (bits per spike). Indeed, the highest and lowest information rates and coding efficiencies were found at similar spike rates. However, these findings are specific to the type of stimuli we used, a randomly varying input signal superimposed on an offset. It is more usual to find that the information rate increases with spike rate whilst the coding efficiency declines because the entropy per spike falls [56], [57]. Non-Gaussian naturalistic stimuli vary more widely than do Gaussians. These larger excursions make the voltage response more nonlinear and engage adaptation mechanisms that, if they affect the signal and noise differently, can change the information rates of both graded potentials and the spike trains they generate [58]. Although there are methods that could allow us to compare the coding and metabolic efficiency of analogue and spiking responses to natural stimuli [6], each modality has its own statistics. Even within a modality different classes of neuron have distinctive firing patterns because they select different components of the input (e.g. retinal ganglion cells [57]). Faced with many particular cases, we chose to start with a general stimulus that identifies factors, such as input signal to noise, that are widely applicable.
As a case in point, in many neurons the mean and standard deviation of the input stimuli and the extrinsic noise are often correlated [59]–[61]. For example, extrinsic noise in synaptic inputs is often correlated with their number and strength, and hence signal amplitude [62]–[64]. Thus, the stimulus space investigated with our models exposes relationships between energy efficiency and information rate that are broadly applicable to a number of different types of neuron. In particular, our models demonstrate that the energy efficiency of spiking neurons can be improved by reducing the mean input and increasing the standard deviation of the signal. Graded neurons achieve this by using predictive coding to eliminate the mean and amplify the remaining signal to fill their output range [65], [66] and these procedures increase both their coding efficiency and their energy efficiency [3]. Our findings demonstrate that spiking neurons can do likewise.
Taken together, our analyses show that the biophysical mechanisms involved in action potential generation contribute significantly to the information loss that accompanies the conversion of a graded input to a spike train. Although we cannot directly relate the proportions of information loss to specific mechanisms, it seems likely that the action potential ‘footprint’ and sub-threshold voltage-gated Na+ channel noise are the major sources of information loss. Viewed as a cost-benefit trade-off, action potentials incur penalties (information loss and energy cost) that are, presumably, balanced against being able to transmit information over considerable distances and preventing noise accumulation during successive processing stages. Reducing the distances over which information is transmitted in the nervous system may favor less conversion of graded signals into spike trains [67]. However, problems associated with accumulating noise during successive processing stages [4] may remain severe. Thus, even in some highly miniaturized nervous systems, neurons with action potentials are likely to be necessary [67].
In conclusion, our modeling of single compartment neurons confirms that a critical step in neural coding, the conversion of an analogue sub-threshold signal to a series of discrete “digital” pulses, is accompanied by substantial information loss. We show that voltage-gated Na+ channels, critical components for the conversion of analogue to digital, reduce the information in sub-threshold analogue signals substantially, and that this loss is compounded by interference from action potentials. Thus, the first step in a hybrid processing strategy to increase efficiency [4], [68], the analogue processing of inputs, is compromised by mechanism used for the second step, the conversion of analogue to digital, and this calls for strategic placement of the spike initiation zone [69]. Some neurons appear to mitigate a small fraction of the loss of information that accompanies the conversion of analogue to digital by transmitting both analogue and digital [70]–[72]. Information may be encoded in the height and width of action potentials [71]–[74] suggesting that spiking neurons may transmit more information than is calculated by treating them as digital pulses. Even in these cases, however, the ‘footprint’ of the action potentials and sub-threshold voltage-gated Na+ channel noise are still likely to cause substantial information loss.
We used a single compartment stochastic Hodgkin-Huxley model of the squid giant axon for our simulations [40], [41]. The model supporting spiking contained two voltage-gated ion channels, transient Na+ and a delayed rectifier K+ along with the leak conductance, while the model producing purely graded signals contained delayed rectifier K+ and leak conductances. The dynamics of the membrane potential is governed by a set of activation and inactivation variables m, h and n with the current balance equation,(1)Cm is the membrane capacitance, are the conductance of the Na+, K+ and leak channels respectively, are the respective reversal potentials, is a time dependent current stimulus and is the input (extrinsic) stimulus noise current. is zero for no input noise simulations. The variables m, h and n follow first order kinetics of the form , where is the steady-state (in) activation function and is the voltage-dependent time constant. The model was driven using a time dependent current – , a 300 Hz Gaussian white noise, filtered using a 40th order Butterworth filter. The voltage resonant frequency of the squid axon model can vary between 100 Hz at 10°C to 250 Hz at 20°C [75]. Therefore, we selected the input cut-off frequency at 300 Hz that is slightly more than the output 3 dB cut-off frequency encompassing the frequency response expected out of an underdamped second-order response (see Figure 3 in Guttman et al. [75]). The mean and the standard deviations of the stimulus were varied in the range 1–10 µA/cm2, enabling comparison to earlier work studying channel noise and its effects on information rates [76]. The stimulus was presented for 1 second and each set of simulations consisted of 60 such trials. is an unfiltered broad-band Gaussian white noise with,(2)where noise variance is computed using(3)Ω denotes the signal-to-noise ratio (SNR).
All Gaussian random numbers were generated using the Marsaglia's Ziggurat algorithm [77]; uniform random numbers were generated using Mersenne Twister algorithm [78]. Deterministic equations were integrated using the Euler-algorithm while stochastic differential equations were integrated using the Euler-Maruyama method, both with a step size of 10 µs. Parameter values are given in Table S1.
Our simulations incorporate Na+ and the K+ voltage-gated ion channels without cooperativity (Figure S8) so that the state transition matrix evolves according to a Markov process [79], [80]. We track the numbers of channels that were either closed or open [79] using the Gillespie algorithm [81]. The Na+ and the K+ channels had 13 states with 28 possible transitions among these states −20 transitions for the Na+ channels and 8 for the K+ channels. As an example, in time interval δt, the probability that the K+ channel remains in state k is , where γk depicts the sum of all transition rates from state k to any possible successive state. During the interval δt no other ion channel changes its state such that the probability of the ion channels remaining in the same state in the time interval δt is ,(4) is the number of Na+ voltage-gated ion channels in state [i, j], [nk] is the number of K+ voltage-gated ion channels in state [k], γij is the total transition rate from state and γk is the total transition rate from state . The transition rate for a particular ion channel state is chosen by drawing a pseudo-random number r1 from a uniform distribution [0, 1] and defining ttrans as . The Gillespie algorithm then selects which of the 28 possible transitions occur in the time interval ttrans [79], [81]. The conditional probability of a particular transition j that occurs in the time interval δt is given by,(5)Here, aj is the product of transition rate associated with transition j and the number of channels in the original state of that transition. The denominator in Eqn. (5) is equal to λ. The particular transition rate is selected by drawing a random number r2 from the uniform distribution [0, 1] and fixing ψ as,(6)The number of voltage-gated ion channels in each state was updated and the membrane potential calculated. An identical algorithm was used for the channel noise in the compartment containing only K+ voltage-gated channels.
Both information-theoretic and linear system analysis are a common place in neuroscience [82], but before providing a detailed exposition for each of these methods, we justify our use of them. The channel capacity for a Gaussian channel [25], [56] allows us to place an upper bound on the Shannon information encoded in the generator potentials under the assumption of an additive Gaussian noise. On the other hand, the “direct method” [26] is a minimal assumption method to derive an estimate of the reduction in entropy per unit time per spike. Although these two calculations enable us to quantify the information loss separately within each domain (graded and spiking), a more appropriate comparison would employ the same metric permitting direct comparison between domains. The Wiener filter [9], [56] permits such a comparison, allowing us to test the fidelity of both the analog and the pulsatile signals using identical linear optimal filtering, giving a lower bound on the information present in the response (e.g. to linearly decode the input stimulus). Thus, if inputs were linearly mapped onto outputs then the information rates from “direct method” and the “Wiener filter” analysis would be identical [82]. The lower our reconstruction error the better our generative model of the output is.
There are several methods that have been used to quantify information rates in spiking neurons. These include histogram based “direct method” [26], context-tree Markov Chain Monte Carlo (MCMC) [83], metric space method [84], binless method [85], compression entropy [6], among others. We have used the widely employed “direct method” to measure the entropy of the responses, primarily due to its simplicity and the separation of mutual information into separate terms capturing variability (spike train entropy) and reproducibility (noise entropy) [26]. The spike train entropy quantifies the variability of the spike train across time. The noise entropy on the other hand, measured the reproducibility of the spike train across trials. These quantities were dependent upon the temporal resolution with which the spikes were sampled, Δt and the size of time window, T. We present a different stimulus current in each subsequent trial (unfrozen noise) to calculate the spike train entropy, while using presentations of the same stimulus current in each subsequent trial (frozen noise) to calculate the noise correlation. We divided the spike train to form K-letter words (K = 2, 4, 6, 8, 12, 16, 24, 32, 48 or 64), where K = T/Δt. We used the responses from the unfrozen noise session, to estimate the probability of occurrence of particular word, P(W). We estimated the total entropy as,(7)We estimated the probability distribution of each word at specified time durations, t so as to obtain P(W|t). Entropy estimates were then calculated from these distributions and the average of the distributions at all times were computed to yield the noise entropy as,(8)〈〉 indicates average over time. The information was then computed as,(9)The spike train entropy and the conditional noise entropy diverge in the limit of Δτ→0, their difference converges to the true finite information rate in this limit [26]. Therefore, we used bias correction methods such that the estimation of entropy was less prone to sampling errors [86]. Using Δt = 1 ms, we varied the spike trains to form words of different lengths. Using these entropy estimates, we extrapolated to infinite word length from four most linear values of the curve of entropy against the inverse of word length.
We used an upper-bound method to calculate the maximum information transferable by the nonspiking responses [25], [56]. This method assumes that the neuronal response and the neuronal noise had independent Gaussian probability distributions in the frequency domain and the noise was additive in nature. In the presence of additive non-Gaussian noise such a method provides us with an upper bound on the channel capacity that is dependent on the entropy power of the non-Gaussian noise distribution [25], [87], [88]. We defined the stimulus S as the mean neuronal response obtained from a frozen noise experiment. The noise in each trial was calculated by removing the average response from the individual responses Ri. Owing to Gaussian assumptions, it required enough data to estimate the mean and variance of the Gaussian probabilities. The actual information might be lower than this bound because a Gaussian distribution has the highest entropy for a given variance. In our simulations, both the response and the noise had an approximately Gaussian distribution. We obtained the mean response power spectrum and the noise power spectrum using the multi-taper spectral estimator and computed their ratio to be the signal-to-noise ratio (SNR) [29]. This is then used to compute the information for the Gaussian channel as,(10)For our simulations, the limits of the integral were taken from k1 = 0 Hz to k2 = 300 Hz. The integral was evaluated using trapezoidal rule.
We performed stimulus reconstruction to test how noise affects the coherence of a linear system [9], [56]. The method involved finding a linear temporal filter to minimize the difference between the real and the reconstructed stimulus. We followed Haag and Borst [89] in the derivation of this filter using Gaussian unfrozen noise as the stimulus set. We used 60 trials that consisted of 1 second period of unfrozen white noise si(t) to obtain the spike trains ri(t) in the form of 1's and 0's with 10 µs resolution. These time domain signals were Fourier-transformed to obtain complex functions Si(f) and Ri(f). Two filters were obtained, either by normalizing the cross-power spectral densities (CPSD) of the stimulus and the spike response by the stimulus power spectral density (PSD) (forward filter) or the spike power spectral density (reverse filter) as demonstrated below, with angle brackets (< >) indicating averages over trials,(11)(12)Using the reverse filter, we estimated the stimulus, as the product between Ri(f) and Greverse(f),(13)The quality of the estimate was evaluated by computing a filter between the original stimulus and the reconstructed stimulus; this is simply the coherence function (γ2(f)) as shown below,(14)(15)The coherence results have been cross-validated using a 65–35 split between the training set and the test set i.e., we used the first 65% of the trials to calculate the reverse filter and then checked its validity on the next 35% of the trials by computing the final filter (Gfnal(f)) or the actual coherence (γ2(f)).
Reconstruction quality was measured using two metrics. First, normalized root mean squared error (nRMSE) between the original stimulus and the reconstructed stimulus was calculated as,(16)A nRMSE value that tends towards zero represents perfect reconstruction. Second, we calculated a coherence based information rate where a higher value indicates better reconstruction,(17)
Energy consumption in our model is defined as the amount of ATP expended during the encoding of the band-limited stimulus current. The Na+–K+ pump hydrolyses one ATP molecule for three Na+ ions extruded out and two K+ ions imported into the cell [11]. We determined the total K+ current by separating the leak current into a K+ permeable leak current and adding it to the delayed rectifier K+ current. We computed the number of K+ ions by integrating the area under the total K+ current curve for the duration of stimulus presentation. In order to derive the energy consumption we calculated the number of ATP molecules used by multiplying the total K+ charge by NA/(2F), where NA is the Avogadro's constant and F is the Faraday's constant.
|
10.1371/journal.pntd.0003605 | Geographic Distribution and Mortality Risk Factors during the Cholera Outbreak in a Rural Region of Haiti, 2010-2011 | In 2010 and 2011, Haiti was heavily affected by a large cholera outbreak that spread throughout the country. Although national health structure-based cholera surveillance was rapidly initiated, a substantial number of community cases might have been missed, particularly in remote areas. We conducted a community-based survey in a large rural, mountainous area across four districts of the Nord department including areas with good versus poor accessibility by road, and rapid versus delayed response to the outbreak to document the true cholera burden and assess geographic distribution and risk factors for cholera mortality.
A two-stage, household-based cluster survey was conducted in 138 clusters of 23 households in four districts of the Nord Department from April 22nd to May 13th 2011. A total of 3,187 households and 16,900 individuals were included in the survey, of whom 2,034 (12.0%) reported at least one episode of watery diarrhea since the beginning of the outbreak. The two more remote districts, Borgne and Pilate were most affected with attack rates up to 16.2%, and case fatality rates up to 15.2% as compared to the two more accessible districts. Care seeking was also less frequent in the more remote areas with as low as 61.6% of reported patients seeking care. Living in remote areas was found as a risk factor for mortality together with older age, greater severity of illness and not seeking care.
These results highlight important geographical disparities and demonstrate that the epidemic caused the highest burden both in terms of cases and deaths in the most remote areas, where up to 5% of the population may have died during the first months of the epidemic. Adapted strategies are needed to rapidly provide treatment as well as prevention measures in remote communities.
| In October 2010, a large cholera outbreak was declared in Haiti and rapidly spread throughout the country, quickly overwhelming the existing health system. Specialized treatment structures were opened rapidly, generally in cities or large villages, and decentralized treatment units or rehydration points were gradually opened later on. To gain insight into the true burden of the cholera outbreak in the community and on potential geographical differences due to accessibility, we conducted a survey in April–May 2011 in a large rural area across four mountainous districts in the Nord department. We interviewed 3,187 households, corresponding to 16,900 individuals, of whom 2,034 (12%) had had diarrhea, probably cholera, since the beginning of the outbreak. The two most remote districts showed higher proportions of population affected by the disease, up to 16.2%, and higher proportions of deaths among patients with probable cholera, up to 15.2%, than the two districts with better accessibility. Remote populations, older patients, severe cases and those not seeking care were at increased risk of dying of the disease. These results show the very high burden of the cholera outbreak in remote areas, emphasizing the need to develop strategies to rapidly provide treatment and prevention measures in remote communities.
| The cholera epidemic in Haiti, which began in 2010 spread rapidly in both urban and rural areas. One month after confirmation of the first case in Mirebalais, in the department of Centre, the whole country had been affected [1,2]. During the first few days, the focus was on case management in hospitals, which were quickly overwhelmed [1]. Gradually, the Ministry of Health (MSPP), together with partners including non-governmental organizations (NGOs), started setting up dedicated treatment facilities ranging from large specialized centers to decentralized oral rehydration points (ORP) in more isolated communities, cholera-specific health education messages, and water and sanitation activities [1,3]. A national training program for cholera management was developed to train clinical staff, nearly all of whom were unfamiliar with the disease [4]. Despite these efforts, over 600,000 cases of cholera and 7,000 deaths were reported by the national health-structure based surveillance system within two years of the first case [5], and, at the time of writing this article, cases are still being reported (http://mspp.gouv.ht/).
In the Nord department, the first cholera cases were officially reported on October 22nd, 2010 (week 42). The first cholera treatment center (CTC) was opened on October 23rd in Cap Haitien, the administrative center of the department, and cholera treatment units (CTUs) were gradually opened in November in the main communes of the department. ORPs only started operating in remote areas in December 2010 and January 2011 (Fig. 1).
Since national surveillance data were based on reports from the health structures and were likely to miss community cases, large retrospective population-based surveys were conducted by Médecins Sans Frontières (MSF) in April and May 2011 to estimate the cholera burden during the first weeks of the epidemic and get insight into health-seeking behavior. Here, we present the results of a survey that was conducted in a large rural, mountainous area across four districts of the Nord department, chosen to facilitate comparison between regions with good versus poor accessibility by road, and with rapid versus delayed response to the outbreak. We also present results of a risk factor analysis carried out in the same area, looking for potential explanatory factors for geographic differences in mortality with the aim of providing information to improve future outbreak response strategies in similar settings.
The procedures followed were in accordance with the ethical standards of the Helsinki Declaration. The National Ethics Committee of Republic of Haiti granted ethical approval and the Ministry of Public Health of the Republic of Haiti gave authorization to perform the survey. Written informed consent for study participation was obtained from all participants.
The Nord department is located on the northern coast of Haiti and encompasses coastal and mountainous areas, with limited road infrastructure (Fig. 1). Health structures are located in urban centers with catchment areas of hundreds of square kilometers. Many houses are not accessible by road and some hamlets are located more than a 10-hour walk from the nearest health structure. The study took place in 2011 in four districts (“communes”) of the Nord department: Plaisance, Pilate, Borgne and Port Margot (Fig. 1).
Villages in the districts of Plaisance and Port Margot are mainly accessible by roads, while villages in the more mountainous districts of Pilate and Borgne are more difficult to reach due to their mountainous terrain. Each district is divided into 6 to 8 sections. According to a 2009 estimate [6], the total population in the four districts was 218,649 inhabitants, of whom 173,903 lived in rural areas targeted by this survey.
From the beginning of the cholera outbreak and until the time of the survey, 29,295 cases and 654 deaths were reported in the Nord department [Rapport journalier MSPP du 22 mai 2011], for an estimated attack rate (AR) of 2.9% and case-fatality rate (CFR) of 2.2%.
MSF, one of the main partners working with the MSPP to treat cholera patients in Haiti, intervened early in Plaisance, where a CTU was opened in epidemiological week 44, 2010, followed by a CTU in week 47 and 5 ORPs in weeks 49 and 50. In Pilate, the intervention was slightly delayed, with a CTU opening in week 47, followed by 2 ORPs in week 49 and another 6 ORPs in week 1, 2011. In Borgne, a CTC opened in week 47, followed by a CTU in week 50, 5 ORPs in week 52, 9 ORPs in week 1, 2011 and 6 ORPs in week 2. In Port Margot, a CTU was opened by the Catholic Church before week 48 and an ORP by the MSPP in week 48, while MSF started late, with one ORP in week 52, 2 ORPs in week 1, 2011 and a CTU and one additional ORP in week 2.
A two-stage, household-based cluster survey was conducted in the study area. The sample size was 16,000 individuals, calculated to estimate an expected crude mortality rate of 0.5 per 10 000 persons per day with a precision of 0.1, an anticipated design effect of 2 and a recall period of 170 days (from October 17th, 2010 to the earliest survey date). In total, 140 clusters of 23 households, with an expected average size of five members per household, were selected in the four districts.
For the first sampling stage, clusters were attributed to each communal section proportionally to the size of the rural population (37 in Plaisance, 34 in Pilate, 42 in Borgne and 27 in Port-Margot). Villages of more than 5,000 inhabitants were considered urban and therefore excluded from the sampling frame. For the second stage, a large number of random geographic points was generated using the R statistical package. These randomly selected points were then mapped using Google Earth; only points with a house found visually within a 50m radius were retained. For each section, the number of points allocated was then randomly selected from the remaining points. The corresponding GPS points were used in the field to locate the initial house of each cluster. The next house was selected by proximity, i.e., next closest house, until 23 households had been visited in each cluster. Households in which no adult was present at the time of the first visit were revisited at least once before the study team left the village.
Data were collected using a standardized questionnaire. After providing written consent, the head of household was asked to provide the age and sex of all household members (defined as persons living under the same roof and sharing meals). For each household member present at the beginning of the recall period, the head of household was asked about episodes of diarrheal illness (defined as at least three watery stools within a 24-hour period) and deaths that occurred during the recall period. More detailed information was collected on the diarrheal episode, or on the most severe one if multiple episodes were reported for the same individual. Information collected included duration and symptoms of the episode, health-seeking behavior (i.e., type(s) of health structure(s) visited or reason for not visiting a health structure), and outcome (i.e., death or survival). Severe cases were defined as those in which patients reported lethargy or altered consciousness during the diarrheal illness. Death was considered related to diarrhea when it was reported as the outcome of the most severe diarrheal episode.
In each cluster, the time and type of transport to the closest village with a health structure (excluding ORPs) was documented.
Double data entry was done in Epidata 3.1 (EpiData, Odense, Denmark) by four trained data entry clerks. Data validation and statistical analysis were performed using Stata 11 (StataCorp, College Station, Texas, USA) and R Statistical Software. As not all clusters achieved a sample of 23 households, weighted analysis was used to adjust for the probability of each household being selected, by dividing the expected household number per cluster (23) by the actual number of households included. In all analyses, we accounted for the clustering of households within the cluster and applied the selection weights. Design effects are reported where relevant.
Each principal outcome was presented as a percentage with its associated 95% confidence intervals. Results were then extrapolated to the overall rural population by applying a weight which multiplies the selection weight by the total rural population of each district divided by the sample size in this district. A geographical representation of each principal indicator was done using a generalized additive model assuming a Poisson distribution and using an isotropic spline to describe the spatial variation of the different indicators [7]. The level of smoothness of the spatial terms was selected using the restricted maximum likelihood method.
Finally, we used a Poisson regression model for the univariate and multivariate analyses of risk factors for cholera morbidity and mortality, and present here crude and adjusted relative risks (RR, ARR) and associated 95% confidence intervals. The district of Plaisance was considered as the reference for comparisons among districts.
The survey took place from April 22nd to May 13th 2011. In total, 138 randomly selected clusters were visited, and information on 3,187 households and 16,946 individuals collected, which corresponded to approximately 9% of the area’s estimated rural population (173,903 inhabitants). Of these, 46 individuals were subsequently excluded from the analysis due to incorrect inclusion criteria (n = 28) or missing data (n = 18). The median number of individuals per household was 5 (range 1 to 20 persons). The male/female ratio was 0.91 and the median age was 21 years (IQR: 11–40).
In total, 2,034 persons (12.0%) reported at least one episode of watery diarrhea during the recall period (Fig. 2). Among them, 1,979 (97.3%, 95% CI: 96.0–98.2) reported a single episode during the recall period (range 1–4). The median length of the most severe episode investigated in each individual who reported watery diarrhea was 3 days (IQR: 2–4; range 1–15). Of those individuals reporting diarrhea, 68.9% (95% CI: 63.1–74.1) reported vomiting, and 38.3% (95% CI: 33.5–43.4) lethargy or altered consciousness during the diarrheal illness, and were therefore considered as severe cases.
The attack rate of watery diarrhea in the area during the recall period was 12.0% (95% CI: 10.8–13.2), with a design effect of 5.9. Extrapolated to the rural population in the four districts, this translated into an estimate of 21,681 individuals (95% CI: 19,440–23,922) suffering from watery diarrhea during the recall period.
The geographical distribution of attack rates showed marked disparities, with attack rates estimated at more than 20% in some sections in the west of Borgne and Pilate and lower than 10% in most sections of the Plaisance and Port Margot districts (Fig. 3). This was reflected in the estimated attack rates by district, which ranged from 8.6% in Port Margot to 16.2% in Borgne (Table 1).
In total, 275 individuals were reported to have died during the recall period, leading to a crude mortality rate estimate of 0.82 deaths per 10,000 persons per day (95% CI: 0.64–1.05), which represented 1.62% (95% CI: 1.26–2.07) of the population during the recall period or 2,925 (95% CI: 2199–3651) deaths of all causes when extrapolated to the rural population of the four districts (Plaisance: 393; Port Margot: 246; Pilate: 746; Borgne: 1540). Most of these deaths (84.8%; 95% CI: 77.5–90.0) were attributed to diarrhea.
Of the 2,034 diarrhea cases, the outcome of the episode was death in 224, for a CFR of 11.0% (95% CI: 8.6–13.9) with a design effect of 3.8. Extrapolated to the rural population of the four districts, this represents 2,375 (95% CI: 1,710–3,040) deaths due to diarrhea during the recall period (Plaisance 256; Port Margot 155; Pilate 609; Borgne 1,355).
The overall CFR in both Borgne and Pilate was significantly higher than in Plaisance, the reference district (Table 1). The highest CFR (up to 30–40%) were found in western Borgne and Pilate (Fig. 3).
Of 2,030 individuals reporting diarrhea and for whom information on health-seeking behavior was available, 1,447 (71.2%, 95% CI: 66.3–75.6) sought care in a health structure. More than 50% of those who sought care visited a specialized CTC or CTU, and only 3% reported using the ORPs (Table 2). Overall, the main reasons for not seeking care were that the health structure was too far or that the diarrhea was not perceived as requiring care or not perceived as cholera. Among the most severe cases, almost two-thirds reported distance as the main reason for not seeking care (Table 2).
The lowest proportion of individuals seeking care was in the remote areas of western Borgne and Pilate (Fig. 3). Of the four districts, the highest proportion of patients seeking care was in Port Margot (83.2%) and lowest in Borgne (61.6%) (Table 1). The reasons for not seeking care also varied by district: distance was cited as a barrier by 52.7% of patients who did not seek care in Borgne but was less cited in Pilate (20.8%), Plaisance (14.5%) and Port Margot (13.5%). In the latter two districts, the main reason for not seeking care was a combination of no perceived need and illness not perceived as cholera (Plaisance: 59.7%; Port Margot: 55.7%, Pilate: 41.4%, Borgne: 42.7%).
A stratified analysis of risk factors by district showed that similar factors contributed to higher CFR across all districts: older age (> = 60 years), greater severity of illness, living in remote areas, and not seeking health care (Table 3). These factors were also found to have a significant association in the univariate analysis (Table 4). Factors associated with highest risk were severity of disease (RR = 8.1) and not seeking care (RR = 5.1). There was no significant difference in case fatality between males and females.
Stepwise introduction of risk factors in a multivariate analysis showed that the differences between districts remained significant when adjusted for age and severity of disease. Due to colinearity between the two variables, remoteness and health-seeking behavior were introduced separately in the model. In each model, older age, severity of disease, and health-seeking behavior or remoteness were associated independently with a higher risk of death (Table 4). Interestingly, when remoteness was introduced in the model, the differences between districts were no longer significant.
The results of this large community-based survey on the burden of cholera during the first six months of the outbreak in a rural and mountainous area in the northern part of Haiti show very high attack rates and case fatality rates. It highlights important geographical disparities in the four districts investigated, and in particular, the higher risk of both disease and death in the most remote areas. Both the attack rate and case fatality rate found through the survey were more than four times higher than those calculated using data recorded by the national surveillance system in the same period in the Nord department. Moreover, the extrapolated number of cases in the rural populations of these four communes only (21,681 for a population of 173,903) almost reached the total number of cases reported in the national surveillance for the whole department until May 22nd (29,295 for a population of 1,004,247), while the extrapolated number of diarrhea-related deaths in the four communes (2,375) was 3.5 times higher than the total number of deaths (654) reported in the whole department over the same period. This acute underreporting of cases and deaths through the national surveillance system derived from health facility-based cases highlights the importance of community data to better estimate disease burden in areas where national surveillance system may encounter major limitations due to the limited access of the population to health structures. Such data are crucial for targeting the most urgent responses to the highest-priority areas. To achieve this goal, local social leaders (head of villages, religious leaders, etc.) and associations should be mobilized early on to participate in both sensitization and community-based surveillance.
Very remote areas were particularly affected by the outbreak, in terms of both number of cases (high attack rates) and diarrhea-related deaths (high CFR). This led to extremely high mortality rates estimates which suggest that up to 5% of the populations in these areas may have died during the first months of the epidemic. Rural areas are generally thought to show lower attack rates than urban and more crowded areas. For example, MSF generally projects attack rates of 0.2%-1 in rural and 1–5% in urban settings, based on a review of MSF programs in previous cholera epidemics [8]. Our data, as well as others suggest that these estimates should be revised [9,10].
In all districts, CFR were particularly high in elderly people (> 60 years old), in patients with severe diarrhea, those living in remote areas accessible only by foot, and those who did not seek care. In a multivariate analysis, older age, severe diarrhea and not seeking care were independently associated with an increased risk of death. We did not find any association with sex, as reported in other studies, while other risk factor identified elsewhere, such as larger household sizes and being in poor health at onset of disease were not investigated here [11,12]. Not seeking care, in contrast, was reported in all studies with a similar adjusted odds ratio of 5.4 in Guinea Bissau. Health-seeking behavior was influenced by the type of information received in Zimbabwe, with person-to person communication by village health workers being more efficient than other sources of information such as friends, family, NGOs or radio [12]. Here, distance rather than lack of information seemed to be the main barrier to health-care seeking in remote areas.
Accordingly, in a separate multivariate model, remoteness was also independently associated with an increased risk of death, with an adjusted risk ratio of 2.20. In addition, in this model, the differences between districts became non-significant, while the risk of death remained higher in Pilate than Plaisance in the multivariate analysis including health-seeking behavior. This finding suggests that in Borgne and Pilate the mortality risk for diarrhea patients (which was more than twice that in Plaisance), could be explained mostly by the fact that these districts have more remote areas accessible only by foot. In addition to reducing the proportion of people seeking care, the delay to reach the treatment center probably also influenced the risk of dying as suggested by the higher, though not significantly, risk of dying in patients consulting more than 12 hours after onset of diarrhea in Guinea Bissau [11].
Considering their high vulnerability, it is important to improve response strategies for remote populations. Rapid implementation of ORPs in remote settings might be a good option. Here, only a small proportion (3%) of diarrhea cases reported using them. We did not investigate reasons for this low attendance, but their late implementation certainly did reduce their efficacy. Other factors such as low awareness of their purpose and location or lack of confidence in the quality of care provided could have participated. Early community involvement could probably improve all these aspects [13]. Mass vaccination campaigns have shown good efficacy to prevent cholera cases [14] and these would be particularly relevant in remote areas where other prevention or treatment strategies are difficult to sustain. Finally, these data illustrate the lack of adequate general health system in rural areas of Haiti, as well as in many other low and middle-income countries. Improving general access to care in these areas would probably be the best step towards reducing the high burden of cholera outbreaks as well as other diseases.
The main limitation of this retrospective survey may have been recall bias, particularly due to the long recall period. In contrast to mortality data, reporting of diarrhea is highly prone to recall bias in infants, and long recall periods are generally not recommended to assess diarrhea [15,16]. However, diarrhea in adults, particularly severe diarrhea, is rare and thus less prone to recall bias and we believe that it was not a major bias in the reported diarrhea cases. This belief is reinforced by the shape of the epidemic curve obtained through the survey, which is similar to those reported by the national surveillance system (http://mspp.gouv.ht). However, information bias might have had a more important impact on respondents’ report of the health structures visited and could have led in particular to an under-estimation of the number of visits to an ORP if patients also visited a higher-level health structure. Another limitation of our risk factor analysis was that it was a post-hoc analysis and we did not explore all risk factors, such as access to water and sanitation, socio-economic status, or access to health information.
In conclusion, we show here that attack rates and case fatality rates of the first cholera epidemic peak were much higher than reported by the national surveillance system, and that people living in very remote areas in the Nord department were particularly at risk for both disease and death during the early phase of the outbreak. Although an initial response focusing on urban and more densely populated areas was appropriate considering the large number of patients treated, this analysis shows that rural areas with poor access to health care and to cholera prevention and treatment information were at the greatest risk. Adapted strategies to rapidly provide access to preventive activities and treatment in remote communities are urgently needed to prevent this disproportionate impact in future cholera outbreaks.
|
10.1371/journal.pntd.0000293 | Impact of Long-Term Treatment with Ivermectin on the Prevalence and Intensity of Soil-Transmitted Helminth Infections | Control of soil-transmitted helminth (STH) infections relies on the periodic and long-term administration of anthelmintic drugs to high-risk groups, particularly school-age children living in endemic areas. There is limited data on the effectiveness of long-term periodic anthelmintic treatment on the prevalence of STHs, particularly from operational programmes. The current study investigated the impact of 15 to 17 years of treatment with the broad-spectrum anthelmintic ivermectin, used for the control of onchocerciasis, on STH prevalence and intensity in school-age and pre-school children.
A cross-sectional study was conducted in communities that had received annual or twice-annual ivermectin treatments and geographically adjacent communities that had not received treatment in two districts of Esmeraldas Province in Ecuador. Stool samples were collected from school-age children and examined for STH infection using the Kato-Katz and formol-ether concentration methods. Samples were collected also from pre-school children and examined by the formol-ether concentration method. Data on risk factors for STH infection were collected by parental questionnaire. We sampled a total of 3,705 school-age children (6–16 years) from 31 treated and 27 non-treated communities, and 1,701 pre-school children aged 0–5 years from 18 treated and 18 non-treated communities. Among school-age children, ivermectin treatment had significant effects on the prevalence (adjusted OR = 0.06, 95% CI 0.03–0.14) and intensity of Trichuris trichiura infection (adjusted RR = 0.28, 95% CI 0.11–0.70), but appeared to have no impact on Ascaris lumbricoides or hookworm infection. Reduced prevalence and intensities of T. trichiura infection were observed among children not eligible to receive ivermectina, providing some evidence of reduced transmission of T. trichiura infection in communities receiving mass ivermectin treatments.
Annual and twice-annual treatments with ivermectin over a period of up to 17 years may have had a significant impact on T. trichiura infection. The present data indicate that the long-term control of onchocerciasis with ivermectin may provide additional health benefits by reducing infections with trichuriasis. The addition of a second anthelmintic drug such as albendazole may be useful for a long-term effect on A. lumbricoides infection.
| Soil-transmitted helminth (intestinal worm) infections are very common in developing countries and are an important cause of illness. Mass de-worming treatments of school children are an important strategy to reduce illness caused by these infections in communities without access to clean water and sanitation. Few studies have examined the effect of repeated mass treatments in the long-term in controlling these infections. The objective of the present study was to assess the impact of the drug ivermectin used for the control of onchocerciasis (river blindness), that has important effects against intestinal worms, on the epidemiology of intestinal worms in children when administered repeatedly for 15–17 years. We compared the epidemiology of infections between children living in communities that received ivermectin with communities that never received the drug. The data suggest that ivermectin has important differential effects on intestinal worms with a greater impact on infections with Trichuris trichiura and little impact on Ascaris lumbricoides and hookworms infections. Our data suggest that long-term ivermectin treatments may provide health benefits through effects on T. trichiura infections but that the addition of second de-worming drug such as albendazole may be required for the control of other intestinal worm infections.
| Soil-transmitted helminths (STHs, also known as geohelminths or intestinal helminths) are important infectious diseases of humans and are estimated to infect over 2 billion humans worldwide [1], being particularly prevalent in poor populations living in tropical and sub-tropical regions of developing countries [2]. Soil-transmitted helminths include Ascaris lumbricoides, Trichuris trichiura, hookworm, and Strongyloides stercoralis, and are considered to be a major cause of morbidity related to impaired nutrition and poor childhood growth [3],[4].
STH control strategies are focused currently on mass treatment with broad spectrum anthelminthic drugs with the aim of reducing morbidity through reductions in parasite burdens that are strongly associated with risk of morbidity [5],[6]. Because of a high risk of re-infection in endemic areas, treatments have to be administered repeatedly for long periods of time. WHO has recommended the use of one of albendazole, mebendazole, levamisole and pyrantel for STH control [7]. The 4 drugs have variable efficacy against different STH parasites.
Ivermectin is a broad-spectrum anthelmintic drug used for the control of filarial infections including onchocerciasis and lymphatic filariasis [8],[9]. Ivermectin has also efficacy against STH infections - it is highly effective against ascariasis and strongyloidiasis [10],[11],[12],[13],[14] for which it has superior or comparable efficacy to albendazole [10],[12], is of moderate efficacy against trichuriasis for which it has comparable efficacy to albendazole [10],[11],[13],[14] but is less effective than albendazole against hookworm [12],[13],[14].
The distribution of ivermectin forms the basis of onchocerciasis control programmes worldwide. In Ecuador, the National Programme for the Elimination of Onchocerciasis has administered annual or twice-annual mass treatments with ivermectin over the past 17 year as the primary control strategy in endemic communities and has achieved high rates of coverage. We have reported previously that ivermectin control has been extremely effective for the control of onchocerciasis and may have eliminated the infection from some infection foci in Ecuador [15].
To investigate the long-term effect of ivermectin treatment used for the control of onchocerciasis on the prevalence of infection and transmission with STH parasites, we compared the prevalence and intensities of STH infections among children living in communities that had received long-term ivermectin treatments with those that had not received ivermectin in Esmeraldas Province in Ecuador. Previous surveys conducted in this area have shown high rates of STH infection [16],[17].
This study was conducted in rural Afro-Ecuadorian communities located in the principal onchocerciasis focus or in adjacent communities in the districts of Eloy Alfaro and San Lorenzo in Esmeraldas province, in northeastern Ecuador. The area is equatorial rain forest at an altitude of below 100 m above sea level and the characteristics of the study area and population are described in detail elsewhere [18].
A cross-sectional study was conducted to compare the prevalence and intensity of STH infections between children from treated and non-treated Afro-Ecuadorian communities. Selection of ivermectin-treated study communities was based on the 6-monthly treatment schedule with ivermectin of the Ecuadorian Elimination Programme for Onchocerciasis. Non-treated communities with a recent census were selected to be as similar as possible to treated communities with respect to geographic location (non-treated communities were geographically close to treated communities within the 2 study Districts), size (communities with schools having >250 pupils were excluded), and similar economic activities (e.g. subsistence agriculture, logging, and hunting). None of the non-treated communities had received previous mass treatments with ivermectin. STH infections were examined in two distinct age groups using the community census: 1) school-age (6–16 years) including children attending and not attending community schools; and 2) pre-school (0–5 years). These age groups separated approximately children eligible to receive ivermectin (>15 kg) from those not eligible, allowing an assessment of the effect of ivermectin treatments on the transmission of STH infections. Assessment of pre-school children was performed for a sample in the treated and non-treated communities.
The field study was conducted between March 2005 and May 2007. Stool collections were performed immediately before 6-monthly ivermectin distributions in treated communities. Annually-updated censuses were used to identify eligible children. A register of the number of treatments received by each child was obtained from the National Programme for the Elimination of Onchocerciasis. A questionnaire was administered to the parent or guardian of each child by trained field workers to obtain information about risk factors for STH infections including age, sex, mother's education level, material goods in the household, household income, household crowding, access to water, sanitation and recent anthelmintic treatments [18].
Single stool samples were collected from each child and examined for STH eggs and larvae using the modified Kato-Katz technique (quantification of Ascaris lumbricoides and Trichuris trichiura infection) [school-age children only] and formol-ethyl acetate concentration (detection of all STH infections including hookworm and Strongyloides stercoralis) [19] [all children].
The National Program for Elimination of Onchocerciasis has used community-based mass distribution of ivermectin as the main strategy for the control of Onchocerca volvulus infection since 1990. There are a total of 119 endemic communities with an estimated 19,420 inhabitants in 2 Provinces in Ecuador and ivermectin treatment was gradually introduced into all communities over the period 1990–1997. The number of treatments and the start of mass distribution vary according to each focus. Mass treatment was initiated in all communities included in this study between 1990 and 1992, and the period of mass ivermectin treatments for study communities ranged 15–17 years. Initially, ivermectin was distributed annually to endemic communities but treatments were increased to 6-monthly in all communities from 2001. Treatment coverage with ivermectin has been high and over the period 1990 to 2003, the annual average treatment coverage per treatment round with ivermectin was 85.2% (range 54.9%–97.9%). The mean treatment coverage for the two years, 2005–2006, was 98%. Eligibility criteria for treatment are: weight greater than 15 kg and free of serious illness (e.g. active tuberculosis, terminal cancer etc), and for women, not pregnant and not nursing infants up to 3 month of age. Ivermectin is distributed by primary health workers in each community at a dose of 150 µg/mg and treatments are directly observed. The population census is used as the basis for drug distribution [15].
STH prevalence was expressed as the percentage of subjects found positive for each parasite. Comparisons between proportions were performed using an adjusted chi-squared test for clustered data. ORs for prevalence data were calculated using multilevel logistic regression models (command, xtlogit) to obtain robust standard errors that take into account correlated data. Potential confounding factors were controlled for in the analysis. A multilevel analysis was used because of the large variation in STH prevalence between study communities. The intraclass correlation coefficients for Ascaris and Trichuris infection in this study were 0.29 (95% CI: 0.21–0.39) and 0.13 (95% CI: 0.09–0.18), respectively.
STH intensity was calculated as eggs per gram of stool. Because stool egg counts were over-dispersed (A. lumbricoides, mean 6,297 epg, variance 5.1×108; T. trichiura, mean 9746 epg, variance 1.3×107), a zero-inflated negative binomial (ZINB) model was used to fit the data. The Vuong test (1.90 (p = 0.03) and 11.55 (p<0.001) for A. lumbricoides and T. trichiura, respectively) indicated that a zero inflated model fit the data better than a standard negative binomial model [20]. A robust variance estimator was used to account for clustering by community, and rate ratios (RRs) and their 95% confidence intervals were estimated. The ZINB model was fitted manually, and potential confounding factors were controlled for in multivariate analyses. RRs for infection intensity indicate the relative increase or decrease in infection intensities in the treated compared to non-treated groups. A single 0.5 unit fall in RR represents a 50% greater infection intensity in the untreated compared to treated group. All analyses were done with STATA (version 9.0).
The study protocol was approved by ethical committee of the Hospital Pedro Vicente Maldonado, Pichincha Province, Ecuador. Informed written consent was obtained by a parent or guardian for all children. All children were offered treatment with a single dose of 400 mg of albendazole at the end of the study.
A total of 3,960 school-age children from 58 communities were assessed and 3,705 provided a stool sample and were included in the analysis. School-age children were analysed from 31 (1,752 children) ivermectin-treated and 27 (1,953 children) neighbouring non-treated communities. The mean cluster size was 54.7 (range: 13–203) children in treated and 69.2 (range: 15–190) in non-treated communities. Demographic, socioeconomic, and environmental characteristics of the children from treated and non-treated communities are shown in Table 1. There were differences in some socioeconomic and environmental characteristics between the two study groups: treated children tended to have a lower household income and fewer material goods and a greater proportion used river water than non-treated children.
The number of doses of ivermectin received by the 1,752 children in the treated communities were: 1–5 doses (12.5%), 6–10 doses (41.3%); 11–15 doses (41.0%) and 16–20 doses (5.2%). Treatment coverage with ivermectin was high and 79.3% (1,390) of children had received >75% of designated treatments over the previous five years. Reported treatments with other anthelmintic drugs were equally widespread in both treatment (77.5%) and non-treatment (78.0%) communities over the previous 6 months. Most treatments were bought directly by parents, were distributed through schools, or through physician consultations. During the course of this study there were no programmes of systematic periodic treatments with other anthelmintic drugs such as albendazole in any of the study communities. School-based treatments with albendazole have been given sporadically by non-governmental organisations, and politically-affiliated groups (e.g. at the time of local and presidential elections every 4 years).
Community treatments may have important effects on the transmission of STHs. To investigate this, we analysed stool samples from children aged 0–5 years that had not received ivermectin. A total of 776 and 925 children from 18 treated and 18 non-treated communities, respectively, were analysed. Of the 776 children aged 0–5 years from treated communities, 484 (62.8%) had not received any dose of ivermectin (for reasons of weight <15 kg), 139 (18.0%) had received 1–2 doses, 96 (12.5%) 3–4 doses, 42 (5.4%) 5–6 doses and 10 (1.3%) 7–8 doses.
Ivermectin treatment was associated with a significant reduction in the prevalence of infection with any STH parasite (62.8% treated vs. 86.3% untreated children; adj. OR 0.27, 95% CI 0.15–0.47, P<0.001) The age-prevalence and age-infection intensity profiles for A. lumbricoides and T. trichiura infections in treated and non-treated children are shown in Figure 1. There were trends of age-associated declines in the prevalence and intensity of both infections. The prevalence (treated 48.9% vs. non-treated 57.3%, adjusted OR 0.96, 95% CI 0.45–2.05, P = 0.92) and intensity (GM infection intensity, treated 30.2 vs. untreated 33.7 epg; adjusted RR 1.51, 95% CI 0.81–2.82, P = 0.19) of A. lumbricoides infection was not different between children from ivermectin-treated and non-treated communities across the age groupings (Figure 1A and B) (Table 2 and 3). Both the prevalence (treated 31.1% vs. non-treated 81.5%, adjusted OR 0.06, 95% CI 0.03–0.14, P<0.001) and intensity (GM infection intensity, treated 3.9 vs. untreated 132.0 epg; adjusted RR 0.28, 95% CI 0.11–0.70, P = 0.007) of T. trichiura infection were greater in non-treated compared to treated school children at all age groups (Figure 1C and D) (Table 2 and 3). Ivermectin treatment did not appear to reduce the prevalence of hookworm infection and, in fact, the prevalence was significantly greater in treated compared to non-treated children (treated 14.8% vs. non-treated 3.9%, adjusted OR 5.53, 95% CI 1.81–16.86, P = 0.003) (Table 2).
Among pre-school children living in the same communities and not eligible to receive ivermectin (in treated communities), there were significantly fewer infections with any STH (untreated 64.8% vs. treated 40.7%; adjusted for age, sex, and treatment, OR, 0.34, 95% CI 0.21–0.56, P<0.001) and T. trichiura infection (untreated 55.4% vs. treated 15.7%; adj. OR = 0.12, 95% CI 0.07–0.21, P<0.001)) but not A. lumbricoides infection (untreated 38.4% vs. treated 33.1%; adjusted OR = 0.67, 95% CI 0.38–1.19, P = 0.17) among treated compared to non-treated children. Pre-school children from the treated communities also had significantly lower infection intensities with T.trichiura (adj. RR = 0.32, 95% CI 0.19–0.54, P<0.001) but not with A. lumbricoides (adj. RR = 1.51, 95% CI 0.76–3.00, P = 0.24) compared to those from non-treated communities (Table 4 and 5). Because 32.9% of children in the treated group had received at least one dose of ivermectin, the analysis was repeated excluding all treated children (925 untreated vs. 471 treated children). The results of these analyses were similar to those obtained for all pre-school children (e.g. age and sex-adjusted OR for T. trichiura prevalence, OR 0.11, 95% CI 0.06–0.22, P<0.001).
At the community level there was a lot of variation in the prevalence of A. lumbricoides between treated and non-treated communities and no evidence of systematic differences in prevalence, but for T. trichiura infection, there was a strong trend for reduced prevalence in treated compared to non-treated communities (data not shown).
The control of soil-transmitted helminth (STH) infections is an important public health strategy targeted at school-age children in many developing countries, and has been prioritised because of the important morbid effects attributed to these infections. Further, such programmes are relatively easy to administer (through schools) and are considered to be highly cost-effective [21],[22]. Because STH infections are caused by faecal contamination of the environment, re-infections may be frequent after a single dose of anthelmintic treatment in highly endemic areas [23], and because treatment programmes rarely address underlying causes (e.g. poor sanitation), treatments may have to be administered periodically for periods of years. There are few reports of the long-term effects of anthelmintic treatment programmes on the epidemiology of STH infections; many studies, both randomized controlled trials [24] and uncontrolled studies [25], have reported the effect of single doses of anthelmintic drugs either alone or in combination with follow-up of up to 1 year post treatment [10]. Few studies have investigated the impact of periodic treatments with anthelmintic drugs on STH infections for longer periods–in such studies, follow-up has ranged 2 [26] and 4 years [27].
In the present study, we measured the impact of up to 17 years of periodic treatments with the broad-spectrum anthelmintic drug, ivermectin, on both the epidemiology of STH infections in school-age children and also the impact of such treatments on the transmission of STH infections in pre-school children. Our results provide evidence that long-term ivermectin treatments may have differential effects on STH infections, with major effects on T. trichiura, but little or no effect on A. lumbricoides and hookworm infections.
The present study had several important strengths. Ivermectin is a broad-spectrum anthelmintic drug that is effective against STH infections and has comparable efficacy (except for hookworm) to albendazole, the most widely used drug in STH control programmes [10],[11],[12]. Periodic treatment in this study was given for periods of between 15 and 17 years, making it, to our knowledge, the longest documented period of periodic treatments used for the evaluation of the impact of anthelmintic treatment on STH infections. Further, the data is from an operational control programme that was able to achieve high rates of treatment coverage, and the data obtained reflect, therefore, what an optimal control programme may be capable of achieving using a strategy of twice-annual treatments. Other strengths were the availability of data on the number of ivermectin treatments at the individual level, censuses that allowed us to identify all children of school age in each community, the collection of data on important confounding factors, and the study of a large sample both at the individual and community levels.
However, the study was cross-sectional rather than prospective and we did not have data on STH infections before the start of the intervention. We tried to select comparable non-treated communities, but there was evidence of differences in some variables between the treated and non-treated communities at baseline (Table 1). Although an analytic strategy allowing for these factors to be controlled in the analysis was used, residual confounding or systematic bias cannot be excluded. Systematic differences between communities may explain the higher prevalence of hookworm observed in treated communities. This observation is unlikely to be explained by the limited of efficacy of ivermectin against hookworm infections [12],[13],[28],[29]. The determinants of the geographic distribution of hookworm and other STH infections are poorly understood but include factors such as climate, socioeconomic factors and human behaviour [30]. We do not have data on climatic variables for these communities, and there were differences between treated communities with respect to some socioeconomic indicators (e.g. median income and material goods in the household) that may explain partly the geographic distribution of hookworm infection.
The study provided evidence that long-term periodic treatments with ivermectin may have important effects on the prevalence and intensity of T. trichiura infections. Previous studies have demonstrated variable efficacy of ivermectin against T. trichiura infections after one (35%–88% cure rate) [10],[11],[13],[14] or two doses (100% cure rate) [13] although studies from Africa showed very limited efficacy after 1–4 doses of ivermectin (0–11% cure rates) [12],[29],[31],[32],[33]. The present study in which communities had received ivermectin for 15–17 years provided some evidence for strong and significant long-term effects against T. trichiura infections among children receiving treatment (reduction in prevalence and intensity of 50.4% and 72%, respectively). Further, the observation of a reduced T. trichiura infection prevalence and intensity (39.7% and 68%, respectively) among children not eligible to receive treatment suggest that long-term ivermectin treatments may suppress the transmission of this infection.
Surprisingly, we did not observe an effect of long-term ivermectin administration on A. lumbricoides infections. Ivermectin is extremely effective against this parasite and single dose cure rates between different studies are consistently greater than 78% [10],[11],[12],[13],[14]. Studies that have examined the effects of ivermectin against ascariasis have shown that a single dose could reduce the prevalence and intensity of A. lumbricoides infection for up to three months after a single treatment [29],[34], although reinfections are an important problem and may occur by 3 months after treatment [29],[31],[35]. Previous studies that have examined the effects of multiple doses of ivermectin against A. lumbricoides infection for periods of 1–2 years have shown no important effect of treatments on the prevalence of A. lumbricoides infection [31],[33]. Further, in this study we did not observe an impact of ivermectin treatment on the prevalence or intensity of infection of A. lumbricoides among children not eligible to receive treatment, indicating little effect on the transmission of infection.
There are three possible explanations for the possible lack of long-term effect of ivermectin on ascariasis. Firstly, fertilized A. lumbricoides adult females are extremely fecund and each may produce in excess of 200,000 embryonated eggs per day [36], considerably more than T. trichiura or hookworm females. Ascaris eggs are extremely resistant to adverse environmental conditions and may remain infectious for years. Thus, in the absence of adequate removal of human faeces, exposure to embryonated eggs in the environment in endemic communities may be difficult to avoid and reinfection inevitable. A city-wide intervention to provide sanitation in the city of Salvador in Brazil had dramatic effects in reducing diarrhoeal incidence [37] and the prevalence of STH infections including ascariasis [38] that have been sustained over many years. A second but perhaps less likely explanation is the development of drug resistance by A. lumbricoides to ivermectin. Mass treatments with ivermectin have been administered in these communities for 15–17 years and the development of drug resistance by A. lumbricoides could explain the lack of impact of ivermectin on this infection. Decreased sensitivity of hookworm infections to mebendazole [39] and pyrantel [40] has been reported previously, and resistance to ivermectin by gastrointestinal nematodes of animals is widespread [41]. Finally, we do not have pre-treatment data on the prevalence of ascariasis in treated communities and it is possible that the prevalence of ascariasis was higher in treated compared to non-treated communities before the start of mass treatment with ivermectin between 1990 and 1992. Data obtained from the pre-treatment period for a group of 158 adults and children from a treated Afro-Ecuadorian community, not included in the present study, indicated a prevalence of A. lumbricoides and T. trichiura infections of 58.2% and 59.5%, respectively [17]. Although such data should be interpreted cautiously–because of differences in age, study population, and the diagnostic method used - they may suggest that the prevalence of ascariasis has not altered greatly since the start of mass treatment and that long-term ivermectin treatments have had a relatively greater effect in reducing the prevalence of trichuriasis than ascariasis.
We have evaluated the impact of long-term treatment with a broad-spectrum anthelmintic drug, ivermectin, used for the control of onchocerciasis, on the epidemiology of STH infections in rural Ecuador. The data indicate that 15–17 years of annual or twice-annual ivermectin treatments was highly effective against T.trichiura infections but may have had little impact on infections with A. lumbricoides and hookworm. Our study provides evidence that control programmes using ivermectin may provide additional health benefits by reducing the prevalence and intensity of infections with trichuriasis. To have a greater impact on ascariasis and hookworm infections, such programmes could consider the addition of twice annual albendazole treatments that can be safely administered with ivermectin [25], although effective reductions in the prevalence of these infections may require more frequent periodic treatments. The combination of ivermectin and albendazole has the advantage also of greater efficacy against trichuriasis than either drug alone [10],[11],[42]. However, long-term sustainable control may require interventions that target the underlying causes of these infections–namely the unsafe disposal and use of human faeces.
|
10.1371/journal.ppat.1002558 | LEDGF/p75-Independent HIV-1 Replication Demonstrates a Role for HRP-2 and Remains Sensitive to Inhibition by LEDGINs | Lens epithelium–derived growth factor (LEDGF/p75) is a cellular cofactor of HIV-1 integrase (IN) that interacts with IN through its IN binding domain (IBD) and tethers the viral pre-integration complex to the host cell chromatin. Here we report the generation of a human somatic LEDGF/p75 knockout cell line that allows the study of spreading HIV-1 infection in the absence of LEDGF/p75. By homologous recombination the exons encoding the LEDGF/p75 IBD (exons 11 to 14) were knocked out. In the absence of LEDGF/p75 replication of laboratory HIV-1 strains was severely delayed while clinical HIV-1 isolates were replication-defective. The residual replication was predominantly mediated by the Hepatoma-derived growth factor related protein 2 (HRP-2), the only cellular protein besides LEDGF/p75 that contains an IBD. Importantly, the recently described IN-LEDGF/p75 inhibitors (LEDGINs) remained active even in the absence of LEDGF/p75 by blocking the interaction with the IBD of HRP-2. These results further support the potential of LEDGINs as allosteric integrase inhibitors.
| Like other viruses, HIV has a limited genome and needs to exploit the machinery of the host cell to complete its replication cycle. The elucidation of virus-host interactions not only sheds light on pathogenesis but also provides opportunities in a limited number of cases to develop novel antiviral drugs. A prototypical example is the interaction between the cellular protein LEDGF/p75 and HIV-1 integrase (IN). Here we generated a human somatic LEDGF/p75 knockout cell line to demonstrate that HIV-1 replication is highly dependent on its cofactor. We show that the residual replication of laboratory strains is predominantly mediated by a LEDGF/p75-related protein, HRP-2. Interestingly, the recently developed HIV-1 IN inhibitors that target the LEDGF/p75-IN interaction interface, LEDGINs, remain active even in the absence of LEDGF/p75. We demonstrate that LEDGINs efficiently block the interaction between IN and HRP-2. In case HIV-1 would be able to bypass LEDGF/p75-dependent replication using HRP-2 as an alternative tether, LEDGINs would remain fully active.
| Integration of viral DNA into the host cell genome is a critical step during HIV replication. A stably inserted provirus is essential for productive infection and archives the genetic information of HIV in the host cell. The presence of a permanent viral reservoir that evades the immune system and enables HIV to rebound once antiretroviral drugs are withdrawn is one of the major remaining hurdles to surmount the HIV epidemic.
Lentiviral integration is catalyzed by the viral enzyme IN in close association with the cellular cofactor LEDGF/p75 [1]–[7]. LEDGF is encoded by the PSIP1 gene, which generates the splice variants LEDGF/p52 and LEDGF/p75 [8]. Both share an N-terminal region of 325 residues containing an ensemble of chromatin binding elements, such as the PWWP and AT hook domain, yet differ at the C-terminus. LEDGF/p52 contains 8 amino acids at its C-terminus [9] and fails to interact with HIV-1 IN [10], [11], whereas LEDGF/p75 contains an IBD (aa 347–429) capable of interacting with lentiviral IN [3], [12], [13]. The cofactor tethers IN to the host cell chromatin, protects it from proteolytic degradation, stimulates its enzymatic activity in vitro and in living cells [1], [10], [13]–[16] and determines HIV-1 integration site distribution [2], [11], [17], [18].
The role of LEDGF/p75 in HIV-1 replication was studied using RNA interference (RNAi) targeting LEDGF/p75 or using LEDGF KO murine embryonic fibroblasts (MEF) [2], [5], [6], [11], [17], [19], [20]. Although both strategies point to a key role for LEDGF/p75 in lentiviral replication, they resulted in somewhat conflicting conclusions. Potent RNAi-mediated knockdown (KD) of LEDGF/p75 reduced HIV-1 replication, yet residual replication was observed [5], [6], [20], which was attributed to imperfect RNAi-mediated KD of LEDGF/p75, with minute amounts of LEDGF/p75 being sufficient to support HIV-1 replication [5], [6]. Whether LEDGF/p75 is essential for HIV-1 replication or not could not be addressed by this approach. Later, two LEDGF KO mice were generated. Since mouse cells are not permissive to spreading HIV-1 infection, HIV-based viral vectors were used. The first effort resulted in mouse LEDGF KO clones following insertion of a gene trap [21]. Data obtained from MEFs isolated from these embryos indicated a strong yet incomplete block in integration of HIV-based lentiviral vectors (LV) [17]. Next, a Cre-conditional LEDGF KO mouse was generated. Challenge of the KO MEFs with LV resulted in reduced but not annihilated reporter gene expression [11]. Although analysis was restricted to single round assays, both studies suggest LEDGF/p75 not to be essential for HIV-1 replication, with the cofactor being involved in integration site selection rather than in promoting integration. Here we present the generation of the first human somatic LEDGF/p75 KO cell line to finally answer the question whether LEDGF/p75 is required for spreading infection of various HIV strains.
Besides LEDGF/p75, a second member of the hepatoma-derived growth factor related protein family [22], Hepatoma-derived growth factor related protein 2 (HRP-2), was shown to interact with HIV-1 IN [12]. Although HRP-2 overexpression relocated IN from the cytoplasm to the nucleus in LEDGF/p75-depleted cells [23], the IN–HRP-2 interaction was weaker than the IN-LEDGF/p75 interaction [12]. Neither transient [20], [24] nor stable HRP-2 KD [6] reduced HIV-1 replication even after reduction of LEDGF/p75, suggesting that HRP-2 is not involved in HIV replication. However, it has not been excluded that in the absence of LEDGF/p75 HRP-2 can function as an alternative molecular tether of HIV integration.
Allosteric HIV-1 IN inhibitors that target the LEDGF/p75-IN interaction interface (LEDGINs) and potently block HIV-1 replication [25] are in preclinical development. The existence of alternative cellular cofactors, such as HRP-2, or alternative escape routes might hamper the clinical development of this class of compounds. To answer these questions, we have generated a human somatic LEDGF/p75 KO cell line. We demonstrate that laboratory-adapted HIV strains are capable of replicating in the absence of LEDGF/p75 but show a drastic replication defect. We show that this residual replication in the absence of LEDGF/p75 is predominantly mediated by HRP-2. Finally, we demonstrate that LEDGINs remained fully active even in the absence of LEDGF/p75 corroborating their allosteric mechanism of action.
To clarify the role of LEDGF/p75 during spreading HIV-1 infection, we generated a human somatic KO in Nalm-6 cells, a human pre-B acute lymphoblastic leukemia cell line [26], [27]. We eliminated the LEDGF/p75 isoform while preserving the LEDGF/p52 splice variant. Deletion of exon 11 to 14 in the PSIP1 gene fuses exon 10 to exon 15 resulting in a frame shift that yields a truncated LEDGF/p75 in which the C-terminus, including the IBD (aa 326–530) is replaced by a 9 aa tail (Figure S1A, referred to as LEDGFKO). Targeting plasmids were designed carrying the genomic flanking regions of LEDGF/p75 exon 11 and 14, interspersed with a floxed selection cassette (Figure 1A). Following transfection of wild-type Nalm-6 cells (Nalm+/+) with the first targeting plasmid and subsequent selection, three heterozygous clones (cl) (denoted as Nalm+/c; cl 31, cl 97 and cl 147, respectively) were obtained (Figure 1B). We continued with Nalm+/c cl 31. Transfection of Nalm+/c cl 31 with the second targeting plasmid resulted in the selection of a homozygous KO clone carrying both resistance cassettes (Nalmc/c 31 cl 73). Selection cassettes were removed by Cre-mediated excision, resulting in seven LEDGF/p75 KO clones, referred to as Nalm−/− cl 1-7.
Correct homologous recombination of the genomic region was verified via genomic PCR (Figure 1C), Southern blot analysis (Figure 1D) and sequencing of the genomic and mRNA region (Figure S1A). The absence of wild-type LEDGF/p75 in the KO cells was corroborated by RT-PCR (Figure S1B and S1C), qRT-PCR (Figure S1D) and Western blot analysis (Figure 1E, arrow). A band of 52 kDa appears in the Nalm+/c and Nalm−/− cell lines; it corresponds to the truncated form, LEDGFKO (Figure 1E, arrowhead), and is absent in wild-type cells. Throughout the manuscript Nalm−/− cl 1 and cl 2 monoclonal cell lines are used. Wild-type Nalm-6 cells, referred to as Nalm+/+, were used as controls, next to Nalm+/c cl 31, referred to as Nalm+/c, the closest clonal ancestor of the Nalm−/− cells.
We first evaluated whether the LEDGF/p75 KO cells (Nalm−/−) support transduction by a single round HIV-based viral vector. We challenged the abovementioned engineered cell lines with a VSV-G pseudotyped HIV reporter virus encoding firefly luciferase under control of the viral long terminal repeat promoter (HIV-fLuc). Transduction efficiency (RLU/µg protein) was 6.7-fold lower in Nalm−/− cells (cl 1 and cl 2) compared to control Nalm+/+ and Nalm+/c cells (Figure 1F) (15±3.7% residual reporter activity; n = 10). Quantitative PCR revealed 2.4-fold lower integrated copies comparing Nalm−/− with Nalm+/c (Figure 1G), whereas late RT products (Figure S1E) and 2-LTR circles remained unaffected (Figure S1F). Together these data indicate a block between reverse transcription and integration.
Since LEDGF/p75 determines lentiviral integration site selection, we analyzed the distribution of HIV-1 integration sites in the absence of LEDGF/p75. A total of 2535 HIV-1 integration sites were obtained in Nalm-6 cells of which 799 in Nalm−/− (Table 1). Random control sites were generated computationally and matched to experimental sites with respect to the distance to the nearest MseI cleavage site (matched random control, MRC) [2]. LEDGF/p75 KO significantly reduced the preference of HIV-1 to integrate in RefSeq genes (P<0.0001 for comparison of Nalm−/− cl 1 or 2 with Nalm+/+ or Nalm+/c) and instead, a preference for CpG islands (P<0.05 for comparison of Nalm−/− cl 1 or 2 with Nalm+/+ or Nalm+/c and P<0.0001 for pooled comparison) emerged (Figure 1H and Table 1). Similar results were obtained using the Ensembl and UniGene annotation (Figure S1G and S1H). HIV-1 integration events in RefSeq genes remained nevertheless significantly favored over MRC in the KO cells (P<0.0001). The target DNA consensus proved to be LEDGF/p75 independent (compare Figure S1I with S1J). The consensus sequence for the different cell lines was similar to that determined previously [28]–[30].
In human LEDGF/p75 KD cells HIV-1 replication is hampered, but not completely blocked which can be attributed to the remaining minute amounts of LEDGF/p75 [5], [6], [20]. Although single round viral vector transduction was severely reduced in LEDGF KO MEFs [11], [17], [21], spreading HIV-1 infection in the absence of LEDGF/p75 could not be tested. To test HIV-1 replication, we introduced the CD4 receptor into the Nalm-6 cells that express CXCR4 [31], a co-receptor for HIV-1 replication. All selected transgenic cell lines (Nalm+/+, Nalm+/c and Nalm−/− cl 1 and cl 2) showed similar growth rates (Figure S6A and S6C) and CD4 and CXCR4 expression levels (Figure S6D and S6E). We then challenged the respective cell lines with the laboratory strain HIVNL4.3 (Figure 2A). Both Nalm+/+ (Figure S2A) and Nalm+/c cells supported viral replication to the same extent (Figure 2A). Peak viral replication was consistently observed between day 7 and 9 post infection depending on the multiplicity of infection (MOI; compare MOI 0.5 and 0.1 in Figure 2A). In Nalm−/− cells infected with HIVNL4.3, low-level p24 production was observed, eventually leading to a breakthrough albeit after a lag-period of 14 to 18 days compared to control cells (Figure 2A, n = 6, a representative experiment is shown). Comparable data showing this delay were obtained with another laboratory strain, HIVIIIb (data not shown).
Next, we challenged the different cell lines with two clinical isolates of HIV-1 (93TH053, denoted as #1 and 96USSN20 [32], denoted as #2). Viral breakthrough was observed 17 to 20 days post infection in the control cell line (Figure 2B and 2C). In the first two weeks after infection of the KO cell lines only a discrete increase in p24 was observed; at 35 days after infection p24 levels were below detection limit (Figure 2B and 2C).
We next evaluated whether the rise in p24 titers observed in Nalm−/− cells after challenge with laboratory HIV-1 strains could be explained by virus release from cells infected in the first round, rather than ongoing replication cycles. Therefore we challenged Nalm+/c and Nalm−/− cells with HIVNL4.3 and resuspended the cells at 8 hrs post infection (Figure S2B) in fresh medium containing either zidovudine (AZT), ritonavir (RIT) or no inhibitor. AZT, a reverse transcriptase inhibitor, blocks infection of new cells but allows monitoring of virus release from already infected cells whereas RIT, a protease inhibitor, blocks processing of GAG-precursor processing in the virus released from infected cells. In Nalm−/− cells as well as in control Nalm+/c cells the p24 production clearly decreased in the presence RIT or AZT. The decrease in p24 in Nalm+/c without inhibitor at day 6 was due to the cytopathic effect of the virus. This indicates that the p24 increase observed in Nalm−/− cells results from spreading infection and not solely from virus release from cells infected in the first round.
The observed delay in multiple round HIV-1 replication in the absence of LEDGF/p75 was further analyzed by quantification of the different HIV-1 DNA species at different time points after infection. Late RT products at 10 hrs post infection and 2-LTR circles at 24 hrs post infection were comparable in Nalm+/c and Nalm−/− cells (Figure S3A and S3B). Addition of the IN strand transfer inhibitor (INSTI) raltegravir (RAL) in Nalm+/c and Nalm−/− cell lines resulted in a comparable increase in 2-LTR circles at 24 hrs post infection. The number of integrated proviral copies (Alu-qPCR, Figure S3C) was severely reduced in the presence of RAL. In Nalm−/− a reduction in the number of integrants was detected after 24 and 48 hrs compared to Nalm+/c cell lines.
We next characterized the virus harvested from Nalm−/− at day 18 after infection with the laboratory strain HIVNL4.3 (referred to as HIV−/−). Challenging Nalm+/c cells with this virus demonstrated that HIV−/− is replication competent (Figure S2C, right panel, HIV−/− on Nalm+/c). In addition, we evaluated whether HIV−/− virus was phenotypically adapted to the absence of LEDGF/p75. HIV−/− replication remained impaired in Nalm−/− compared to Nalm+/c cells (Figure S2C, right panel). The proviral IN sequence of HIV−/− was unaltered compared with the consensus sequence of HIVNL4.3 (data not shown). Control HIV harvested from Nalm+/c cells (denoted as HIV+/c) demonstrated a phenotype that was comparable to that of HIVNL4.3 (Figure S2C, left panel). Serial passaging (N = 10) of HIV-1 on LEDGF/p75 KO cells did not result in phenotypic adaptation or changes in the proviral IN sequence (data not shown).
Although residual HIV-1 replication in KO cells was only detectable after infection with laboratory strains, we performed additional experiments to understand this phenotype. Residual viral replication in the absence of LEDGF/p75 can either be explained by cofactor independent replication, or by the presence of a second cofactor that substitutes for LEDGF/p75. Like LEDGF/p75, HRP-2 also harbors a PWWP-domain and an IBD-like domain shown to interact with HIV-1 IN in vitro [12]. In order to determine whether HRP-2 can act as an alternative co-factor for HIV integration, we targeted the HRP-2 mRNA using miRNA-based short hairpins (miR HRP-2). As controls we employed a vector lacking the miRNA expression cassette (denoted as control) (Figure S7B). We generated stable HRP-2 KD cells, termed Nalm+/+miR HRP-2, Nalm+/cmiR HRP-2 and Nalm−/−miR HRP-2 and matched controls Nalm+/+control, Nalm+/ccontrol and Nalm−/−control. HRP-2 KD cells showed 65, 75 and 80% depletion of HRP-2, respectively, as determined by qPCR (Figure 3A). No effect on cellular growth kinetics was observed (data not shown). Upon single round transduction with HIV-fLuc no difference was observed in Nalm+/c cells with or without HRP-2 KD (Figure 3B, left panel), whereas luciferase activity was reduced 5-fold in the Nalm−/−control cell line (20.0±1.5%, n = 3) due to LEDGF/p75 KO. An additional 2.4-fold reduction was observed in Nalm−/−miR HRP-2 when compared to Nalm−/−control (8.4±0.6%, n = 3) (Figure 3B, left panel) that correlated with a 2-fold reduction in integrated copies (Figure 3B, right panel).
We next challenged these cells with the laboratory strain HIVNL4.3 at different MOI (Figure 3C–E). In the control Nalm+/+ and Nalm+/c cell lines, we observed a minor reduction in viral replication upon HRP-2 KD but only at lower MOI (compare Figure 3C and 3D, E). However, HRP-2 KD in LEDGF/p75 KO cells additionally inhibited HIV-1 replication 2- to 3-fold compared to control cells (Figure 3C–E, compare Nalm−/−control and Nalm−/−miR HRP-2, detail panel). We generated a second LEDGF/p75 KO HRP-2 KD cell line to corroborate our results. Single round transduction with HIV-fLuc resulted in an additional 4.7-fold reduction of luciferase reporter activity when compared with LEDGF KO cells (Figure S5F), whereas HIVNL4.3 replication was affected 10-fold at day 8 post infection when comparing LEDGF/p75 KO and LEDGF/p75 KO HRP-2 KD cells (compare Figure S5D with S5E, condition without compounds). To corroborate that additional KD of HRP-2 results in an increased block of integration in LEDGF/p75 KO cells, we analyzed the number of integrated viral copies at 24 hrs and at 5 days post infection, the latter in the presence of RIT (Figure 3F and 3G, respectively). A 2-fold drop in proviral copies upon HRP-2 KD was observed.
To extend our findings in LEDGF/p75 KO cells, we tested whether HRP-2 KD resulted in additional reduction of viral replication in LEDGF/p75 KD HeLaP4 (Figure S4), PM1 (Figure 4A–C) and SupT1 (Figure 4D–F) cell lines. First, wild-type HeLaP4 (wild-type) and LEDGF/p75 KD (miR LEDGF) cells [18] were transduced with miR HRP-2 or miR control vectors, the latter containing a miRNA-hairpin directed against monomeric red fluorescent protein (DsRed) mRNA [33] (Figure S7C). Following zeocin selection, single HRP-2 KD (wild-type/miR HRP-2) and double KD (miR LEDGF/miR HRP-2) cells showed 20–25% of residual HRP-2 mRNA levels compared to the control cell lines (wild-type, wild-type/miR control and miR LEDGF/miR control cells) as determined by qPCR (Figure S4A and S4C). Loss of HRP-2 protein was corroborated by Western blot analysis and immunocytochemistry (data not shown). Of note, LEDGF/p75 levels remained unaffected upon additional HRP-2 KD (data not shown) and growth rates of the respective cell lines were comparable (Figure S6B and S6C). KD of HRP-2 in wild-type HeLaP4 cells did not affect multiple round HIV-1 replication (Figure S4B), confirming previous findings by Llano et al. [6]. LEDGF/p75 KD on the other hand reduced HIV-fLuc transduction 5-fold (luciferase reporter activity = 19.2±3.5% of wild-type) (Figure S4D). Additional KD of HRP-2 in LEDGF/p75-depleted cells diminished HIV-fLuc reporter activity an additional 3-fold, to 6.3±2% of control cells (miR LEDGF/miR control) (Figure S4D). This reduction was accompanied with a 2-fold decrease in the number of integrated copies (Figure S4E). Transfection of the cell lines with the plasmid encoding HIV-fLuc (pHIV-fLuc) did not demonstrate any difference (Figure S4F), ruling out transcriptional effects upon HRP-2 KD. Next, we infected double KD (miR LEDGF/miR HRP-2) cells and control (miR LEDGF/miR control) cells together with wild-type and LEDGF/p75 back-complemented (LEDGF BC) cells with the laboratory strain HIVNL4.3 (Figure S4G). Viral replication was inhibited in miR LEDGF cells and rescued upon LEDGF/p75 back-complementation (Figure S4G, compare wild-type and LEDGF BC). Additional KD of HRP-2 in LEDGF/p75 depleted cells (miR LEDGF/miR HRP-2) inhibited viral replication more than LEDGF/p75 KD alone (miR LEDGF/miR control). The latter demonstrated a breakthrough around day 30 post infection (Figure S4G, open diamonds), whereas cells with double KD did not demonstrate viral breakthrough (Figure S4G, open squares). Analysis was ended at 48 days post infection. Comparable data were obtained in HeLaP4 cell lines generated with other LV constructs (Figure S7B and S7D) using hygromycin B selection or eGFP sorting (data not shown). The additional block of HIV-1 replication upon HRP-2 KD in LEDGF/p75 depleted cell lines was also measured by quantifying the number of integrated proviral copies. At day 39, 45 and 48 post infection the number of integrated copies was low in double KD (miR LEDGF/miR HRP-2) cells compared to the control LEDGF/p75 KD (miR LEDGF/miR control) cells (Figure S4H) with proviruses numbering 0.032 (±0.012) and 0.038 (±0.012) per RNaseP genomic copy on day 39 and 48 respectively, compared to 1.39 (±0.18) and 0.79 (±0.23) in the control LEDGF/p75 KD cell lines. In addition, we quantified different HIV-1 DNA species at different time points post infection in wild-type, LEDGF/p75 KD (miR LEDGF/miR control) and double KD cells (miR LEDGF/miR HRP-2). We observed no difference in late RT products at 10 hrs post infection (Figure S4I). The number of 2-LTR circles in LEDGF/p75 KD (miR LEDGF/miR control) and both LEDGF/p75 and HRP-2 KD (miR LEDGF/miR HRP-2) cells was elevated compared to wild-type cells (Figure S4J). Together with the data in the LEDGF/p75 KO cells, these data indicate that HRP-2 KD blocks HIV-1 at a step between reverse transcription and integration but only after potent depletion of LEDGF/p75.
Next, we expanded our findings to relevant T-cell lines, PM1 and SupT1. We generated cell lines with stable KD of LEDGF/p75, HRP-2 or both, together with their respective controls (constructs shown in Figure S7B). For PM1 cells KD efficiency was 85–92% for LEDGF/p75 (Figure 4A) and 79–81% for HRP-2 (Figure 4B), for SupT1 cells it amounted to 81–88% for LEDGF/p75 (Figure 4D) and 75–80% for HRP-2 (Figure 4E). In both cell lines HRP-2 KD alone did not affect HIV-1 replication, whereas a clear reduction in HIV-1 replication was observed upon LEDGF/p75 KD (Figure 4C and 4F, left panel, for PM1 and SupT1 respectively). Consistent with our findings in LEDGF/p75 KO cells and LEDGF/p75 depleted HeLaP4 cells, also in PM1 and SupT1 cells, HRP-2 KD in LEDGF/p75 depleted cells further hampered HIV-1 replication (Figure 4C and 4F, detail panels, for PM1 and SupT1 respectively).
Recently, we reported a new class of antiretrovirals termed LEDGINs that bind to the LEDGF/p75 binding pocket of HIV-1 IN and block HIV-1 integration and replication in cell culture [25]. We assayed their activity in the LEDGF/p75 KO cells. We challenged Nalm+/+ and Nalm+/c cells together with Nalm−/− cells with the laboratory strain HIVIIIb in the presence of different concentrations of LEDGIN 7 [25]. LEDGIN 7 blocked HIV-1 replication in all cell lines in a concentration dependent manner (Figure 5A and 5C). Similar data were obtained with the laboratory strain HIVNL4.3 (Figure S5A). The toxicity profile in Nalm-6 cells corresponded to that elaborated previously in MT4 cells [25]. No significant toxicity was observed in the concentrations used (data not shown). Of note, LEDGINs were also active against HIV harvested from LEDGF/p75 KO cells (HIV−/−, data not shown). RAL served as a positive control, demonstrating equal inhibition of HIV-1 replication in the different cell lines (Figure 5B and 5D). Dose response curves (Figure 5E and 5F) enabled determination of IC50 values, listed in Table S1.
We have shown that residual replication of HIV-1 laboratory strains in LEDGF/p75 KO cells is predominantly mediated by HRP-2 and that LEDGINs block residual HIV-1 replication in KO cells. This can be explained by allosteric inhibition of LEDGINs or by the fact that binding of LEDGINs to the IN-surface also impedes the interaction with HRP-2 or a combination of both. We evaluated whether LEDGINs inhibit the HRP-2-IN interaction in an AlphaScreen assay. Since IN binds HRP-2 via its IBD (aa 470–593) [12] in vitro, we measured the interaction between recombinant HIV-1 IN and the C-terminal part of HRP-2 (aa 448–670). We generated maltose binding protein (MBP) tagged fusions containing either the C-terminal end of LEDGF/p75 (aa 325–530) or HRP-2 (aa 448–670). These recombinant proteins, MBP-LEDGF/p75325–530 and MBP-HRP-2448–670, bound to His6-IN with apparent KD's of 6.6 nM (±4.6 nM) or 89.8 nM (±18.1 nM), respectively (Figure 6A). In line with previous observations [25], LEDGINs inhibited the IN-LEDGF325–530 interaction (Figure 6B; IC50 = 2.60±0.99 µM). LEDGINs also inhibited the IN-HRP-2448–670 interaction, albeit with a 10-fold lower IC50 (Figure 6B; IC50 = 0.23±0.14 µM). This 10-fold increased potency for LEDGIN 7 to block interaction of IN with MBP-HRP-2448–670 compared to MBP-LEDGF325–530 correlates well with the 13-fold lower affinity of MBP-HRP-2448–670 for IN, as shown in Figure 6A.
Next, we evaluated whether LEDGINs remain active in LEDGF/p75 KO HRP-2 KD cells. The residual HIV-1 replication was sensitive to inhibition by LEDGINs (Figure S5E).
Since the identification of LEDGF/p75 as a binding partner of HIV-1 IN in 2003 [1], we and other groups have demonstrated its importance for HIV-1 replication [3]–[7], [10], [11], [34], [35]. Our current understanding of the mechanism of action proposes LEDGF/p75 to act as a molecular tether between the lentiviral preintegration complex and the host cell chromatin; the chromatin reading capacity of LEDGF/p75 thereby determines integration site distribution [2], [11], [17], [18]. Given the methodological restrictions associated with the RNAi and mouse KO studies of the past, we decided to investigate the role of LEDGF/p75 in HIV-1 replication by generating a human somatic LEDGF/p75 KO cell line. A second rationale for this study follows the recent development of LEDGINs, small molecules that efficiently target the interaction between HIV-1 IN and LEDGF/p75 by interaction with the LEDGF/p75 binding pocket in HIV-1 IN [25]. Since LEDGINs block HIV-1 replication, the interest in the question whether or not LEDGF/p75 is essential for viral replication was revived.
Our studies demonstrate that residual HIV-1 replication in LEDGF/p75 KO cells can be observed using laboratory-adapted HIV-1 strains (Figure 2A). These observations are reminiscent to data obtained in LEDGF/p75 KD cell lines [5], [6], [20], although important differences can be noticed. First, when clinical HIV-1 isolates were used, we observed sterilizing infections in LEDGF/p75 KO cells (Figure 2B and 2C). Sterilizing infection has never been reported with RNAi mediated LEDGF/p75 KD. Although the effect might be in part explained by a lower infectivity of these clinical isolates, it emphasizes the importance of LEDGF/p75 for HIV-1 replication. In addition, LEDGF/p75 KO results in a more pronounced shift of HIV-1 integration out of RefSeq genes when compared to control cells (25.7% difference when comparing LEDGF/p75 KO to control cells; Table S3, see column 8), whereas integration in LEDGF/p75 KD cells was only moderately affected (1.6–8.4% compared to control cells, Table S3, see column 8) [2].
A next application of our KO cell line was the investigation of the role of HRP-2 in HIV-1 replication. The cellular function of HRP-2 is currently unknown. Like LEDGF/p75, HRP-2 contains a PWWP domain at its N-terminus [12], [22], [36], [37] and a basic C-terminus, that harbors an IBD-like domain. GST pull-downs showed that the homologous IBD region in HRP-2 (amino acids 470–593) interacts with IN [12]. Vanegas and colleagues reported earlier that HRP-2 overexpression relocated HIV-1 IN from the cytoplasm to the nucleus in LEDGF/p75 depleted cells [23]. Although HRP-2 was investigated previously as a potential alternative for LEDGF/p75, no effect in multiple round HIV-1 replication was observed after HRP-2 KD alone or in combination with LEDGF/p75 KD [6], [20], [24]. However, these observations may have been obscured by the remaining LEDGF/p75 after incomplete RNAi mediated KD. Therefore we revisited the mechanism of residual replication of HIV-1 laboratory strains in LEDGF/p75 KO cell lines. We demonstrate that both single round transduction and multiple round replication is additionally hampered upon HRP-2 KD in LEDGF/p75 KO cells. HIV-1 engages HRP-2 as an alternative for LEDGF/p75, but this low affinity IN binding partner (Figure 6A) can only substitute for LEDGF/p75 after depletion of the latter (Figure 3, 4 and S4), suggesting a dominant role for LEDGF/p75 over HRP-2. Several reasons can be proposed. Cherepanov et al. [12] demonstrated that considerably less IN could be co-immunoprecipitated by HRP-2 than LEDGF/p75, implying that the IN–HRP-2 interaction is weaker than the IN-LEDGF/p75 interaction. In line with these observations, Vanegas et al. reported that Flag-LEDGF/p75 but not Flag-HRP-2 co-immunoprecipitated IN from cell lysates [23]. Here we demonstrate using AlphaScreen technology that the IBD containing C-terminal end of HRP-2 has an approximately 13-fold lower affinity for HIV-1 IN than the corresponding part in LEDGF/p75 (Figure 6A). Next, LEDGF/p75 demonstrates a speckled nuclear localization pattern and binds to mitotic chromatin. Vanegas et al. demonstrated that contrary to LEDGF/p75, HRP-2 does not bind to mitotic chromatin [23] questioning its role as a chromatin-tethering molecule. However, since LEDGF/p75 KD also affects viral replication in non-dividing macrophages [20], the binding capacity of LEDGF/p75 to condensed mitotic chromatin might not be relevant for HIV-1 replication.
The preference of HIV-1 to integration in genes [38] is reduced upon LEDGF/p75 KO corroborating previous observations in LEDGF/p75 KD cells [2], [11], [17], [18] and underscoring LEDGF/p75 as the major targeting factor for HIV-1 integration. In line with this tethering role for LEDGF/p75, chimeras carrying alternative chromatin binding motifs fused to IBD could retarget HIV-1 integration [18], [39], [40]. In addition, De Rijck et al. [41] demonstrated that the LEDGF/p75 chromatin binding mirrors HIV-1 integration site distribution. HIV-1 integration in RefSeq genes remained significantly different from MRC throughout (P<0.0001) and more directed towards CpG islands in LEDGF/p75 KO cells. Both observations support the idea of an alternative targeting mechanism for HIV-1 acting in the absence of LEDGF/p75. Since additional HRP-2 KD resulted in an additional 2-fold reduction in integrated copies compared to LEDGF/p75 depletion, HRP-2 is a candidate. The integration site distribution pattern of HIV-1 derived vectors remained unaltered after additional HRP-2 KD in LEDGF/p75 KD HEK293T cells [2], but LEDGF/p75 depletion may have been insufficient in those experiments.
Apart from LEDGF/p75 and HRP-2, no other human protein contains a PWWP-domain in conjunction with an IBD. However, other proteins or protein complexes could take over the tethering activity in the absence of LEDGF/p75 and HRP-2 by combining an IBD-like domain with a chromatin-binding function. The IBD belongs to a family exemplified by the Transcription Factor IIS (TFIIS) N-terminal domain (InterPro IPR017923 TFIIS_N) ([12] and based on an updated search using the HHpred algorithm [42], [43]). Sequence comparison of the respective predicted IN-binding loops of these domains, suggests it is however unlikely that IN binds to these IBD-like proteins as it does to the IBD of LEDGF/p75 or HRP-2 (data not shown). Therefore the residual HIV-1 replication observed in the LEDGF/p75 KO HRP-2 KD cells may 1) still be HRP-2 mediated since the KD of HRP-2 is not complete, 2) be mediated by an unknown third cellular cofactor or complex, or 3) occur independently from cellular cofactors.
The question remains whether HRP-2 is of any importance for HIV infection in patients? The HRP-2 phenotype only becomes evident in vitro using laboratory strains and upon strong depletion or KO of LEDGF/p75. Taking into account the lower affinity of HRP-2 for HIV-1 IN, interaction likely only takes place in the complete absence of LEDGF/p75. The LEDGF/p75IBD is highly conserved within humans and across species [12]. Only a few SNPs have been identified [44]. Although relative LEDGF/p75 and HRP-2 expression levels still need to be verified in relevant human cells, to date there is no evidence for LEDGF/p75 depletion in humans and a substituting role of HRP-2 in HIV-1 infection.
Previous reports demonstrated a moderate increase in 2-LTR circles upon LEDGF/p75 KD [5], [6], whereas 2-LTR circles were not significantly different in LEDGF KO MEFs [11]. In this study, we observed no clear difference in the number of 2-LTR circles upon LEDGF/p75 KO. Possibly, the complete absence of LEDGF/p75 affects other steps besides integration that might result in reduced nuclear import and circle formation. Alternatively, cellular pathways involved in 2-LTR formation may be affected. Opposing effects on circle formation by reduced import and reduced integration may finally result in an equal 2-LTR circle number. Alternatively, the sensitivity of 2-LTR circle quantification may be too low to detect a small increase.
In the last part of the manuscript we demonstrate that LEDGINs block the residual replication observed in LEDGF/p75 KO cell lines (Figure 5C) and block the interaction in vitro between HRP-2IBD and IN (Figure 6B). Figure 6D illustrates how LEDGINs fit in the pocket at the IN core dimer interface. LEDGINs block the interaction with two interhelical loops of the IBDs of LEDGF/p75 (Figure 6E) or HRP-2 (Figure 6F). The inhibition of the interaction with HRP-2 can explain why residual replication of HIV-1 in LEDGF/p75 KO cells is still sensitive to LEDGINs. Since LEDGF/p75 has been reported to act as an allosteric modulator of the IN activity in vitro [1], [12], [45], [46], it is plausible that inhibition of the LEDGF/p75-IN interaction not only interferes with its function as a molecular tether but also results in an allosteric inhibition of IN activity. In fact, inhibition of in vitro IN activity in the absence of LEDGF/p75 by potent LEDGINs has been reported [25]. The allosteric mode of inhibition by LEDGINs can as well explain inhibition of HIV-1 replication in LEDGF/p75 KO HRP-2 KD cells [25]. In vivo both mechanisms are intrinsically coupled. LEDGINs compete with LEDGF/p75 as a molecular tether and at the same time interfere with integrase activities probably by affecting conformational flexibility in the intasome. Whereas transdominant inhibition of HIV-1 replication by IBD overexpression [4], [35] presumably also acts through this dual mechanism [46], RNAi-mediated depletion of LEDGF/p75 likely only affects tethering and/or targeting. We should however be cautious to translate the results in KO cells to human patients. Since no individuals without functional LEDGF/p75 expression have been documented, LEDGINs will always have to compete with LEDGF/p75 for the IN binding pocket to inhibit integration.
Somatic KO cell lines are cumbersome to generate. This is why few studies used this technology to study the role of cellular cofactors in virus replication. Previously, the role of cyclophilin A in HIV replication was confirmed in a human somatic KO cell line [47] as well as the roles of CBF1 [48] and TB7 [49] in Epstein-Barr virus replication. Our work supports the value of generating human KO cell lines for cofactor validation and drug discovery in general.
Nalm-6 cells, SupT1 cells, obtained from the ATCC (Manassas, VA) and PM1 cells, a kind gift from Dorothee von Laer (Innsbruck Medical University, Innsbruck, Austria), were maintained in RPMI 1640 – GlutaMAX-I (Invitrogen, Merelbeke, Belgium) supplemented with 8% heat-inactivated fetal calf serum (FCS; Harlan Sera-Lab Ltd.) and 50 µg/ml gentamycin (Gibco, Invitrogen). HEK293T cells, obtained from O. Danos (Genethon, Evry, France), and HeLaP4 cells, a kind gift from Pierre Charneau, Institut Pasteur, Paris, France, were grown in DMEM (Invitrogen) supplemented with 5% FCS, 50 µg/ml gentamycin and 0.5 mg/ml geneticin (Invitrogen). All cells were grown in a humidified atmosphere with 5% CO2 at 37°C.
For growth curve analysis, Nalm-6 cells were seeded at 100,000 in 5 ml of corresponding medium and HeLaP4 cells at 200,000 per well in a 6-well plate. Cell growth was followed on consecutive days by cytometry (Coulter Z1, Beckmann Coulter). Experiments were performed in triplicate.
The HIV-based lentiviral transfer plasmid pCHMWS_CD4_IRES_Bsd encodes the CD4 receptor driven by a human early cytomegalovirus (CMV) promoter followed by an EMCV internal ribosomal entry site (IRES) and a blasticidin resistance cassette (Bsd). The plasmid was generated by PCR amplification of human CD4 from a T-cell cDNA library using CD4-Fwd and CD4-Rev, followed by digestion with BamHI and XbaI, and cloning into pCHMWS_LEDGF_BC_IRES_Bsd [18], digested with BamHI and SpeI.
The lentiviral transfer plasmids for miRNA-based KD were generated based on miRNA-R30 as previously described [50], [51] (Table S2). For HRP-2 KD, miR HRP-2 was adapted from the sequence validated previously [6]. As negative controls a non-functional, scrambled miRNA30-based short-hairpin sequence (miR scrambled) and a functional, short-hairpin sequence targeting the monomeric red fluorescent protein from Discosoma corallimorpharia, DsRed (miR DsRed) were designed [33]. PCR fragments were introduced into the XhoI–BamHI sites from a modified pN3-eGFP plasmid (Clontech, Saint Quentin Yuelines, France) duplicated and cloned into the XhoI-KpnI sites at the 3′ end of the enhanced green fluorescent protein (eGFP) reporter cDNA, driven by a Spleen focus forming virus LTR (SFFV) promoter, resulting in pCSMWS_eGFP_miR_HRP2 and pCSMWS_eGFP_miR_scrambled. To generate pCSMWS_Zeo_miR_HRP2, the zeocin resistance cassette (Zeo) was amplified with primers Zeo-Fwd and Zeo-Rev from pBUD (Invitogen), digested with NheI-Pfl23II and inserted into the XbaI–Pfl23II digested pCSMWS_eGFP_miR_HRP2 plasmid. To generate pCSMWS_Zeo_miR_DsRed, miR_DsRed was cloned into the XhoI/KpnI digested pCSMWS_Zeo_miR_HRP2 plasmid. To generate pCSMWS_Hygro_miR_HRP2, the hygromycin B resistance cassette (Hygro) was amplified using Hygro-Fwd and Hygro-Rev as primers and pBud (Invitrogen) as a template. The resulting products were digested ClaI-XhoI and ligated into pCSMWS_Zeo_miR_HRP2.
For bacterial expression of C-terminal His6-tagged HIV-1 IN and MBP-tagged-LEDGF325–530, the plasmids pKBIN6H [10] and pMBP-Δ325 [4] were used, respectively. To construct pMBP-HRP-2448–670, the sequence corresponding to aa 448 to 670 of HRP-2 was PCR amplified with primers HRP2-Fwd, and HRP2-Rev, using p3xFlagHRP-2 (a kind gift from E. Poeschla) as a template. The resulting products were digested and ligated into pMAL-c2E (New England Biolabs Inc., USA). The integrity of all plasmids was confirmed by DNA sequencing.
LV production was performed as described earlier [18], [52]. Briefly, vesicular stomatitis virus glycoprotein (VSV-G) pseudotyped lentiviral vector particles were produced by PEI transfection in HEK293T cells using the different transfer plasmids. Single round HIVNL4.3ΔNefΔEnvfLuc (HIV-fLuc) virus was prepared by co-transfection of HEK293T cells with pNL4-3.LucR–E– (pHIV-fLuc, National Institutes of Health AIDS Research and Reference Reagent Program) and pMD.G, that codes for VSV-G.
For lentiviral transduction experiments, Nalm-6 cells were typically plated at 150,000 cells per well in a 96-well plate and transduced overnight. After 72 hrs, 50% of cells were reseeded for luciferase expression quantification and/or FACS analysis. The remainder was cultured for quantitative PCR and integration site analysis during at least 10 days to eliminate non-integrated DNA. HeLaP4 cells were plated at 20,000 cells per well in a 96-well plate and transduced overnight. After 72 hrs, 50% of cells were reseeded for luciferase quantification. The remainder was cultured for quantitative PCR or integration site analysis as described for Nalm-6 cells.
Targeting plasmids for generation of PSIP1 KO were designed and cloned as described previously [27], [53] utilizing the MultiSite Gateway System (Invitrogen) as described [54]. Briefly, a 2.3 and a 2.0 kb fragment for the left and right arms of the targeting plasmids, respectively, were amplified by genomic PCR using primers LEDGF/p75 attB4 and LEDGF/p75 attB1 for the left arm, and primers LEDGF/p75 attB2 and LEDGF/p75 attB3 for the right arm (Table S2). The resulting fragments were cloned into pDONR/P4-P1R and pDONR/P2R-P3 via recombination, resulting in p5′-ENTR-left and p3′-ENTR-right, respectively. The fragments p5′-ENTR-left, p3′-ENTR-right, pDEST DTA-MLS, and pENTR lox-Puro or pENTR lox-Hygro, were then ligated using recombination to generate the final targeting plasmids pTARGET-LEDGF/p75-Hyg and pTARGET-LEDGF/p75-Puro, respectively.
Cell lines were generated as previously described [27]. Briefly, targeting plasmid was transfected with Nucleofector I (Amaxa, Inc., Gaithersburg, MD, USA) using 2×106 Nalm-6 cells and 2 µg of DNA. At 24 hrs after transfection, cells were seeded into 96-well plates at 103 cells per well, in culture medium supplemented with either 0.2 µg/ml puromycin (BD BioSciences, San Jose, CA, USA) or 0.35 mg/ml of hygromycin B (Clontech, Mountain View, CA, USA). After 2–3 weeks, individual drug resistant colonies were propagated and analyzed by genomic PCR using primers A and B or C (Figure 1A), generating a 3272 bp AB-fragment for Nalm+/c, Nalmc/c, Nalm−/− and a 3123 bp AC-fragment for Nalmc/c, Nalm−/− (Figure 1C). Genomic PCR of the targeted region was performed with primers D and E generating a 1624 bp DE-fragment in Nalm+/+, Nalm+/c and a 288 bp DE-fragment in Nalm−/− (Figure 1C). Targeting efficiency was calculated as the ratio of the number of cell clones where the LEDGF/p75 allele was disrupted by homologous recombination to the number of drug-resistant cell clones (Figure 1B). Sequencing of the genomic KO region was performed as follows. The PCR fragments obtained after amplification with primers gFB and gRB spanning a 1885 bp region around exon 11–14 in wild-type cells or a 571 bp region in KO cells, followed by nested PCR with primers gFA and gRA resulting in a 1694 bp in wild-type or 380 bp region in KO cells, were cloned into the pCRII-TOPO plasmid (Invitrogen, Merelbeke, Belgium) and sequenced with primers M13-Fwd and M13-Rev (Figure S1A). Total RNA extracted from KO clones (RNeasy 96 kit, Qiagen) was used for cDNA synthesis using oligo-dT primers (High capacity cDNA RT kit, Applied biosystems). Correct recombination was verified at mRNA level for LEDGF/p52, LEDGF/p75 and truncated LEDGF/p75 by PCR on cDNA using primers RNA-A and RNA-B for LEDGF/p52, LEDGF/p75 and LEDGFKO, RNA-A and RNA-C for LEDGF/p75 and LEDGFKO, RNA-A and RNA-D for LEDGF/p52, resulting in fragments of 245 bp, 1.606 kb or 1.163 kb and 1.011 kb, respectively (Figure S1C). Additionally, the cDNA of the truncated protein was sequence verified (Figure S1A) as follows: a PCR product generated by primers d243 and RNA-C, followed by nested PCR with primers d244 and LEDGF-R-exon15, was cloned into pCRII-TOPO (Invitrogen) and sequenced with M13-Fwd and M13-Rev. Primers are listed in Table S2.
Stable CD4 expressing Nalm-6 cell lines were generated by transducing wild-type Nalm+/+, Nalm+/c cl 31 and Nalm−/− cl 1 and cl 2, with the lentiviral vector pCHMWS_CD4_IRES_Bsd and subsequent selection with blasticidin (3 µg/ml; Invitrogen, Merelbeke, Belgium). CD4 expression was verified by flow cytometry using R-Phytoerythrin-conjugated mouse anti-human CD4 monoclonal antibody (BD pharmigen) according to the manufacturer's protocol.
Stable monoclonal LEDGF/p75 KD cells were generated previously [18]. Additional HRP-2 KD was obtained by transduction of HeLaP4 wild-type cells and LEDGF/p75 KD cells with pCSMWS_Zeo_miR_HRP2, pCSMWS_Hygro_miR_HRP2 or pCSMWS_eGFP_miR_HRP2. Transduced cells were selected with zeocin (200 µg/ml) or hygromycin B (200 µg/ml) or by FACS sorting of eGFP positive cells respectively. Control cell lines were generated likewise by transduction with vectors encoding pCSMWS_Zeo_miR_DsRed, pCSMWS_eGFP_IRES_HygroR and pCSMWS_eGFP_miR_scrambled, respectively. Stable PM1 and SupT1 LEDGF/p75 KD cell lines were generated with LV, encoding a miRNA cassette targeting LEDGF/p75 under control of an SFFV promoter (unpublished data). Additional HRP-2 KD and control PM1, SupT1 and Nalm-6 cell lines were generated by transducing the cells with vectors made with pCSMWS_Hygro_miR_HRP2 and pCSMWS_eGFP_IRES_HygroR, respectively.
Integration sites were amplified by linker-Mediated PCR as described previously [17], [18]. For integration sites to be authentic, sequences needed a best unique hit when aligned to the human genome (hg18 draft) using BLAT. The alignment began within 3 bp of the viral long terminal repeat end, and had >98% sequence identity. Reanalysis of previously obtained integration sites [2], [11], [17] was performed in parallel. Statistical methods are described previously [55]. Integration site counts were compared using a two-tailed Fisher's exact test. Analysis was carried out using Prism 5.0 (GraphPad Software).
The origin of HIVNL4.3 [56] and HIVIIIb has been described [57]. Clinical isolate #1 was obtained through the AIDS research and reference reagent program, Division of AIDS, NIAID, NIH: HIV-1 93TH053 from the UNAIDS network for HIV isolation. Clinical isolate #2 was obtained through the AIDS research and reference reagent program, Division of AIDS, NIAID, NIH: HIV-1 96USSN20 from Drs Ellenberger, P. Sullivan and R.B. Lal [32]. The p24 antigen titer was determined for each virus stock. The MOI was determined using flow cytometry analysis of intracellular p24 antigen 24 hrs after infection in control Nalm+/c cells. Cells were stained with pycoerythrin-anti-p24 (KC57-RD1; Beckman Coulter) using the Fix&Perm (Invivogen) cell fixation and cell permeabilization kit following the manufacturer's protocol.
HIV-1 infection of Nalm-6 cells was typically performed with 1*106 cells in 5 ml of medium with the indicated virus and MOI. After 6–12 hrs of infection, cells were washed twice with PBS and resuspended in the initial volume of culture medium. Infection of HeLaP4 cells was performed as described previously [18]. HIV-1 replication was monitored by quantifying p24 antigen in the supernatant daily via ELISA (Alliance HIV-1 p24 ELISA kit; Perkin Elmer). Cells were split 1/6 every 5–6 days if experiments exceeded 10 days. PM1 and SupT1 cells were infected at 0.01 pg p24/cell.
Proviral DNA extraction of infected cells was performed using the QIAamp blood kit (Qiagen) according to the manufacturer's protocol. PCR amplification and sequencing of IN encoding sequences were done as described previously [58].
Zidovudine (AZT) and ritonavir (RIT) were purchased and raltegravir (MK518) was kindly provided by Tibotec (Mechelen, Belgium). LEDGIN 7 was synthesized as described [25].
Cells harvested from a 96-well plate were lysed with 50 µl lysis buffer (50 mmol/l Tris pH 7.5, 200 mmol/l NaCl, 0.2% NP40, 10% glycerol). The lysate was assayed according to the manufacturer's protocol (ONE-Glow; Promega, Madison, WI). Luciferase activity was normalized for total protein (BCA; Pierce, Rockford, IL). All conditions were run at least in triplicate in each experiment.
HeLaP4 cells were transfected with pHIV-fLuc using Lipofectamine 2000 (Invitrogen, Merelbeke, Belgium) according to the manufacturer's protocol with minor modifications. Briefly, 70,000 cells were seeded in a 96 well plate and transfected after one day with a mixture of 333 ng DNA and 0.66 µl Lipofetamine 2000 for 4 hrs and washed afterwards twice with PBS. 48 hrs post transfection cells were harvested for luciferase activity quantification.
Quantification of LEDGF/p75 mRNA levels was performed as described previously [18]. Similar settings were used to determine HRP-2 mRNA levels. HRP-2 primer/probe set: HRP2 s4, HRP2 as4 and HRP2 probe. In all cases, RNaseP was used as endogenous house-keeping control (TaqMan RNaseP Control Reagent; Applied Biosystems). All samples were run in triplicate for 3 minutes at 95°C followed by 50 cycles of 10 seconds at 95°C and 30 seconds at 55°C. Data were analyzed with iQ5 Optical System Software (BioRad, Nazareth, Belgium). To quantify the different HIV-1 DNA species qPCR for total viral DNA, 2-LTR circles and integrated copies was performed as described [59], [60], with minor modifications. Nalm-6 cells were seeded one day prior to infection at 2*105 cells per ml. After 4 hrs of incubation with HIV, medium was replaced by RPMI containing 10% FCS. Quantification of proviral copies as shown in Figure 3G was performed accordingly, only RIT (at 50 times IC50) was added to the culture medium after the washing step to ensure only a single replication cycle could take place and genomic DNA was isolated after 5 days to dilute all non-integrated forms. Non-infected cells were incubated in parallel. To quantify the number of integrated copies as shown in Figure S4H, cells were cultured for 10 days in medium containing RIT and AZT both at 25 times IC50 following day 39, 45 or 48, before harvesting genomic DNA. Quantitative Alu-PCR for quantification of proviral copies was done in two steps [60]. The first phase amplifies from Alu sequences to U3 sequences absent in self-inactivating (U3-deleted) HIV-1 vectors using 400 nM AluSINIIfwd, 400 nM qAluRout_SB704. Amplification conditions were 95°C for 30 sec, 60°C for 40 sec, 72°C for 1 min 30 sec, ×13 cycles. The second phase amplifies a nested product using 300 nM sense primer Q-Alu-F-in, 300 nM antisense Q-Alu-R-in and 200 nM Alu-probe. PCR conditions were 95°C for 10 sec, 55°C for 30 sec, ×50 cycles.
Southern blot analysis was performed as described previously in [61], [62]. Briefly, genomic DNA was digested with BamHI, separated by electrophoresis on a 0.7% agarose gel and blotted on positively charged nylon membranes (Biodyne B; Pall Corp., Pensacola, FL, USA). The probe covered a 1024 bp genomic region around exon 10 of PSIP1 and was amplified by PCR using LPROBE-Fwd and LPROBE-Rev, and labeled with α-32P-dCTP (Megaprime DNA labeling system, GE Healthcare, USA). Signals were detected using autoradiography.
Western blotting was performed as described previously [5]. Briefly, cellular extracts were separated by sodium dodecyl sulfate–polyacrylamide gel electrophoresis. LEDGF was detected using a purified IgG1 mouse anti-human LEDGF monoclonal antibody (mAb, C26) (BD Biosciences Pharmingen, San Diego, CA). Equal loading was verified with β-tubulin (T-4026; Sigma-Aldrich, St Louis, MO). Visualisation was performed by chemiluminescence (ECL+; Amersham Biosciences, Uppsala, Sweden).
Recombinant HIV-1 IN containing a C-terminal His6 tag was purified as described previously [10]. LEDGF325–530 and HRP-2448–670 fragments were expressed in E. coli as maltose binding protein (MBP) fusions. The purification of pMBP-LEDGF325–530 from BL21(DE3) bacterial cells was done as described previously [4]. For purification, cells were resuspended in lysis buffer (50 mM Tris-HCl, pH 7.2, 500 mM NaCl, 5 mM dithiothreitol, 1 mM EDTA, 0.2 mM phenylmethylsulfonyl fluoride, 0.1 U/ml DNase). After complete lysis by ultrasonication, the supernatant was cleared by centrifugation and recombinant proteins were bound to amylose resin (New England Biolabs Inc, United Kingdom). The resin was washed with 20 bed volumes wash buffer (50 mM Tris-HCl, pH 7.2, 500 mM NaCl, 5 mM dithiothreitol), and the MBP-tagged proteins were eluted in 1 ml fractions wash buffer supplemented with 10 mM maltose. The fractions were analyzed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis for protein content, pooled, and concentrated by dialysis (overnight, 4°C) against storage buffer (50 mM Tris-HCl, pH 7.2, 500 mM NaCl, 50% (vol/vol) glycerol). All protein concentrations were measured using the Bradford assay (Bio-Rad).
AlphaScreen measurements were performed in a total volume of 25 µL in 384-well Optiwell microtiter plates (PerkinElmer). All components were diluted to their desired concentrations in assay buffer (25 mM Tris-HCl pH 7.4, 150 mM NaCl, 1 mM MgCl2, 0.1% Tween-20 and 0.1% BSA). Anti-MBP coated donor beads were generated by dialyzing biotin-labelled anti-MBP (Vector Laboratories) to the assay buffer and incubating 10 nM of this antibody with the desired amount of Streptavidin donor beads (PerkinElmer) for 1 h at room temperature. For the KD determinations, HIV-1 IN-His6 was titrated against a background of 10 nM MBP-LEDGF325–530 or MBP-HRP-2448–670. This amount provided minimal binding curve perturbation while maintaining a good signal-to-noise ratio. When performing IC50 determinations, LEDGIN 7 was titrated against a background of 500 nM IN-His6 and 10 nM MBP-LEDGF325–530 or MBP-HRP-2448–670. After addition of the proteins and/or compounds, the plate was incubated for 1 h at 4°C and 20 µg/mL anti-MBP donor and Ni2+-chelate acceptor beads (PerkinElmer) were admixed, bringing the final volume to 25 µL. After 1 h of incubation at RT, protected from light, the plate was read on an EnVision Multilabel Reader in AlphaScreen mode (PerkinElmer). Results were analyzed in Prism 5.0 (GraphPad software) after non-linear regression with the appropriate equations: one-site specific binding, taking ligand depletion into account for the KD measurements and sigmoidal dose-response with variable slope for the IC50 determination.
The Genbank (http://www.ncbi.nlm.nih.gov/genbank) accession numbers for the proteins discussed in this paper are LEDGF/p52 (NM_021144.3), LEDGF/p75 (NM_033222.3) and HRP-2 (NM_032631.2).
|
10.1371/journal.pbio.2003962 | Design of synthetic bacterial communities for predictable plant phenotypes | Specific members of complex microbiota can influence host phenotypes, depending on both the abiotic environment and the presence of other microorganisms. Therefore, it is challenging to define bacterial combinations that have predictable host phenotypic outputs. We demonstrate that plant–bacterium binary-association assays inform the design of small synthetic communities with predictable phenotypes in the host. Specifically, we constructed synthetic communities that modified phosphate accumulation in the shoot and induced phosphate starvation–responsive genes in a predictable fashion. We found that bacterial colonization of the plant is not a predictor of the plant phenotypes we analyzed. Finally, we demonstrated that characterizing a subset of all possible bacterial synthetic communities is sufficient to predict the outcome of untested bacterial consortia. Our results demonstrate that it is possible to infer causal relationships between microbiota membership and host phenotypes and to use these inferences to rationally design novel communities.
| Symbiotic microbes influence host development and health, but predicting which microbes or groups of microbes will have a helpful or harmful effect is a major challenge in microbiome research. In this article, we describe a new method to design and predict bacterial communities that alter the plant host response to phosphate starvation. The method uses plant–bacterium binary-association assays to define groups of bacteria that elicit similar effects on the host plant. By constructing partially overlapping bacterial communities, we demonstrated that it is possible to modify phosphate accumulation in the plant shoot and the induction of plant phosphate starvation genes in a controlled manner. We found that bacterial colonization of the plant root does not predict the capacity to produce this phenotype. We evaluated the predictive performance of different statistical models and identified one best able to predict the behavior of untested communities. Our work demonstrates that studying a subset of all possible bacterial communities is sufficient to anticipate the outcome of novel bacterial combinations, and we establish that it is possible to deduce causality between microbiome composition and host phenotypes in complex systems.
| The composition of plant-associated microbial communities influences plant health and development [1][2]. This has raised interest in the use of microbes for biotechnology and agriculture [3][4]. However, it is challenging to measure the contribution of individual microbes from a complex microbiota to host health. Thus, a number of in vitro screening strategies are commonly applied to identify candidate plant-interacting microbes; however, none of the traits typically screened are correlated with a plant-beneficial outcome [5]. Another common prescreening strategy involves performing plant–bacterium binary-association assays [6], but only a few have been successfully translated into agricultural settings [7][8][9], suggesting that these assays also fail to capture critical aspects of nature’s complexity. Moreover, it is well established that microbial consortia can produce strong and unexpected effects on host health [10][11], and such emergent properties are hard to predict, hindering the rational design of microbial consortia with desired host outputs. Previous strategies to address this conceptual problem included the exhaustive study of possible communities assembled from a small number of microbiota constituents in zebrafish [12] and the analysis of randomized combinations of bacteria in mice [13]. Other approaches often begin with an exhaustive evaluation of all combinations of, for example, the nutrients nitrate and ammonium and the hormones auxin, cytokinin, and abscisic acid on plant root growth and development [14]. Although exhaustive approaches can provide a complete picture of interactions within complex systems, they are unfeasible for systems with more than a handful of variables, given the astronomical number of possible factorial combinations. Even in the rare cases in which functional microbial consortia have been assembled, most studies focus on a single community that is considered a treatment, and rarely is an effort made to dissect the contribution of its constituents. This makes it impossible to establish predictable generalizations beyond the tested communities or conditions used. An instance that dissected the components of a consortium consisted of only 2 bacterial strains [15]. These findings reinforce the necessity for reduced complexity and modular model systems to associate microbial community composition with host phenotypes.
Our approach is summarized in Fig 1. In short, we first characterized the relationship between in vitro bacterial assays and plant–bacterium binary-association assays, and we used the latter to define functional bacterial blocks. These blocks are groups of bacteria that, by themselves, have a similar influence on a host phenotype. Then, we defined a subset of all the possible communities by constructing partially overlapping synthetic communities (SynComs) of 2 blocks each, tested the effect of these consortia on multiple plant phenotypes, and characterized the plant transcriptional response to these consortia. We evaluated the predictive performance of different statistical models on communities that the models had not seen before. We selected a neural network (NN) because it maximized predictive performance and used this model to design novel synthetic communities that maximized the change in 1 plant phenotype. Finally, we tested the model designs by constructing the novel communities it suggested and validated nearly all of the predicted host phenotypic outputs.
We focus on bacterial manipulation of the plant response to phosphate (Pi) starvation, a commonly limiting nutrient for plant growth [16]. Pi is an essential macronutrient for plants and also for microbes [17][18] and is limited in soil [19]. Microbial communities living in the proximity of the plant take up Pi from the environment using a highly efficient Pi transport system [20][21]. Therefore, the available Pi in the close vicinity of plants is subject to direct and intense competition for uptake between microbes and plants [18]. Although the response of Arabidopsis thaliana seedlings in axenic conditions to phosphate starvation is well characterized [22], the elucidation of the regulatory mechanisms of this response in the presence of the plant microbiome is only recently emerging [23] [24].
We systematically evaluated the performance of a large collection of root bacterial isolates using in vitro screening and binary plant–bacterium association assays as predictors for the effect of derived bacterial consortia on plant phenotypes in response to phosphate starvation. We confirm that bacterial in vitro assays have no correlation with bacterial effects on plant phenotypes. However, we found that plant–bacterium binary-association assays are informative for designing small synthetic communities. Surprisingly, the effects of bacterial consortia on host physiology were mostly additive and independent of bacterial abundances, suggesting that functional stacking within a microbial consortium can determine its effect on host phenotypic response. Finally, we successfully validated novel synthetic communities designed by an NN that led to predictable changes in plant shoot Pi content. Our results provide a useful road map from binary host–microbe assays to the design and testing of useful small consortia to predictably alter host phenotypes.
In response to Pi deficiency, plants change root exudate metabolite profiles and root architecture to explore Pi-rich soil patches [25]. This may lead to bacterial soil community shifts [26]. In order to learn how root exudate profiles change in response to Pi, we harvested root exudates from A. thaliana plants in response to 2 short and complementary nutritional transitions that mimic the dynamics of Pi stress [27] (S1A Fig; Materials and methods 1a, 1b, 1d, 1e). We demonstrated that our Pi transitions were sufficient to induce a reconfiguration of plant exudate primary metabolic profiles (S1C and S1D Fig, and S1 Table).
We next tested whether these exudates modified the in vitro growth capacity of a collection of 440 bacterial strains isolated from the roots of Brassicaceae grown in soil that is not overtly Pi deficient (nearly all from A. thaliana) ([28], S1B Fig, and Materials and methods 1c). We identified a range of bacterial growth behaviors (Materials and methods 1f, 1g) and found that the bacterial growth differences between phosphate conditions are much weaker than the differences between strains (S2 Fig). As expected, phylogeny explained most of the growth differences between strains (S2A Fig and Materials and methods 1h). Most of the bacterial growth parameters provided the same information, so we selected the area under the growth curve (optical density [OD] versus time) (AUC) as a growth marker for subsequent analyses.
Hierarchical clustering of AUC differences between in vitro conditions identified 10 groups of bacteria that represented different response patterns to exudates derived from roots grown in different Pi concentrations and media supplemented or not with Pi (S2B Fig and S2 Table). We found that root exudates could enhance or inhibit bacterial growth and that this effect could be either general or specific to one type of exudate (S2B Fig). Thus, consistent with previous findings [26], plant-derived root exudates modulated the growth of bacterial root isolates depending on the plant’s Pi starvation status.
We selected a subset (n = 183) of the strains from the in vitro assays for determining whether they exerted a functional role on the plant under different phosphate conditions. We selected bacterial isolates that belonged to all of the different response patterns (S2B Fig) and that were most responsive to both Pi levels and the presence of exudates (Materials and methods 1g, S2 Table). We measured the change in plant shoot Pi content, a direct marker of phosphate starvation responsiveness [22], in response to the presence of each of 183 individual strains, when compared to axenically grown plants. We evaluated shoot Pi content under 4 Pi conditions that represented a 2 × 2 design matrix of 2 Pi levels used for plant germination (full; 1 mM; and depleted, about 5 μM Pi) and 2 Pi concentrations (30 μM Pi and 100 μM Pi) to which seedlings were switched, concomitant with the application of each bacterial strain (Fig 2A and Materials and methods 2). The use of 2 germination conditions in the experimental design allowed us to evaluate the effect of the activation of the phosphate starvation response and the shoot Pi content on the plant–bacterium interaction under different Pi concentrations.
On average, bacteria had a slightly negative effect on plant shoot Pi content, visualized as a small tail in the bacterial treatment graphs (pink) in Fig 2B. This effect was stronger when the environmental Pi concentration was lower (Fig 2B and 2C and S3 Table). These findings are consistent with our previous results that a bacterial synthetic community drives a context-dependent competition with the plant for Pi [23]. Overall, we found that more strains had a negative than a positive effect on shoot Pi content (Fig 2B, S3 Table, and Materials and methods 2e). Specifically, there were significantly more strains that had a stronger negative effect on plant shoot Pi content in the most limiting Pi conditions (germination in Pi depleted, followed by transfer to 30 μM Pi) (Fig 2B and S3 Table), in which the phosphate starvation response should be active. Conversely, the least Pi-deprived condition (germination in full Pi, followed by transfer to 100 μM Pi) exhibited a significant enrichment of strains that positively affected shoot Pi content (Fig 2B and S3 Table). These results are consistent with bacterial effects on Pi content in the shoot being modulated by the nutritional status of the plant. Importantly, germination conditions did not alter bacterial colonization (Fig 2D), and the effect of individual strains on plant shoot Pi content was independent of the ability of root-inoculated bacteria to colonize the shoot and independent of bacterial titers in different plant organs (Fig 2D; S3A Fig and Materials and methods 2c).
We detected a weak phylogenetic signal in the ability of bacterial strains to modulate plant Pi content that was significant in only 2 of the 4 conditions (Fig 2C and Materials and methods 2f). Accordingly, we found no correlation between the effect of individual bacterial isolates on shoot Pi content and their in vitro growth phenotype in response to switched Pi levels and root exudates (S4 Fig). Overall, our survey of plant–bacteria binary associations and the resulting distribution of bacterial effects on shoot Pi content argue that the majority of plant–bacteria interactions are competitive, at least in the context of phosphate starvation response.
We recently demonstrated that the A. thaliana phosphate starvation response is largely antagonistic to immune system function [23]. We therefore asked whether activation of the plant phosphate starvation response could modulate the outcome of binary bacteria–plant interactions. We analyzed shoot Pi content in plants pretreated with phosphite (Phi) (KH2PO3) and then transferred to either 30 μM Pi or 100 μM Pi in the presence of each of 30 selected bacterial strains that either reduced, increased, or had no effect on the shoot Pi content (10 strains per class; S3E Fig). Phi is a nonmetabolizable analog of Pi and its accumulation delays the phosphate starvation response, resulting in low accumulation of Pi in the shoot [29](S3B and S3C Fig). We found that germinating plants on Phi (low shoot Pi, phosphate starvation response off) dramatically reduced the number of bacterial isolates that diminished shoot Pi content, compared to germination on low Pi (low shoot Pi, phosphate starvation response on) (S3D and S3E Fig). Additionally, we observed that under Phi pretreatment, none of the strains significantly increased shoot Pi content compared to germination on high Pi (high shoot Pi, phosphate starvation response off) (S3D and S3E Fig). Importantly, Phi treatment did not alter bacterial colonization (Fig 2D, S3A Fig and Materials and methods 2c). These findings indicate that activation of the plant phosphate starvation response results in different modes of bacterial interactions with the plant that are independent of shoot phosphate content. These results indicate that an active phosphate starvation response can modulate the outcome of both positive and negative interactions with bacteria, likely mediated via coregulation of the plant immune system. An analogous mechanism has been described for the interaction between A. thaliana and a beneficial fungus [24].
We sought to establish whether the results from binary associations are indicative of bacterial effects when a more complex bacterial community is present. We used a microcosm reconstitution approach, in which we inoculated plants with defined complex bacterial synthetic communities (Materials and methods 3a). A subset of 78 strains analyzed in the binary-association experiments was grouped into 3 functional groups consisting of positive (P1-P3), indifferent (I1-I3), and negative (N1-N3) bacteria, depending on their effect on shoot Pi accumulation. For the positive and negative groups, we focused on strains that had a statistically significant effect on plant Pi accumulation, after correcting for multiple testing (Materials and methods 3i). Each functional group was further divided into another 3 blocks of 8–9 bacterial strains, according to the magnitude of their individual effects (Fig 3A, S3 and S4 Tables, and Materials and methods 3i). We then combined pairs of these blocks to define 14 partially overlapping bacterial synthetic communities (Fig 3B). This scheme was designed to maximize the probability of observing extreme plant phenotypes by stacking functionally similar blocks and to gain information from combining the most extreme phenotypic blocks defined in the binary-association assays.
We evaluated shoot Pi content, primary root elongation, shoot size, and total root network in A. thaliana plants grown in association with the 14 bacterial synthetic communities in the same growth conditions used for the binary-association analysis (Fig 2A and Materials and methods 3b). We found that synthetic communities, like individual bacterial strains, were more likely to reduce plant shoot Pi content, and that synthetic communities made of negative blocks led to lower shoot Pi accumulation than those composed of positive blocks (Fig 3C and 3D and S5 Fig). For example, in general, the estimated negative effect on shoot Pi accumulation for negative blocks is significantly larger than for positive or indifferent blocks (Fig 3C and 3D, S5a Table, and Materials and methods 3j). At the synthetic community level, the effect of the negative strains was clearly dominant; only 2 communities containing negative blocks (I3N1 and N2N3) showed a nonsignificant reduction in shoot Pi content and this in only 1 of the tested conditions. Importantly, the only significantly positive effect with respect to no bacteria involved 2 positive blocks and was weak (P1P3) (Fig 3C, S5b Table, and Materials and methods 3j). We also observed that the majority of the cases in which a synthetic community did not significantly reduce the shoot Pi accumulation occurred under the less Pi-restricted condition (100 μM) (Fig 3C, S5b Table, and Materials and methods 3j), consistent with the results from the individual strains (Fig 2B). This trend was generally consistent for the other plant phenotypes analyzed (Fig 3C and 3D and S5 Fig). Overall, the reduction in shoot Pi content associated with negative blocks correlated with less shoot area, shorter primary roots, and bigger root networks (top and bottom rows in Fig 3C and 3D, and S5 Fig), morphological changes that match the canonical phosphate starvation response in axenic conditions [22][30]. In contrast, positive bacterial blocks caused less intense plant phosphate starvation response phenotypes. These effects were more obvious in plants grown at low environmental Pi concentration (Fig 3C and 3D and S5 Fig). Thus, the binary-association assays were generally informative with regard to the behavior of bacteria in more complex biotic backgrounds.
Interestingly, we observed that a number of synthetic communities, for example P2P3 and P1P2, led to increased shoot area compared to axenically grown plants, despite exhibiting reduced shoot Pi content (Fig 3C and S5 Fig). In contrast, plants treated with P1P3 in +Pi_100 μM Pi condition, had shoot Pi content similar to Pi-sufficient plants but unexpectedly exhibited a reduced shoot area (Fig 3C and S5 Fig). Thus, bacterial consortia can decouple shoot Pi-content accumulation from the growth inhibition responses typically associated with the canonical phosphate starvation response [22][30][31].
We estimated the common (additive) effects of each block of strains across different bacterial backgrounds (e.g., in different synthetic communities) (Materials and methods 3j). Surprisingly, we found that additive contributions of the bacterial functional blocks are sufficient to explain most of the plant phenotypic variation observed (Fig 4). We found that synthetic community membership (i.e., ignoring bacterial relative abundances) typically explained more than 50% of the plant phenotypic variance (Fig 4). This indicates that intrablock bacterial interactions contribute at least as much as interblock interactions to the plant phenotypes tested. Furthermore, the effect of bacterial blocks on the phenotypes analyzed is generally consistent across different synthetic communities, despite each strain’s relative abundance being dependent on the microbial context (S6 and S7 Figs).
We found that the bacterial abundances in either agar substrate or in the root endophytic compartment were poorly correlated with plant phenotypes (Materials and methods 3e, 3k). Despite the consistent taxonomic profiles of the inoculum, we observed that bacterial communities of agar and root samples were dominated by variable bacterial taxa, depending on the specific combination of bacterial blocks present (S6A and S6B Fig). This suggests that bacteria–bacteria interactions are important in shaping the final community. Furthermore, we found clear taxonomic differences between root and agar samples. Most notably, Streptomyces strains (order Actinomycetales) were particularly good root colonizers despite their limited success on agar, while Pseudomonadales strains were relatively more successful in agar than in root samples (S6A and S6B Fig). These results recapitulate previous findings in natural soils, indicating that Actinobacteria are enriched in A. thaliana roots [32][33]. Phosphate concentrations in the media had only a minimal effect on the final community composition (S6B Fig).
We then quantified the information gained by incorporating relative abundance data (S6 Fig) into our additive model (Materials and methods 3k). Surprisingly, in all cases (16/16) the plant phenotypic variance explained by microbiota composition decreased when we incorporated relative abundance (S7 Fig). While in some cases, the differences might not be statistically significant, together, this result demonstrates significantly better performance by the model that ignores relative abundance (p-value = 0.000481; 2-tailed Wilcoxon signed-rank test). Our results indicate that bacterial blocks disproportionately modulate shoot Pi content with respect to their strain abundances, an observation analogous to that seen in bacteria modulating zebrafish immune responses [12].
The synthetic communities differentially modulated plant phenotypes related to phosphate starvation response. Therefore, we examined the transcriptomes of plants growing with different synthetic communities. We first explored the expression of a literature-based core set of 193 phosphate starvation response transcriptional markers [23]. Plants did not exhibit induction of phosphate starvation response markers in axenic conditions, even when Pi was low (Fig 5A) [23]. However, some synthetic communities induced the canonical transcriptional response to Pi starvation in plants grown on 30 μM Pi (Fig 5A). Plants that showed transcriptional activation of the phosphate starvation response displayed lower shoot Pi accumulation. However, we also observed that some synthetic community treatments lead to low shoot Pi content and no activation of the transcriptional phosphate starvation response (S8 Fig). The effect of synthetic communities was in general dependent on the presence of negative bacterial strain blocks (Fig 5A). In contrast, synthetic communities consisting of only positive blocks of bacteria did not induce the phosphate starvation response transcriptional signature in any condition analyzed (Fig 5A). No induction of the phosphate starvation response genes was observed when the Pi stress was released (following transfer to 100 μM Pi) except for the bacterial combination P3N3, which exhibited induction on 100 μM Pi (Fig 5A). In accordance with the shoot Pi content data (Figs 3C, 3D and 4), we found that additive effects of bacterial blocks could explain the level of transcriptional induction (Fig 5B). The specificity in the bacterial modulation of plant phenotypes suggests that the changes observed in the plant in response to the synthetic communities are linked to bacterial block activities.
We next explored the overall plant genome-wide transcriptional response to bacteria consortia, Pi conditions, or both. Our design allowed us to test both the response to synthetic communities and to individual bacterial blocks between and within conditions (Materials and methods 3h). As anticipated, plants growing with bacterial synthetic communities on low Pi generally induced phosphate starvation responsive genes and modified the expression of immune system–related genes (S9 Fig, S6 and S7 Tables) [23]. Overall, there was not a common response to bacterial presence, with only 45 and 35 genes being significantly up- or down-regulated by more than half of the bacterial blocks, respectively (S6 Table). The number of genes differentially expressed in response to different bacterial blocks did not correspond with the strain composition of the blocks; blocks P2, N3, and I1 altered the expression of the most genes, and blocks I3, N1, and P3 influenced the least (S6 Table). In particular, block I3 only altered the expression of 17 genes, despite being detected in plant roots and surrounding agar (S6 Fig). At the functional level, most of the bacterial blocks induced the expression of the plant defense response, specifically up-regulating genes for salicylic acid biosynthesis (S10 Fig), consistent with overall Bacteria versus No Bacteria comparisons (S9C Fig).
We also investigated differences between the genes induced by different bacterial blocks. Comparison of genes differentially expressed between positive and negative blocks across all conditions showed that positive blocks had higher expression of genes involved in energy production, while negative blocks specifically induced abiotic stress–responsive gene sets, specifically abscisic acid–related genes (Fig 6A, S6 and S7 Tables). Negative blocks of bacteria also increased the expression of a specific sector of the jasmonic acid response involved in glucosinolate biosynthesis (Fig 6A–6C, S6 and S7 Tables). The glucosinolate pathway modulates the interaction of A. thaliana with a beneficial fungus at low Pi [24], and its expression is regulated by the master regulator of phosphate stress response, PHR1 (PHOSPHATE STARVATION RESPONSE1) [23]. When the environmental Pi was low (30 μM Pi), we observed many more differentially expressed genes between positive and negative blocks (Fig 6A), with negative blocks driving higher expression of genes of both the phosphate starvation and defense response.
We then focused our analysis on the Pi-limiting conditions (30 μM). In this condition, synthetic communities containing negative blocks showed a strong induction of the phosphate starvation response (Fig 5A). We asked whether the different negative blocks (N1, N2, and N3) differed in their effects. There were almost no expression differences between the 2 most negative blocks (N2 and N3), but we identified 103 genes differentially regulated by bacterial blocks N1 and N3. These genes were mostly stress-related genes, including general abiotic stress and defense response, the expression of which was comparatively reduced in the phenotypically more negative block N3 (Fig 6D, S6 and S7 Tables). This result indicates that under phosphate starvation, all negative blocks activate a similar set of phosphate starvation response genes but differentially suppress other stress responses.
We found that some genes were induced in response to specific block combinations. For example, we found that PHOSPHATE2 (PHO2), a ubiquitin-conjugating E2 enzyme in A. thaliana required for the degradation of Pi transporters at high Pi [34], is highly expressed only in plants exposed to the synthetic community P3N3 in all Pi conditions analyzed (Fig 6E). This finding may explain the strong transcriptional response to Pi starvation caused by this synthetic community (Fig 5A). The auxin-regulated gene AUXIN-REGULATED GENE INVOLVED IN ORGAN SIZE (ARGOS) [35] showed a weak positive correlation with the induction of the phosphate starvation response, and it was induced in plants grown with the synthetic communities P3N3 and N2N3 (Fig 6F). ARGOS controls organ size in A. thaliana and its transgenic expression results in enlarged aerial organs [35]. This could serve to counterbalance the negative effect on shoot size that low Pi typically causes.
Our design of synthetic communities emphasized placing every bacterial functional block into at least 2 microbial backgrounds; therefore, we should be able to estimate bacterial effects that are independent of background. In principle, this estimation could be used to design novel synthetic communities with predictable outputs. We found that additivity of bacterial effects could explain most, but not all, of the host phenotypic variation. Therefore, we built 3 different quantitative predictive models capable of capturing different levels of complexity and evaluated their performance. We constructed a simple linear model (LM), a linear model that included pairwise interactions between bacterial functional blocks (INT), and a Neural Network model (Fig 7A and Materials and methods 4). We focused on shoot Pi content, which had the strongest signal-to-noise ratio (SNR) of all plant phenotypes tested (S11 Fig, Materials and methods 4b). To evaluate the predictive performance of each model, we used a form of cross validation in which the data from each synthetic community were held out one at a time, and the remaining synthetic communities were used to train each of the 3 models; those trained models were then used to predict the plant phenotypic output of the held-out consortium. We found that the NN had the lowest cross-validated prediction error of the 3 models and that the difference was statistically significant (p-value = 0.0073) (Fig 7B). Neural Networks are popular predictive models because they can capture more complex and nuanced relationships that simpler (linear) models cannot; however, this can come at a cost of reduced interpretability. We performed a sensitivity analysis (Materials and methods 4f) by calculating the effect that changing each variable would have on shoot Pi content according to the NN and the 2 linear models (LM and INT). We found a general agreement between the 3 models; for example, all models showed that Pi level in the media and the presence of negative bacterial blocks had the strongest effect on shoot Pi content, but the NN produces much more fine-grained results, because it is able to predict the change differentially across each condition (Fig 7C).
In order to validate the prediction accuracy of the NN, we chose the 25 bacterial block replacements that were predicted to result in the largest increase in shoot Pi content and experimentally tested whether an increase was produced by these synthetic communities (Fig 7D, S8 Table, and Materials and methods 4g). There was a significant correlation (ρ = 0.42, p-value = 0.0375) between predicted and observed shoot Pi content changes caused by the bacterial block replacements (Fig 7E). Strikingly, we found that 23 out of 25 bacterial block replacements increased shoot Pi content on average (p-value = 0.004; 1,000 permutation tests with synthetic community labels randomly permuted) (S9 Table). Moreover, the improvement in shoot Pi content was statistically significant in 16 out of 25 bacterial block replacements (p-value = 0.032; 1,000 permutation tests with synthetic community labels randomly permuted) (S9 Table). Only 1 out of 25 bacterial block replacements significantly decreased Pi content (S9 Table). Again, we noted little correlation between bacterial abundances and their effect on Pi content (S12 Fig and S9 Table). Compared to linear models (LM and INT), the NN had significantly lower prediction errors (p-value ≤ 4.65 × 10−7) (Fig 7F). In summary, we were able to rationally design novel synthetic communities that lead to predictable plant phenotypic outputs.
While plant responses to stress have been shown to be influenced by associated microbial communities, causal relationships in plant–microbe interactions in a community context and measured phenotypes have proven difficult to establish. This limitation is, in fact, generally true across complex host–microbial interaction systems [12][13][36]. Here, we demonstrate that binary-association assays can inform the design of synthetic bacterial communities that lead to predictable plant phenotypes, an observation seen only once previously, in one animal system [12]. The host phenotypic output of the bacterial synthetic communities was consistent with the output expected from binary interactions. Validation of predictions from an NN confirmed that we could predictably alter certain plant phenotypes by changing the plant’s microbiota membership.
Other phenotypes and host–microbiota systems can likely be studied with this approach. The only requirements are that a microcosm reconstitution system is available and that functional bacterial blocks can be defined, so that synthetic communities that maximize the expected range of phenotypic variance can be constructed (Fig 3A and 3B). In practice, other aspects that are likely to influence the tractability of a system are the functional bacterial diversity and the SNR of the phenotypes being measured. In the case of plant phosphate starvation, we found that bacterial abundances provided no information, and while it is too early to say if this is a general feature, the only other work that directly manipulated a well-defined microbiota to establish its effect on a host phenotype reached a similar conclusion [12]. If this trend continues across other host–microbiota systems, then our approach has the added advantage that strains need not be distinguishable by a specific marker gene. While a simple additive model typically explained more than 50% of the host phenotypic variation, we found value in utilizing an NN approach that was able to capture more complex relationships but remained interpretable and significantly increased our prediction accuracy for novel communities. Our framework is based on empirical validation and thus remains flexible enough to allow for simpler or more complex models, depending on the case.
We achieved high prediction accuracy across an untested set of synthetic communities, thus demonstrating that selecting a subset of the possible communities by partial overlap of bacterial functional block pairs is sufficient to characterize this system. This method requires no design of specific heuristics. Thus, this methodology provides an opportunity to expand the capacity for mechanistic understanding not only of biological networks that control plant phenotypes [37][14] but of other complex ecological systems [12][13][36].
Furthermore, by focusing on block replacements as testable hypotheses, we provide a simple outcome that can be extracted from both linear and nonlinear predictive models. This can guide the next set of experimental designs, thus providing nonlinear methods like deep learning a stronger empirical grounding, rendering them less of a “black-box.”
We demonstrate the utility of our approach by defining mechanistic aspects of the plant phosphate stress response in the presence of combinations of bacterial blocks. We observed that bacteria range in their effect on phosphate content in the plant between severely decreasing and moderately enhancing it. These data are consistent with our previous findings that bacterial interactions with the plant are controlled by negative regulation exerted by the phosphate starvation response on the plant immune system [23]. A similar mechanism was described for the interaction under phosphate-limiting conditions of A. thaliana with the beneficial fungus Colletotrichum tofieldiae [24]. Thus, our results provide additional evidence for mechanisms by which plants and bacteria compete in times of nutritional stress.
The use of multiple bacterial synthetic communities led us to define interesting particular aspects of the phosphate stress response. We observed that certain synthetic communities, such as P2P3 and P1P2, drive an increase in the shoot area compared to axenically grown plants, despite the low shoot Pi content that they engender. These data recapitulate a previous observation [31] on the effect of altering the activity of PHOSPHATE1 (PHO1), a gene required for Pi loading into the xylem [38]. These authors found that shoot Pi content could be uncoupled from the developmental responses typically linked to Pi scarcity. We corroborated that reduced shoot growth is not necessarily a direct consequence of Pi limitation. The observations that both bacterial activity and the modulation of PHO1 expression can uncouple plant phenotypes during the response to low Pi leads us to hypothesize that microbes could interdict PHO1 transport activity, thus modifying Pi translocation from roots to shoots.
Additionally, we found that synthetic community P3N3 uniquely induced a strong transcriptional response to phosphate starvation in the majority of the conditions tested. Plants exposed to this bacterial combination showed a high-level induction of PHO2, a ubiquitin-conjugating E2 enzyme required for the degradation of Pi transporters at high Pi [34]. This discovery may explain the intense transcriptional response to Pi starvation caused by this particular synthetic community.
We observed much more variability in the bacterial colonization patterns than in their effects on plant phenotypes. Synthetic communities tended to be dominated by 1 block, but the identity of that block did not correlate with plant phenotype. On the other hand, synthetic communities had remarkably consistent effects on plant phenotypes, and synthetic community membership was sufficient to predict host phenotype. These observations suggest that bacteria–bacteria interactions are critical for microbial community assembly, which is probably a highly dynamic process in which the microbial background determines which bacteria perform well. On the other hand, the effect of bacteria on plant phenotypes is probably due to functional stacking, in which many phenotypically redundant strains with potentially different niches maximize the chance of attaining the desired host phenotypic output. This “lottery model” has been proposed as a major driving mechanism of host colonization by its microbiota at the taxonomic level [39], and it would be interesting to test whether a similar process governs functional assembly.
In conclusion, we provide a general method for the study of various biological host–microbiome systems through rational selection of a tractable subset of the possible combinations of bacteria from a defined collection. We demonstrate that complex relationships among host phenotypes, the microbiota, and the abiotic environment can be captured using deep learning techniques. By testing each block of bacterial strains in multiple synthetic communities, and successfully validating predictions derived from an NN, we demonstrated that it is possible to both infer causality and attain generality when it comes to predicting host phenotypes in this complex system. This approach contributes to the rational design and deployment of microbes to improve responses of hosts to biotic and nutritional stresses.
|
10.1371/journal.pgen.1006513 | Protein Phosphatase 6 Protects Prophase I-Arrested Oocytes by Safeguarding Genomic Integrity | Mammalian oocytes are arrested at prophase of the first meiotic division in the primordial follicle pool for months, even years, after birth depending on species, and only a limited number of oocytes resume meiosis, complete maturation, and ovulate with each reproductive cycle. We recently reported that protein phosphatase 6 (PP6), a member of the PP2A-like subfamily, which accounts for cellular serine/threonine phosphatase activity, functions in completing the second meiosis. Here, we generated mutant mice with a specific deletion of Ppp6c in oocytes from the primordial follicle stage by crossing Ppp6cF/F mice with Gdf9-Cre mice and found that Ppp6cF/F; GCre+ mice are infertile. Depletion of PP6c caused folliculogenesis defects and germ cell loss independent of the traditional AKT/mTOR pathway, but due to persistent phosphorylation of H2AX (a marker of double strand breaks), increased susceptibility to DNA damage and defective DNA repair, which led to massive oocyte elimination and eventually premature ovarian failure (POF). Our findings uncover an important role for PP6 as an indispensable guardian of genomic integrity of the lengthy prophase I oocyte arrest, maintenance of primordial follicle pool, and thus female fertility.
| Formation of haploid gametes from diploid germ cells requires a specialized reductive cell division known as meiosis. In contrast to male meiosis that takes place continuously, a unique feature of female meiosis in mammals is the long arrest in meiosis I, which lasts up to 50 years in humans. Because the size of the germ cell pool determines the reproductive lifespan of females, it is important to discover mechanisms preserving the germ cell pool during the lengthy meiotic arrest. In this study, we examined the physiological role of a member of the PP2A-like serine/threonine phosphatase subfamily, protein phosphatase 6, in mouse oocytes during ovarian follicular development. This is the first study linking PP6 to the maintenance of the female germ cell pool and fertility. We find PP6 is an indispensable protector of arrested oocytes by safeguarding genomic integrity during their dormancy in the mouse ovary.
| In mammals, females are born with a finite number of oocytes contained within primordial follicles that serve as the source of ova for the entire period of reproductive life. To produce mature eggs, dormant primordial follicles are recruited into the growing follicle pool, a process termed as initial follicular recruitment or activation. Activated follicles subsequently develop into primary follicles, secondary follicles, and antral follicles [1]. Throughout this follicular growth process, oocytes grow while being arrested in prophase of meiosis I with homologs held together by chiasmata. Only a few dominant antral follicles reach the preovulatory stage and release a mature egg for fertilization after a gonadotropin surge during each estrus cycle [2]. When the ovarian follicle reserve is exhausted in women, menopause occurs. However, disorders during folliculogenesis could lead to follicle depletion in advance and cause premature ovarian insufficiency (POI) or premature ovarian failure (POF), which is a main cause of female infertility in humans and affects nearly 1% women under the age of 40 [3].
Protein phosphorylation, mediated by a conserved cohort of protein kinases and phosphatases, regulate follicular activation and growth, meiotic cell cycle arrest and progression, chromosome dynamics, and ovulation [4]. Numerous studies using genetically modified mice reveal that protein kinases play important roles during folliculogenesis/oogenesis. For example, the PTEN/PI3K/AKT signaling pathway regulates follicular activation and survival [5]. Recently, we reported that LKB1 acts as a gatekeeper of the ovarian primordial follicle pool [6]. In contrast, there is limited information about the roles of protein phosphatases. Among the serine/threonine phosphoprotein phosphatases (PPPs), PP2A, PP4 and PP6 form a subfamily called PP2A-like protein phosphatases, which share high homology in the catalytic subunit and account for the majority of cellular serine/threonine phosphatase activity [7, 8]. PP2A is involved in regulating chromosome condensation, DNA damage repair, G2/M transition and sister chromatid cohesion [9]. Our recent knockout mouse model revealed that oocyte PP2A is dispensable for folliculogenesis, though PP2A has been reported to dephosphorylate AKT and AMPK, important kinases for folliculogenesis [10]. Although PP6 was discovered almost 20 years ago, progress has been slow regarding its functions in cells, not to mention its specific functions in meiotic cells.
The PP6 holoenzyme consists of a catalytic subunit, PP6c, one of the three regulatory subunits including SAPS1, 2, 3 (also known as PP6R1, PP6R2 and PP6R3, respectively), and one of the three ankyrin repeat subunits including ARS-A, -B, -C [11, 12]. PP6 is conserved among all eukaryotic species from yeast to humans, attesting to its fundamental importance. Mutations in PP6c are found to exist in 9–12.4% melanomas surveyed and may act as drivers for melanoma development [13, 14]. The PP6 yeast homologue, Sit4/Ppe1, is required for G1/S progression and equal chromosome segregation [15, 16], and plays a role in signaling through the target of rapamycin (TOR), a key nutrient-sensing kinase [17]. Human PP6 has an established role in DNA damage response with its ability to modulate signaling by DNA-dependent protein kinase (DNA-PK), homology recombination-mediated repair of DNA double strand breaks (DSBs) [18, 19], as well as its interactions with Aurora A kinase [20, 21]. More recent studies suggest a broader role for PP6 in pre-mRNA splicing [22], control of apoptosis in immune cells [23], formation of adherens junctions through interaction with E-cadherin [24], and modulation of signaling through the Hippo pathway [25]. Overall, these data suggest that PP6 integrates signaling from multiple pathways.
Genetically modified mouse models are powerful tools for studying gene function in vivo [26, 27]. We recently reported that a conditional knockout of PP6 in oocytes from growing follicles (by crossing Ppp6cF/F mice with Zp3-Cre mice) causes female subfertility by disrupting MII spindle organization and MII completion after fertilization [28]. Here, we crossed Ppp6cF/F mice with Gdf9-Cre mice to generate mutant mice with specific deletion of Ppp6c in prophase I-arrested oocytes from the primordial follicle stage. We find that PP6 plays a critical role in germ cell survival and follicular development by safeguarding genomic integrity of prophase I-arrested oocytes.
To explore the in vivo roles of PP6 during folliculogenesis/oogenesis, we generated mutant mice (referred to as Ppp6cF/F;GCre+ mice), in which exon II-IV of the Ppp6c gene were targeted, by crossing Ppp6cF/F mice [28] with transgenic mice expressing Gdf9 promoter-mediated Cre recombinase [5] (Fig 1A). In Gdf9-Cre mice, Cre is specifically expressed in oocytes of primordial follicles and later stage follicles since postnatal day 3 [27]. By immunoblotting analysis, we confirmed successful depletion of PP6c protein in GV oocytes from Ppp6cF/F;GCre+ females (Fig 1B).
To investigate the effect of oocyte-specific knockout of PP6c on female fertility, a breeding assay was carried out by mating Ppp6cF/F or Ppp6cF/F;GCre+ female mice with males of proven fertility for 6 months. As shown in Fig 1C, female Ppp6cF/F;GCre+ mice were completely infertile. The infertility appeared to be due to anovulation in adult mutant mice, whereas control mice ovulated normal numbers of eggs (8.7±0.8) in the natural ovulation assays (Fig 1D).
To understand the defects of the mutant mice, we first observed the morphology of ovaries from both Ppp6cF/F and Ppp6cF/F;GCre+ mice. At 1 month-of-age, both histological morphology of ovaries and numbers of follicles were similar between Ppp6cF/F and Ppp6cF/F;GCre+ ovaries (S1A and S1B Fig), indicating that comparable numbers of follicles are formed in the wild-type and mutant ovaries. However, after 1 month-of-age, the time of onset of sexual maturity, mutant ovaries started to show differences and became smaller than the controls. In ovaries of 2-month-old Ppp6cF/F;GCre+ mice, there were few growing follicles (Fig 2B) in contrast to control ovaries that contained many healthy-looking growing follicles (Fig 2A). However, large clusters of primordial follicles could still be observed on the ovarian surface area of 2-month-old mutant ovaries (white arrowheads, Fig 2C and 2C’), compared with control ovaries where such clusters barely could be found. Consistently, the number of primordial follicles in 2-month-old Ppp6cF/F;GCre+ ovaries was more than double that of Ppp6cF/F ovaries (Fig 2M and S1C Fig). The numbers of large growing follicles, especially Type 5 and Type 6 follicles, were significantly decreased, corresponding to only 16.7% and 36.5% of those in control ovaries (S1C Fig). At 3 months-of-age, clusters of primordial follicles were no longer observed on the ovarian surface; instead, many primary follicles appeared in the same location, indicating delayed activation of the arrested primordial follicles (yellow arrows, Fig 2F and 2F’). Consistent with these observations, quantification of ovarian follicles revealed a significant reduction of primordial follicles and an increase in type 3 and type 4 follicles in 3-month-old mutant ovaries (S1D Fig). Nevertheless, large growing follicles (including type 5 and type 6 follicles) were significantly fewer than those in control ovaries (S1D Fig), though both control and mutant ovaries might contain similar numbers of primordial follicles and activated follicles (Fig 2M and 2N). These later activated follicles, however, could not serve as the source of ova for mutant mice, probably because they died soon after activation with only empty follicle-like structures left at the ovarian surface (yellow arrowheads, Fig 2I, 2I’, 2L and 2L’). At 4 months-of-age, only a few primary follicles and small secondary follicles were seen at the cortical region of mutant ovaries (Fig 2H), and the other types of follicles (including primordial follicles, type 5,6 and 7 follicles) were disappearing (S1E Fig). By 6 months-of-age, almost all types of follicles were depleted in Ppp6cF/F;GCre+ ovaries (Fig 2K and 2L; S1F Fig), which is termed POF. In general, from 1 month to 2 months postpartum, more than half of the primordial follicles in the Ppp6cF/F ovaries decreased due to both follicular activation and atresia. In contrast, loss of primordial follicles in Ppp6cF/F;GCre+ ovaries was slower because they failed to be activated upon puberty and stayed arrested until 2 months postpartum, after which time they were rapidly eliminated either through death following delayed activation or degeneration (Fig 2M). Activated follicles in Ppp6cF/F;GCre+ ovaries only survived for a short time and none could develop to the preovulatory stage (Fig 2N).
The histological analysis suggested that absence of PP6c in oocytes caused defects in follicular activation and growth. To confirm these observations, we performed immunostaining of the germ cell marker MVH (mouse VASA homolog) on 2-month-old ovarian sections. As shown in S2A Fig, in normal control ovaries, primordial follicles were mostly scattered around the cortical region whereas in ovaries of adult Ppp6cF/F;GCre+ mice a significant number of primordial follicles remained in clusters (S2B Fig), indicating abnormal development of primordial follicles. This finding confirmed that the natural incidence of follicular activation after puberty was disrupted by Gdf9-Cre mediated Ppp6c deletion. At 2 months-of-age, although Ppp6cF/F;GCre+ mice still had a large number of growing follicles, these follicles failed to mature and ovulate. As shown by TUNEL assay on ovarian sections, increased granulosa cell apoptosis and follicle atresia (yellow arrowheads, S2D Fig) were detected in ovaries of 2-month-old Ppp6cF/F;GCre+ mice compared to ovaries in control mice (S2C Fig). Furthermore, when we tried to stimulate follicle growth with exogenous PMSG, the mutant mice still could not respond normally because almost all the antral follicles initiated atresia by premature luteinization and formed numerous atretic corpora lutea (CLs) (yellow arrowheads, S2F Fig) instead of developing into preovulatory follicles (red asterisks, S2E Fig). The above data demonstrated that defective follicular development after puberty, including blocked primordial follicle activation and compromised growth of activated follicles, accounted for the infertility of Ppp6cF/F;GCre+ mice.
mTOR signaling is essential for oocyte survival and awakening from dormancy within primordial follicles [5, 29, 30]. Considering an analogous involvement of PP6 in TOR signaling in yeast and plants [17, 31], it is possible that PP6 maintains oocyte survival by regulating the mTOR pathway. Accordingly, we performed immunoblotting analysis with PD35 GV oocytes. Surprisingly, the activity of the AKT/mTORC1/S6K signaling pathway was significantly enhanced, as indicated by elevated levels of phosphorylated AKT (S473), phosphorylated mTOR (S2448), phosphorylated S6K (T389) in Ppp6cF/F;GCre+ oocytes (Fig 3A and S3B Fig); phosphorylated rpS6 (S240/244) did not show obvious changes (Fig 3A and S3B Fig). This finding was not consistent with our phenotypes based on previous reports because enhanced AKT/mTOR signaling is responsible for the over-activation of primordial follicles in Pten and Tsc1/2 mutant mouse models. In contrast, our Ppp6c mutant mice did not show any signs of premature activation of the entire primordial follicle pool, instead showing blockage/delay of follicular activation, although the activity of AKT/mTOR signaling was higher than in controls. Thus, up-regulation of the AKT/mTOR pathway could result from feedback effects to defective oocyte growth or local effects attributed to PP6 in regulating mTOR activity as suggested for other organisms [17, 31].
Recently, we reported that Lkb1fl/fl; Gdf9-Cre mice exhibit over-activation of primordial follicles starting from the onset of sexual maturity and defective follicle growth at later stages. The phenotypes of Lkb1fl/fl; Gdf9-Cre mice appear opposite to those of Ppp6cF/F;GCre+ mice. Accordingly, examined the activity of AMPK, the main substrate of LKB1, in Ppp6cF/F;GCre+ oocytes and observed that the level of phosphorylated AMPK (T172) was significantly increased (Fig 3B and S3B Fig); it is decreased in Lkb1 mutant oocytes. To ascertain whether PP6 interacts with the AMPK pathway we generated double knockout mice for both Lkb1 and Ppp6c (Lkb1F/F;Ppp6cF/F;GCre+). As expected, Lkb1 deletion within a Ppp6c deletion background rescued the blockage of primordial follicle activation at 2 months-of-age (Fig 3E). Unanticipated was that double knockout ovaries resembled Lkb1 mutant ovaries (Fig 3C) by exhibiting large sizes and over-activation of primordial follicles at 2 months-of-age. One difference, however, was that growth of activated follicles in double knockout ovaries was slower with secondary follicles containing unhealthy oocytes and showing apoptotic signals indicating extensive follicle atresia at 2 months-of-age (yellow arrowheads, Fig 3E), which is similar to 2-month-old Ppp6cF/F;GCre+ ovaries (Fig 3D). In contrast, most activated follicles reached the antral follicle stage in Lkb1 mutant ovaries at the same age (Fig 3C). These results showed that knockout of Lkb1 could partially rescue the follicle development phenotype of the PP6c mutant ovaries, which strongly suggested involvement of AMPK in follicle development and PP6c participating in regulating the AMPK pathway. However, knockout of Lkb1 did not rescue Ppp6cF/F;GCre+ oocytes from death, suggesting that PP6 might not only regulate primordial follicle activation but also maintain survival of oocytes within primordial follicles. Therefore, we concluded that there were additional reasons for the PP6c mutant phenotype and pursued this possibility as described below.
Because PP6 is involved in the DNA damage response via its ability to dephosphorylate γH2AX and antagonize DNA-dependent protein kinase (DNA-PK) [18, 19] and unrepaired meiotic or induced DNA double-strand breaks (DSBs) could cause oocyte elimination and female infertility by triggering DNA damage response pathway [32], we wondered if loss of PP6c leads to DNA damage in our case. Thus, we collected oocytes from PD35 ovaries and performed western blot analysis. As shown in Fig 4A and S3 Fig, the levels of γH2AX were significantly elevated in mutant oocytes indicating accumulated DSBs. However, the DNA damage response pathway was significantly reduced because the activity of CHK1/2-p53 signaling cascade was much lower than in controls (Fig 4A and S3 Fig). We also confirmed accumulation of γH2AX in small oocytes by immunofluorescence analysis. As indicated in Fig 4B, mutant ovaries contained more and higher nuclear signals of γH2AX within primordial follicles (yellow arrows) when compared to controls (white arrows). Oocyte maturation in vitro of mutant oocytes was also compromised. As shown in Fig 4C, the incidence of GVBD (56.7±7.9%) and PBE (48.7±14.2%) were lower than controls (76.6±2.0%; 84.2±2.6%, respectively). Moreover, after 8 h of in vitro maturation of Ppp6cF/F;GCre+ oocytes spindles were disorganized with scattered chromosomes, in contrast to the well-organized MI spindles with chromosomes all aligned at the equatorial plate in Ppp6cF/F oocytes. Even after 13 h of in vitro culture, when control oocytes had extruded the first polar body, most mutant oocytes still showed defective spindle organization and aberrant chromosome alignment, and could not complete meiosis I successfully. Taken together, these data demonstrate that loss of PP6c resulted in DSBs accumulation and severely impaired oocyte quality but deactivated the DNA damage response pathway in oocytes until puberty, which could explain why primordial follicle activation is delayed and mutant oocytes are damaged but still survive until 2 months postpartum.
Depleting PP6c sensitizes cells to induced DNA damage [19, 33]. Thus, we speculated that this susceptibility also exists in our PP6c-deficient oocytes and leads to eventual oocyte elimination. Because endogenous DNA damage might be low and long-term, mutant oocytes would wait for repair first and then die slowly within 6 months postpartum. Experimentally increasing DNA damage in mutant oocytes could therefore trigger more rapid apoptosis and accelerate oocyte elimination if PP6c-deficient oocytes are defective in mounting a DNA damage response.
To test this proposal, zeocin was used to induce DSBs in vivo by intraperitoneal injection. Forty mg of zeocin was injected per mouse once every day for 5 days [34] after which the mice were allowed 5 days of recovery and then sacrificed around 2 months postpartum. Typically, Ppp6cF/F;GCre+ ovaries were smaller than the Ppp6cF/F ones at 2 months-of-age; however, after zeocin treatment, mutant ovaries were even smaller whereas the size of control ovaries was similar to untreated ones (Fig 5A), suggesting that oocyte elimination was faster after zeocin treatment. This conclusion was confirmed by histological analysis of ovaries and follicle counting. After zeocin treatment, the number of primordial follicles in mutant ovaries (~57%) as well as the number of activated follicles (~44%) decreased dramatically compared to those of untreated mutant ovaries; in control groups, treated ovaries also showed fewer numbers of follicles compared to untreated ones, but these changes were not significant (Fig 5B). Consistently, as shown in Fig 5C, treated mutant ovaries contained more atretic follicles (yellow arrows), with many primordial follicles devoid of oocytes (yellow arrowheads), whereas control ovaries had plenty of healthy-looking growing follicles (white arrows) and primordial follicles (white arrowheads). The above observations showed that in response to induced DNA damage, Ppp6cF/F;GCre+ ovaries showed no primordial follicle arrest but oocyte death and follicle depletion, indicating that PP6c-deficient oocytes were more sensitive to DNA damage.
To investigate the molecular causes for the results described above, we performed in vitro zeocin treatment in PD35 GV oocytes. GV oocytes were treated with zeocin (200 μg/ml for 1 h), then washed and cultured in M2 medium containing 2.5 μM milrinone overnight for recovery. These GV oocytes were collected for western blot analysis. As shown in Fig 6A and S4 Fig, in comparison to Ppp6cF/F oocytes after treatment, Ppp6cF/F;GCre+ oocytes showed lower levels of γH2AX but a highly active CHK1/2-dependent DNA damage checkpoint response, with p53-induced cell apoptosis. We also performed in vivo zeocin treatment in young mice (5 days of zeocin injection and 5 days of recovery) and collected ovaries for western blot at ~PD35 when mutant ovaries still had similar numbers of follicles as controls (Fig 6B and S4 Fig). The levels of MVH, a marker of germ cells, were similar in both groups indicating mutant ovaries still contained comparable numbers of oocytes to controls. Mutant ovaries, however, showed an enhanced CHK2-p53 DNA damage response pathway activity, suggesting PP6c-deficient oocytes could not repair induced DNA damage and would die eventually. Based on the above results, the main cause for the PP6c depletion phenotype appeared to be an increased susceptibility to DNA damage of PP6c-deficient oocytes.
Collectively, these findings support the notion that PP6 is a critical regulator for oocyte survival and follicle development by restraining phosphorylation of H2AX to normal levels and participating in AMPK pathway regulation.
In female reproduction, production of high quality eggs requires both successful follicular development and precise completion of oocyte meiosis. Previously, we studied the roles of PP6c in meiosis completion by crossing Ppp6cF/F mice with Zp3-Cre mice. By crossing Ppp6cF/F mice with Gdf9-Cre mice to generate mutant mice with a specific deletion of Ppp6c in oocytes from the primordial follicle stage we were able to investigate the roles of PP6c in follicular development. We find that Ppp6c mutant female mice show defective folliculogenesis and are infertile. Importantly, PP6c depletion caused persistent phosphorylation of H2AX. Thus, susceptibility to DNA damage and defective DNA repair mechanisms turned out to be the main underlying causes for the observed infertility. In addition, PP6c may control follicular activation by regulating the AMPK pathway.
During embryonic development, primordial germ cells in female mammals enter meiosis I and finish a crucial process called synapsis that requires homologous recombination (HR), a high-fidelity DNA double-strand break (DSB) repair process. Aberrant homolog synapsis or DSB repair triggers checkpoints that eliminate defective meiotic oocytes [35–37]. Loss of oocytes defective in DSB repair occurs soon after birth, which is controlled by the DNA damage checkpoint including the CHK2-p53/p63 pathway [32]. Oocytes are subsequently arrested at the dictyate stage of prophase I in the form of dormant oocytes enclosed in primordial follicles [38]. Such prophase I arrest usually takes weeks or months, or even longer in mice, and after primordial follicular activation, undergo a prolonged period of follicular growth before meiosis resumption and ovulation [39, 40]. The lengthy dormancy and growth of oocytes makes maintenance of genomic integrity during follicular development more challenging and important for generating healthy gametes. However, the underlying molecular mechanisms to protect genomic DNA after embryonic HR and DSB repair remained undiscovered. The DNA damage checkpoint usually acts around the time oocytes enter meiotic arrest but presumably persists, because resting primordial follicles are highly sensitive to ionizing radiation (IR) [41]. In our study, oocyte-specific knockout of PP6c from primordial follicle stages results in increased γH2AX in arrested oocytes and the whole germ cell pool is then progressively eliminated by DNA damage checkpoint pathway within 6 months postpartum. These findings make PP6 a competitive candidate for safeguarding genomic DNA integrity of female germ cells during the long prophase I arrest.
As noted above, PP6 is implicated in the cell response to DNA damage. The phosphorylated form of H2AX on S139 (γH2AX) is a marker of DSBs. PP6c exhibits phosphatase activity against γH2AX in in vitro phosphatase assays. In human cancer lines, depletion of PP6c or PP6R2 leads to persistent high levels of γH2AX after DNA damage and defective homology-directed repair (HDR) [19]. PP6c is recruited to DSB sites by DNA-PK, and PP6 is also required for efficient activation of DNA-PK, which is essential for non-homologous end joining (NHEJ)-mediated repair of DSBs [18, 42]. A recent study also showed that Ppp6c-deficient mouse keratinocytes exhibit a high frequency of both p53- and γH2AX-positive cells, suggestive of DNA damage, as well as up-regulated expression of p53, PUMA, BAX, and cleaved caspase-3 proteins following UVB irradiation [33]. Our in vivo data show that absence of PP6c also leads to higher levels of γH2AX (Fig 4) and defective DNA repair in oocytes, especially massive oocyte death after induced DNA damage (Figs 5 and 6), suggesting that PP6 has a conserved role in DNA damage response, which is essential for gamete production and fertility maintenance.
As members of the well-known PP2A-like subfamily, PP6 shares common features with PP2A or PP4. As phosphatases, they all are involved in a diverse set of biological pathways due to their wide range of substrates. Until now, PP6 was implicated in regulation of DNA damage response, cell cycle progression, apoptosis, pre-mRNA splicing, signaling through the mTOR pathway and Hippo pathway, and others [19, 22, 23, 25, 28, 33]. Among these multiple functions, mTOR pathway regulation was first considered as the potential cause of the phenotype in our study. mTOR signaling regulates follicular activation and oocyte survival because oocyte-specific deletion of its upstream genes, Pten or Tsc1/2, lead to premature activation of the entire primordial follicle pool, resulting in POF due to enhanced mTORC1-S6K-rpS6 signaling [5, 29, 30]. Although PP6c-deficient oocytes also show similar enhanced AKT/mTOR signaling, the ovarian phenotype of Ppp6c mutant mice is not similar at all, because PP6c mutant ovaries show blocked/delayed follicular activation instead of premature activation, also at later time points.
Although the AKT/mTOR pathway is activated in PP6c mutant ovaries, primordial follicles are not activated, perhaps because the downstream effectors of mTOR pathway are not responding. As seen from the western blot results (Fig 3A), the activities of the AKT/mTORC1/S6K signaling are significantly enhanced in Ppp6cF/F;GCre+ oocytes, but as the downstream effector that enhances protein translation, rpS6 does not show an obvious change of activity. Thus, the effects of AKT/mTOR pathway activation are somehow blocked at the execution phase and therefore do not activate primordial follicles in mutant ovaries. In light of these findings, we turned to another important folliculogenesis regulator, the LKB1-AMPK pathway.
Our previous study reported that Lkb1 mutant female mice show over-activation of primordial follicles after puberty [6], at a similar time point as that of Ppp6c mutant mice. Because Western blot results also show up-regulated p-AMPK in PP6c-deficient oocytes, opposite to that in LKB1-deficient oocytes, we generated double knockout of Lkb1 and Ppp6c in oocytes to try to rescue the phenotypes of Ppp6c mutant mice. Indeed, the blocked/delayed follicular activation was rescued, which means mis-regulation of AMPK pathway could be a partial reason for the PP6c mutant phenotypes. Moreover, the double knockout ovaries show over-activation of primordial follicles, more similar to Lkb1 single knockout, but accelerated oocyte death and slower follicle growth, suggesting that absence of PP6c might affect oocyte quality and survival more directly than just control follicular activation. Thus, PP6’s role in DNA damage response could be the main cause. Consistent with this proposal is that PP6c depletion caused increased γH2AX, a marker of DSBs, and defective DNA repair in oocytes, with accelerated oocyte death with induced DNA damage. Interestingly, oocyte defects resulting from PP6c depletion are relatively low in natural circumstances, and oocyte death occurs only when both endogenous and exogenous harm accumulated with time to a certain degree, which could explain why the whole oocyte elimination process took up to 6 months in Ppp6c mutant ovaries. Thus, PP6c could control oocyte quality through its role in DDR pathway as well as regulate follicular activation through participating in the AMPK pathway. Nevertheless, we cannot exclude other possibilities, e.g., regulation of pre-mRNA splicing and Hippo pathway, that could also contribute to the phenotypes in our mutant mouse model.
Female meiosis is error-prone in humans. Our previous study reported that Zp3-Cre mediated PP6c depletion in growing oocytes leads to defective MII spindle function and unfaithful chromatid segregation in meiosis II without affecting folliculogenesis, indicating that PP6 can act as antagonizer to oocyte aneuploidy during the MII exit. Here we demonstrate that Gdf9-Cre mediated PP6c depletion in dormant oocytes causes defective folliculogenesis and massive germ cell elimination at early stages, indicating that PP6 can also safeguard oocyte genomic integrity and regulate folliculogenesis during the long prophase I arrest. Furthermore, isolated GV oocytes from Ppp6cF/F;GCre+ mice before POF occurs show severely impaired in vitro maturation because of DNA damage, in sharp contrast to the unaffected meiotic maturation progress of Ppp6cF/F;ZCre+ oocytes. Although these two knockout mouse models are both oocyte-specific knockouts, they exhibit completely different phenotypes that presumably reflect differences between timing of Zp3-Cre and Gdf9-Cre expression. Both ZP3 and GDF9 are specifically expressed in oocytes. The synthesis of ZP3 starts in primary follicles from PD5, reaches a maximum in growing follicles, and decreases in full-grown oocytes, which makes Zp3-Cre only suitable for deletion of gene expression in oocytes from primary follicle stages on. However, Gdf9-Cre is expressed in oocytes from primordial follicle stage. This difference in expression is presumably why Ppp6cF/F;GCre+ mice display primordial follicle defects whereas Ppp6cF/F;ZCre+ mice do not.
In summary, we provide evidence that PP6 acts as a critical guard of genomic integrity in lengthy prophase I arrest of oocytes and is an indispensable regulator of folliculogenesis, and thus female fertility. Our data may provide valuable information for the design of therapeutics for POF.
Animal care and handling were conducted according to the guidelines of the Animal Research Committee of the Institute of Zoology, Chinese Academy of Sciences. The institutional committee which is licensed by Beijing Municipal Experimental Animal Administration approved this study.
Mice lacking Ppp6c in oocytes (referred to as Ppp6cF/F;GCre+) were generated by crossing Ppp6cF/F mice [28] with Gdf9-Cre mice. Both transgenic mouse lines have C57BL/6J genomic background. The mice were housed under controlled environmental conditions with free access to water and food. Light was provided between 08:00 and 20:00.
Commercial antibodies were used to detect PPP6C (rabbit, A300-844A; Bethyl Laboratories, Inc.), α-tubulin (mouse, DM1A; Sigma-Aldrich), MVH (rabbit, ab13840; Abcam), γH2AX (rabbit, 9718; Cell Signaling Technology, Inc.), p-CHK1 (S345) (rabbit, BS4041; Bioworld Technology, Inc.), p-CHK2 (T68) (rabbit, BS4043; Bioworld Technology, Inc.), p-p53 (S15) (rabbit, 12571; Cell Signaling Technology, Inc.), CHK1 (rabbit, BS1052; Bioworld Technology, Inc.), p-AKT (S473) (rabbit, 4060; Cell Signaling Technology, Inc.), p-AMPK (T172) (rabbit, 2535; Cell Signaling Technology, Inc.), p-mTOR (S2448) (rabbit, 5536; Cell Signaling Technology, Inc.), p-S6K (T389) (rabbit, 9234; Cell Signaling Technology, Inc.), p-rpS6 (S240/244) (Rabbit, 5364; Cell Signaling Technology, Inc.), GAPDH (rabbit, 5174; Cell Signaling Technology, Inc.) and β-actin (mouse, sc-47778, Santa Cruz). Secondary antibodies were purchased from ZhongShan Golden Bridge Biotechnology Co., LTD (Beijing).
Ovaries used for histological analysis were collected from adult female mice. They were fixed in 4% paraformaldehyde (pH 7.5) overnight at 4°C, dehydrated, and embedded in paraffin. Paraffin-embedded ovaries were sectioned at a thickness of 8-μm for hematoxylin and eosin (H&E) staining. One or both ovaries from more than three mice of each genotype were used for the analysis. Paraffin-embedded ovarian tissue sections were deparaffinized, immersed in retrieval solution (10 mM sodium citrate), heated in an autoclave, blocked with 10% normal goat serum, and then incubated overnight with primary antibodies (anti-MVH and anti-γH2AX at 1:200 dilution). For immunofluorescence, localization of the primary antibody was performed by incubation of the sections with the corresponding secondary antibodies (Invitrogen) at 1:500 dilution for 1h at room temperature. Finally, nuclei were stained with DAPI. For immunohistochemistry, the Vecta stain ABC kit (Vector Laboratories, CA, USA) was used to detect the signal of primary antibody. Analysis of apoptosis in ovarian follicles was carried out by TUNEL assay using the ApopTag Plus in situ apoptosis detection kit (Chemicon International, Temecula, CA, USA). At least three different samples from each genotype were analyzed in parallel.
Oocytes for immunofluorescent staining were fixed in 4% paraformaldehyde in PBS for 30 min at room temperature. The fixed oocytes were then transferred to membrane permeabilization solution (0.5% Triton X-100) for 20 min and blocking buffer (1% BSA-supplemented PBS) for 1 h. The oocytes were then incubated overnight at 4°C with FITC conjugated anti-α-tubulin at 1:2000 dilution. Nuclei were stained with DAPI. Finally, oocytes were mounted on glass slides and examined with a laser scanning confocal microscope (Zeiss LSM 780 META, Germany).
Quantification of ovarian follicles was performed as previously described [43]. Briefly, to count the numbers of follicles, paraffin-embedded ovaries were serially sectioned at 8-μm thickness and every fifth section was mounted on slides. Then these sections were stained with hematoxylin and eosin for morphological analysis. Ovarian follicles at different developmental stages, including primordial, primary (type 3 and type 4), secondary (type 5) and antral follicles (type 6 and type 7) were counted in collected sections of an ovary, based on the well-accepted standards established by Peterson and Peters [44]. In each section, only those follicles in which the nucleus of the oocyte was clearly visible were scored and the cumulative follicle counts were multiplied by a correction factor of 5 to represent the estimated number of total follicles in an ovary.
For the natural ovulation assay, 2–4 month-old female mice were mated with fertile males overnight. Successful mating was confirmed by the presence of vaginal plugs. Fertilized eggs were harvested from oviducts, counted and analyzed after removal of the cumulus mass with 1mg/ml hyaluronidase (Sigma-Aldrich) in M2 medium (Sigma-Aldrich).
GV stage oocytes were isolated from ovaries of ~PD35 female mice and cultured in M2 medium under paraffin oil at 37°C, 5% CO2 in air.
For in vitro treatment of zeocin, fully-grown GV oocytes were first treated with zeocin (200 μg/ml, Invitrogen) for 1 h in M2 medium supplemented with 2.5 μM milrinone and then blocked by the same concentration of milrinone for recovery. Oocytes were collected after 12 hours of recovery for western blot.
Ovary lysate was prepared from minced ovaries after removal of suspended granulosa cells by centrifugation for western blot analysis. Thirty μg ovary protein or 200 oocytes were mixed with SDS sample buffer and boiled for 5 min at 100°C for SDS-PAGE. Western blot was performed as described previously [45], using antibody dilutions as below, antibodies against PPP6C, MVH, γH2AX, p-CHK1 (S345), p-CHK2 (T68), CHK1 at 1:500, antibody against p53, p-p53 (S15), p-AKT (S473), p-AMPK (T172), p-mTOR (S2448), p-S6K (T389), p-rpS6 (S240/244) at 1:1000, and antibodies against GAPDH and β-actin at 1:2000.
To induce DNA DSBs in vivo, zeocin was injected into the abdominal cavity of female mice once every day for 5 days, and physiological saline (vehicle) was injected as control. Zeocin (100 mg/ml, Invitrogen) was diluted in physiological saline to give a final concentration of 400 mg/ml, and 0.1 ml (40 μg zeocin) was injected per mouse. At least 3 mice were injected in each group. Mice were sacrificed 5 days after injection and ovaries were fixed for histological analysis or lysed for western blot.
In breeding assays, Ppp6cF/F and Ppp6cF/F;GCre+ female mice with sexual maturity were continually mated to Ppp6cF/F male mice with known fertility for 6 months. Cages were checked daily for counting the number of litters and pups.
All experiments were repeated at least three times. Student’s t test was used for statistical analysis and performed using SPSS. Data were expressed as mean ± SEM and values are statistically significant at *P<0.05; **P<0.01.
|
10.1371/journal.pcbi.1003501 | An Integrated Model of Multiple-Condition ChIP-Seq Data Reveals Predeterminants of Cdx2 Binding | Regulatory proteins can bind to different sets of genomic targets in various cell types or conditions. To reliably characterize such condition-specific regulatory binding we introduce MultiGPS, an integrated machine learning approach for the analysis of multiple related ChIP-seq experiments. MultiGPS is based on a generalized Expectation Maximization framework that shares information across multiple experiments for binding event discovery. We demonstrate that our framework enables the simultaneous modeling of sparse condition-specific binding changes, sequence dependence, and replicate-specific noise sources. MultiGPS encourages consistency in reported binding event locations across multiple-condition ChIP-seq datasets and provides accurate estimation of ChIP enrichment levels at each event. MultiGPS's multi-experiment modeling approach thus provides a reliable platform for detecting differential binding enrichment across experimental conditions. We demonstrate the advantages of MultiGPS with an analysis of Cdx2 binding in three distinct developmental contexts. By accurately characterizing condition-specific Cdx2 binding, MultiGPS enables novel insight into the mechanistic basis of Cdx2 site selectivity. Specifically, the condition-specific Cdx2 sites characterized by MultiGPS are highly associated with pre-existing genomic context, suggesting that such sites are pre-determined by cell-specific regulatory architecture. However, MultiGPS-defined condition-independent sites are not predicted by pre-existing regulatory signals, suggesting that Cdx2 can bind to a subset of locations regardless of genomic environment. A summary of this paper appears in the proceedings of the RECOMB 2014 conference, April 2–5.
| Many proteins that regulate the activity of other genes do so by attaching to the genome at specific binding sites. The locations that a given regulatory protein will bind, and the strength or frequency of such binding at an individual location, can vary depending on the cell type. We can profile the locations that a protein binds in a particular cell type using an experimental method called ChIP-seq, followed by computational interpretation of the data. However, since the experimental data are typically noisy, it is often difficult to compare the computational analyses of ChIP-seq data across multiple experiments in order to understand any differences in binding that may occur in different cell types. In this paper, we present a new computational method named MultiGPS for simultaneously analyzing multiple related ChIP-seq experiments in an integrated manner. By analyzing all the data together in an appropriate way, we can gain a more accurate picture of where the profiled protein is binding to the genome, and we can more easily and reliably detect differences in protein binding across cell types. We demonstrate the MultiGPS software using a new analysis of the regulatory protein Cdx2 in three different developmental cell types.
| Profiling the activity of regulatory proteins in multiple cell types is important for understanding cellular function, as a single regulator can bind to distinct sets of genomic targets depending on the cellular context in which it is expressed. Characterizing the determinants of such binding specificity is key to understanding how a single regulator can play multiple roles during development and other dynamic cellular processes. For example, pre-existing genomic context such as chromatin accessibility or the binding of other regulators may determine the binding of some developmental transcription factors (TFs) [1]–[3], while other ‘pioneer’ TFs may find their binding targets independently of the established chromatin state [4], [5].
Here we introduce MultiGPS, an integrated machine learning approach for the analysis of condition-specific binding events from multi-condition ChIP-seq data. MultiGPS performs binding event analysis across multiple conditions, sharing information across conditions to produce accurate joint binding estimates while simultaneously allowing for condition-specific binding events. MultiGPS employs a flexible framework for incorporating prior information into binding event discovery, allowing models of joint binding and sequence dependence to be used. The novel multi-experiment modeling approach of MultiGPS identifies the read enrichment associated with binding events that are bound in specific conditions, enabling principled methods of discovering differential binding [6]–[9].
Most current strategies for defining consistent ChIP-seq binding event locations across multiple experiments either analyze each experiment independently or pool reads for analysis. For example, the ENCODE2 project used standard ChIP-seq event finders on each experiment independently, and then merged event locations across experiments using a fixed-sized window to define event identity [10], [11]. Related methods specifically developed for multi-condition ChIP-seq analysis require that binding events be called in each condition individually as a preprocessing step, then apply statistical models to matched regions to detect differential effects [9], [12]. Other multi-condition approaches focus on ChIP-seq signals arising from broad regions of enrichment, such as histone modifications. These methods instead search for larger genomic regions where coverage patterns differ across experiments [8], [13]–[15]. In contrast, MultiGPS uses a joint multi-experiment model that considers the read data from all experiments to produce accurate location estimates of punctate binding events.
Approaches that first identify binding events and then attempt to merge locations across conditions may inappropriately combine distinct binding events that happen to be located within the same window. In genomic regions with a high density of binding events, the problem of matching sites across conditions is difficult and may lead to erroneous comparisons between binding strengths. Furthermore, the experiment-by-experiment event calling approach fails to use the full power of the experimental data when a large fraction of binding events are shared across conditions. An alternative method is to pool ChIP-seq reads from all experiments and then use a single event finding run to yield a consistent set of binding event locations that can be subsequently quantified in each individual experiment. However, this pooling approach may not discover weak condition-specific binding locations that are swamped by noise from other experiments in the pooled set of reads. Additionally, applying a single detection threshold in the pooled read set may bias the binding event calls to experiments that had higher sequencing coverage, better antibody batches, or fewer technical sources of error. Similarly, varying experimental parameters such as the fragmentation distribution could render the pooled read dataset harder to analyze by algorithms that assume a single, consistent set of experimental properties.
MultiGPS combines the theoretical benefits of pooling and separate ChIP-seq experimental analysis by using a Bayesian prior to couple the analysis of independent experiments together. This multi-experiment model is one aspect of a novel modeling approach that enables external sources of information to be included as priors in binding event identification (see Methods). In this work, we use the following priors, while recognizing that other directions are also possible:
MultiGPS detects binding events independently in each experiment in each step of its iterative optimization, allowing it to model experiment-specific parameters such as the distribution of reads around binding events and the properties of background noise. The iterative optimization procedure analyzes each experimental condition in turn, using binding event locations from other experiments to form an inter-experiment prior term for a single experiment optimization. MultiGPS therefore encourages the base locations of binding events to align across experiments when appropriate, and automatically produces coherent sets of binding events that are linked across experiments without any potentially noisy windowed analysis. To our knowledge, MultiGPS is the first ChIP-seq analysis approach that uses read data from multiple experiments in a joint and fully integrated method for identifying consistent and accurate binding event locations.
As a case study of our framework's sensitive and accurate multi-condition analysis, we applied MultiGPS to Cdx2 binding data in three developmentally relevant cellular contexts and found that condition-specific Cdx2 binding events are predicted by preexisting chromatin state. Surprisingly, condition-independent Cdx2 binding events that are bound in multiple contexts do not appear to be predetermined by accessibility or other chromatin signatures, and instead may be predicted on the basis of cognate motif occurrence. Our results suggest that Cdx2 can act as a pioneer factor at a subset of sites, while also being influenced by preexisting genomic context at other sites. Therefore, our results have consequences for understanding where TFs will bind when introduced into an established regulatory state during development, or when induced artificially during cellular programming techniques.
We find that MultiGPS's inter-experiment and motif priors encourage binding location consistency on CTCF biological replicate experiments. The binding events that are called in both CTCF replicates should by definition be located at the same base location. As we can see in Figure 1a, when MultiGPS is run without either prior, predicted binding events do not typically align to each other or to cognate motif instances. Each prior alone makes a significant, though incomplete, improvement in binding event accuracy (Figure 1b–c). The inter-experiment prior is able to significantly improve the distance to the nearest motif when compared to sites identified without any positional priors (p<5×10−5, Mann-Whitney U test comparing binned distance to nearest motif match). The motif prior significantly improves the distance to the nearest site in another experiment (p<1×10−12, Mann-Whitney U test comparing binned distance to the nearest event in another experiment). In these two comparisons, we used information sources not considered by the prior as validation (motif distance for the inter-experiment prior and inter-experiment distance for the motif prior). The use of both priors together fully utilizes available sequence and multi-experiment information and allows almost all binding events in this example to be aligned to consistent (typically motif-associated) locations (Figure 1d). These comparisons are not meant as absolute performance assessments for the MultiGPS modeling approach, but instead as relative measurements of the benefit of using additional types of prior information within a single modeling framework.
MultiGPS facilitates the detection of differential binding events by accurately quantifying read count levels associated with each binding event in each analyzed experiment. Since at present no ChIP-seq datasets exist for which absolute binding levels are known across multiple conditions, we generated simulated ChIP-seq datasets to test the relative performance of MultiGPS in defining differential binding events. In our simulated data, the distribution of reads at binding events mirrors the properties of real ChIP-seq datasets (see Methods). A subset of binding events is chosen to be differentially enriched across conditions, and while we chose to set the absolute level of differential enrichment to be constant at all differential events (4-fold in Figure 2, 8-fold in Figure S1), simulated sampling noise leads to a wide array of apparent fold differences (Figure 2a, blue dots).
Using the simulated data, we compared MultiGPS with other approaches for determining differential binding events. We used MultiGPS (without the motif prior since no sequence information was used to simulate the data), MultiGPS in single-condition mode (i.e. without using either inter-experiment or motif priors), and the single-condition event finders MACS [18] and SISSRs [19] to predict binding events in each simulated condition. All methods made comparable numbers of binding event predictions in each dataset (Figure S2). For the methods other than MultiGPS, differential binding events were defined using: a) binding event list comparison, where differential binding events are those that are detected in one condition and no binding event is detected within 200 bp in the other condition; b) using the software DBChIP [9]; or c) by counting reads that occur within the enriched regions and inputting the resulting tables into edgeR [6] (using the same parameters as used by edgeR within MultiGPS).
The results illustrate the problems with defining differentially bound events using binding event list comparison. Regardless of which event finding method was used to provide input binding events, list comparisons have poor sensitivity when predicting differentially bound events with higher mean read counts (Figure 2b, dashed lines). Such events are more likely to be detected in both conditions and hence would be treated as non-differential binding events regardless of quantitative differences in ChIP enrichment levels. Conversely, binding event list comparisons have low specificity when predicting differentially bound events with lower mean read counts (Figure 2c, dashed lines). Low enrichment binding events may have read counts that are just above a binding event detection threshold in one condition, and just below in another, even if there is no significant quantitative difference in the underlying ChIP enrichment levels. Such events would appear as false positive differential binding event predictions according to the binding event list comparison approach.
In contrast, approaches that test differential binding using statistical analyses of read count tables have uniformly high specificity across our test datasets (Figure 2c, solid lines). These methods also have higher sensitivity when predicting differential binding events with higher mean read counts (Figure 2b, solid lines) or involving greater absolute differences in binding levels (Figure S1b, solid lines). EdgeR attains the highest overall sensitivity using the read count tables generated by MultiGPS, thus illustrating the advantages of MultiGPS' probabilistic approach to quantifying read enrichment at binding sites across conditions.
MultiGPS models experiment-specific parameters such as the distribution of reads around binding events and the properties of background noise. To investigate whether these parameters yield improved quantification of binding event ChIP enrichment, we ran the complete MultiGPS model on 14 ChIP-seq experiment sets in which the ENCODE2 project has performed replicated ChIP-seq of a given protein in all three human Tier 1 cell lines. While no gold standard exists for measuring the accuracy of ChIP-enrichment quantification, we reasoned that accurate quantification estimates should be correlated across biological replicate experiments. For each of the 14 experiment sets, MultiGPS yields per-replicate estimates of binding enrichment for binding events discovered in any cell line. We compared these values to those produced by the widely used approaches of counting read occurrences in a window around the binding event locations (here we use a 400 bp window centered on the MultiGPS-defined binding event locations), or by using the peak heights defined by MACS [18] analyses of the same data. Quantified read counts were compared across biological replicate pairs using Spearman's rank correlation, a nonparametric assessment of statistical dependence that makes no distributional assumptions that could artificially favor one model over another. Note that MACS does not produce per-replicate read counts or peak heights at each event, and so to compare MultiGPS with MACS we ran MACS on each replicate separately and compared read counts and heights at only those binding events detected in both replicates by MACS and MultiGPS. Read counts at these reproducibly detected binding events may be more highly correlated than read counts associated with the wider sets of binding events tested in the comparison between MultiGPS estimates and windowed read counts.
As shown in Figure 3, MultiGPS improves the cross-replicate correlation of binding event quantification estimates in most tested datasets, implying that MultiGPS has reduced the effects of inter-replicate noise in comparison to the window counting approaches. We expect that reducing the degree of over-dispersion between replicates will yield greater sensitivity in detecting significant differences between conditions. Indeed, in all 14 tested datasets we find substantially greater numbers of statistically significant differentially enriched binding events between cell lines when we run edgeR [6] on the MultiGPS quantification table as opposed to the table of read counts produced by the window approach (Table S1). Therefore, MultiGPS improves the quantification of binding event ChIP-enrichment and the detection of condition-specific binding events.
To demonstrate the ability of MultiGPS to analyze biologically relevant condition-specific binding events, we examined if MultiGPS improves upon the independent analysis of experiments when identifying Cdx2 events in multiple conditions. Cdx2 is a mammalian caudal-type homeobox protein that plays a key role in regulating the development of diverse tissue types. For example, Cdx2 is a master regulator of the intestinal lineage when expressed in endoderm [20], and also plays a key role in defining caudal motor neuron fate when expressed in motor neuron progenitors (pMNs) [21]. In addition, over-expression of Cdx2 in embryonic stem (ES) cells forces cells to differentiate into the trophectoderm lineage [22], [23]. We thus wanted to elucidate how Cdx2 performs its different regulatory functions in these three developmental contexts. Does it bind to the same genomic targets in all cell types, or does it bind distinct targets in each context? If the latter, how is such specificity achieved? To determine the context-dependent binding activity of Cdx2, we performed ChIP-seq analysis of Cdx2 after it was over-expressed in ES cells, endoderm, and pMNs. We call these cell types after Doxycycline-dependent Cdx2 induction ES+Dox Cdx2, endoderm+Dox Cdx2, and pMN+Dox Cdx2, respectively. Since Cdx2 is not natively expressed in any of these three cell types, our experiments provide a useful model of how a transcription factor responds to a new cellular environment.
We found that MultiGPS outperformed an independent binding event analysis (i.e. using independent runs of MultiGPS without the use of priors) on the three Cdx2 conditions using a binding event list comparison approach to determine differentially bound sites. While this is a common approach in the literature, it leads to highly misleading results. As can be seen in Figure 4, the binding event list comparison suggests that 95% of pMN+Dox sites are not bound in ES+Dox cells. However, the apparent degree of differential binding is largely caused by the disparity in the total numbers of binding events predicted in each condition (3,704 in ES+Dox and 36,651 in pMN+Dox). The difference in the total number of events is in turn caused by differences in read coverage between the conditions and the thresholds employed to determine bound events. In addition, the binding event list comparison approach may miss differences at events when the level of ChIP enrichment varies significantly between conditions. To perform a more principled analysis of Cdx2 differential binding, we analyzed the ChIP-seq data collection using MultiGPS (Table 1). With the coupled MultiGPS method only 24% of all pMN+Dox Cdx2 binding events are significantly differentially enriched in pMN+Dox cells compared with ES+Dox cells (p<10−3), while 37% of all ES+Dox Cdx2 binding events are significantly differentially enriched in ES+Dox cells compared with pMN+Dox (p<10−3).
Since MultiGPS identifies a large proportion of condition-specific Cdx2 binding events without finding any evidence for a corresponding change in Cdx2's DNA-binding preference, we asked whether ES cell genomic context could predict the observed condition-specific binding of Cdx2 after induction. To answer this question, we examined the ES genomic patterns at the locations of Cdx2 sites that are significantly enriched in ES+Dox cells according to MultiGPS. Interestingly, we found that ES+Dox-specific Cdx2 sites are enriched for ES signatures of chromatin accessibility (DNaseI hypersensitivity), enhancers (H3K4me1 and H3K27ac ChIP-seq), and TF binding (Oct4, Sox2, and Nanog ChIP-seq), but not active transcription (H3K4me3 ChIP-seq) (Figure 5). Conversely, pMN+Dox-specific Cdx2 sites and endoderm+Dox-specific Cdx2 sites show no enrichment for these ES cell chromatin signatures (Figure 5 & Figure S3).
To more rigorously test the capacity of ES cell genomic context to predict ES+Dox-specific Cdx2 binding events, we trained support vector machines (SVMs) to classify Cdx2 binding events vs. unbound Cdx2 motif instances using the read count information from a collection of 55 ES experiments (2 DNaseI-seq, 13 histone modification ChIP-seq, 35 TF, co-activator and chromatin modifier ChIP-seq, and 5 Pol2 ChIP-seq experiments). Cross-validation was used to generate disjoint training and test sets (see Methods). Our SVMs discriminate ES+Dox-specific Cdx2 sites from unbound sites with an area under true-positive vs. false-positive curve (AUC) of 0.95–0.96, suggesting that the pre-existing genomic context in ES cells is highly predictive of future Cdx2 binding. Conversely, our SVMs are unable to discriminate pMN+Dox-specific Cdx2 sites from unbound Cdx2 motif instances using ES genomic context (AUC = 0.63, Figure 6). Our results therefore suggest that condition-specific Cdx2 binding events are more likely to be located in genomic regions that already displayed regulatory activity or accessibility before Cdx2 expression was induced.
Since condition-specific Cdx2 binding events appear highly correlated with immediately pre-existing genomic context, we reasoned that the condition-independent Cdx2 sites that are bound in multiple conditions might also display the same associations. For example, Cdx2 sites that are bound in two conditions may represent locations that happened to have pre-existing regulatory activity or accessibility in both conditions. Surprisingly, the Cdx2 sites bound in both ES+Dox and pMN+Dox conditions are not enriched for accessibility, enhancer chromatin marks, or TF binding in ES cells (Figure 5). Furthermore, SVMs trained as before are unable to discriminate between these shared Cdx2 sites and unbound motif instances using ES genomic context information (AUC = 0.61, Figure 6). These results suggest that the condition-independent Cdx2 sites are not determined by pre-existing genomic context, in contrast with the condition-specific sites.
Given that the condition-independent Cdx2 sites do not seem to have any distinguishing chromatin features before Cdx2 induction, we asked how Cdx2 recognizes these sites regardless of genomic context. We hypothesized that such sites may have sequence features that enable condition-independent binding. To test this hypothesis, we trained SVMs to discriminate condition-independent Cdx2 sites from condition-specific Cdx2 sites using only 4-mer word frequencies in 200 bp windows around the sites. Surprisingly, even these crude sequence features were sufficient to discriminate between the two types of sites (AUC = 0.89–0.92, Figure 7a), suggesting that some sites contain sequence information that enables condition-independent Cdx2 binding. We next used the discriminative motif finders DEME and DECOD [24], [25] to determine which sequence motifs discriminate between Cdx2 site types. Interestingly, both tools returned the primary Cdx2 motif as being the most discriminative, even though most condition-specific and condition-independent sites contain instances of the same primary motif. This apparent contradiction is resolved by considering features of the motif instances in each set of Cdx2 sites. SVMs trained with just three simple primary Cdx2 motif-related metrics – the maximum motif score in the 200 bp window around sites, the number of motif instances above a threshold, and a score that integrates motif scores across the entire 200 bp window [26] – were able to discriminate between condition-independent and condition-specific sites with reasonable accuracy (AUC = 0.81, Figure 7b). In other words, the strength and multiplicity of motif instances are somewhat predictive of condition-independent Cdx2 binding.
Taken together, our results suggest that sequence information allows Cdx2 to act as a pioneer TF at some sites, overriding the lack of pre-existing accessibility or chromatin markers.
MultiGPS provides a principled platform for the analysis of differential protein-DNA binding across multiple experimental conditions by preferring consistent binding locations across related experiments while also modeling condition-specific experimental parameters. Rather than treating reads from all experiments as equivalent, MultiGPS models experiment-specific read distributions around binding events. MultiGPS can thus correctly analyze collections of related ChIP experiments that were performed according to different protocols such as mixtures of related ChIP-seq and ChIP-exo [27] experiments. As demonstrated above, MultiGPS improves the quantification of ChIP enrichment at binding events in comparison with the typically used window-counting approaches, thus enabling more sensitive analyses of differential binding enrichment between conditions.
Since MultiGPS prefers but does not force binding events to align across experiments, it may also be used to study possible forms of differential binding activity that we did not illustrate. For instance, it may be of interest to examine locations where the underlying read evidence overrides the MultiGPS inter-experiment prior, resulting in differing reported binding locations across experiments. Such locations may represent shifts in binding location between conditions, which may be useful for studies of nucleosome positioning or regulators that might bind alternate nearby locations in different conditions.
We demonstrated that MultiGPS can characterize condition-specific binding and then used MultiGPS to characterize the nature of both condition-specific and condition-independent binding of Cdx2. Our results suggest that many condition-specific Cdx2 binding events are located in regions that had pre-existing regulatory activity, thus agreeing with hypotheses proposed to explain the observed binding of other developmental TFs [1]–[3]. However, Cdx2 also appears to act as a ‘pioneer’ at a subset of sites that are bound condition-independently. Our analysis suggests that such sites on average contain stronger and more frequent Cdx2 motif instances than condition-specific sites, thus suggesting a possible mechanism by which condition-independent sites can be bound regardless of preexisting genomic context. These findings also accord with our recent demonstration that TF combinations can override pre-existing cellular state to synergistically bind composite motifs during motor neuron programming [28], perhaps pointing to a deeper relationship between sequence information and ‘pioneer’ binding activity.
In our previously described GPS [16] and GEM [17] approaches to binding event detection, ChIP sequencing data are modeled as being generated by a mixture of binding events along the genome, and an Expectation Maximization (EM) learning scheme is used to probabilistically assign sequencing reads to binding event locations. The assignment of reads is achieved via an empirically estimated multinomial distribution, Pr(rn|x), which gives the probability of observing read rn from a binding event located at genomic coordinate x. Conceptually, every base position is treated as a potential binding event, although the use of a sparse prior [29] has the effect of allowing only a small subset of these potential binding events to take responsibility for observed reads and survive the EM training process.
In MultiGPS, we decouple the relationship between a binding event's index and its spatial (genomic) location. Specifically, we introduce a vector of component locations μ where μj is the genomic location of event j. We initialize a large number of potential events, M, such that the events are evenly spaced in 30 bp intervals along the genome. Note, however, that the use of a sparse prior will again result in only a subset of events remaining active in the model after training (i.e. components having mixing probability πj>0; see MAP estimation of π below). In the new mixture model, the likelihood of observing the N total ChIP read locations r is given by:where Pr(rn|μj) is the distribution over ChIP-seq read positions conditioned on membership in a binding event at location μj. This distribution is initialized to a strand-specific shape typical of many ChIP-seq datasets (see Figure S5), and is iteratively re-estimated during EM training using the distribution of reads observed around high-confidence binding site locations. The above expression calculates the observed data likelihood of a mixture model by taking the product over all reads, where each read averages over each possible binding event that may have caused it. This extension of the model allows us to apply prior knowledge directly to the positions of the binding events (μ), without affecting the binding event strength estimation or the sparsity-promoting prior, which continues to act on raw expected read counts.
We introduce a Bernoulli prior over each genomic location where each element ki of the parameter k corresponds to the probability that location i is a binding event (that is, i μ). This prior assumes that there can be only one or zero binding events at a single position and that binding positions are selected independently along the genome according to this weighting. The prior assigns likelihoods to a set of binding events on a genome of size L as follows:
As in the original framework, the latent assignments of reads to binding events are represented by the vector z. The complete-data log posterior can thus be derived as follows:Here, C is a normalization constant that does not involve any of the terms to be optimized. It can be seen that the overall binding event sparsity-inducing negative Dirichlet prior α acts only on the mixing probabilities π, which controls the total fraction of reads assigned to each binding event, and the positional prior k acts only on the binding event locations μ. Therefore, the E-step that calculates the relative responsibility of each binding event in generating each read is unchanged from our original framework, following standard mixture model approaches:Furthermore, the maximum a posteriori probability (MAP) estimation of π is also unchanged:where Nj is the effective number of reads assigned to binding event j. The α parameter can thus be interpreted as the minimum number of ChIP-seq reads required to support a binding event remaining active in the mixture model. We set the value of α per experiment to be the maximum number of reads that would be expected to occur (p>10−7) in a window equal to the effective range of the binding distribution should the experiment's reads be distributed uniformly along the mappable portion of the genome.
We can estimate μ component-wise since it only participates in sums in the log likelihood. However, no closed form solution exists since the prior k has no parametric form. We can determine the MAP (integer) value of μj by simply enumerating over all possible values of μj. Specifically, the MAP value of μj is: . If the maximization step results in two components sharing the same location, they are combined in the next iteration of the algorithm.
One practical use for the positional prior k is to bias the estimated binding locations towards biologically appropriate base positions. For example, a TF's position weight matrix scores along the genome can be directly encapsulated in k in the above framework. As described previously for our GEM approach [17], we can estimate binding motifs from current estimates of binding locations, and reciprocally use those motifs as prior information to re-estimate binding event locations. Note that motif priors are incorporated quite differently in GEM and MultiGPS. In practice, MultiGPS uses MEME [30] to discover a set of over-represented motifs in the top 500 most enriched binding events (80 bp windows), chooses the motif with the highest true-positive vs. false-positive AUC for discriminating bound regions from random sequences (if any motif AUC≥0.7), and incorporates the genomic log-odds scores for that motif in the positional prior.
Unlike our previously described approaches, MultiGPS incorporates an additional mixture component that explicitly models noise (i.e. reads arising from nonspecific binding and independent of any binding event). Whereas binding component read distributions have approximately finite support (and therefore only allow binding events to take responsibility for reads in their local vicinity), the noise component is defined as having a global distribution. The form of the noise distribution can be defined as uniform or can be parameterized using the read density observed in a control experiment. In the latter case, the shape of the noise distribution is defined by smoothing the control experiment's read counts using a 50 bp sliding window (adding fractional pseudocounts to 50 bp windows that contain no control reads).
For a more efficient and stable training process, some parameters in MultiGPS are re-estimated only periodically, including the form of the binding event read distribution, the noise component mixing probability (πM+1), and the binding motif position weight matrix. We can therefore think of MultiGPS as an instance of a generalized EM algorithm. Generalized EM algorithms increase the expected log likelihood in each M step without necessarily achieving a maximum in each iteration (as in the original EM algorithm) [31]. Convergence to a local optimum is guaranteed with generalized EM algorithms, as it is with the EM algorithm [31].
As with GPS and GEM, MultiGPS filters predicted binding events to require that their associated read counts are significantly enriched (p<10−3, Benjamini-Hochberg corrected Binomial test) over the corresponding read count from an appropriately normalized control experiment, such as a mock-IP experiment. The control experiment normalization factors are estimated via regression on the read count ratios in 10 Kbp windows. Control read counts are associated with individual binding events via maximum likelihood assignments using the trained model (i.e. assigning control reads to binding events without changing the π and μ parameters learned from the ChIP data).
MultiGPS can be run in a multi-condition analysis mode by providing multiple input datasets and structured annotation as to how these datasets are related (i.e. which datasets represent technical or biological replicates of others, which collections of datasets represent distinct experimental conditions, and which datasets serve as controls for others). MultiGPS then runs semi-independent mixture model training across all provided data. Since reads from distinct conditions are not pooled, MultiGPS can maintain condition-specific and replicate-specific parameters, including distinct binding event read distributions per replicate, distinct noise component read distributions and mixing probabilities per replicate, and distinct binding motifs per condition. However, the goal is to report binding event locations that are consistent across conditions. This is achieved using another form of prior information during the maximization of binding event locations μ.
We motivate our approach by imagining a TF that binds to N locations in cellular condition A and N locations in cellular condition B. In typical analysis scenarios, the number of bound locations will be much fewer than the number of bases on the genome (i.e. ), and a non-zero set of S locations will be bound in both A and B conditions. We present the model for two conditions with a symmetric number of binding sites here for notational simplicity, but note that the same process can be applied to any number of conditions with more complex binding site sharing patterns. A schematic example (not to scale) of bound and unbound bases in two conditions as a fraction of the genome is shown in Figure 8.
Now, the distribution that generates binding positions is extended from the single-condition case of a Bernoulli distribution to a multivariate Bernoulli distribution. As suggested by the schematic in Figure 8, this distribution generates a sample from {(0,0),(0,1),(1,0),(1,1)} at each base in the genome, where each element in a sample corresponds to whether a binding site is present at that position in that condition. This generative model induces the following distribution over genome positions i with respect to binding site positions in conditions A and B:We parameterize the above distribution during each iteration of the MultiGPS algorithm by choosing appropriate values for N and S (L is fixed, being the length of the genome). While N can be taken from MultiGPS' current estimate of the number of binding events in each condition, we do not typically know S. We therefore define S by setting the ratio S/N as described below.
We need to know the contribution of the location prior in the optimization step for the binding site locations μ. For the multi-condition analysis, we jointly optimize two binding sites when they fall within 100 bp of each other (range chosen empirically as the maximum range for which the inter-experiment prior will have an effect at most binding events, see Figure S7). The model optimization step determines whether the two binding positions in question are separate (and therefore two site-specific positions contribute to ) or shared (and therefore one shared site contributes to ). All other bases will be the same during this optimization since all other binding sites are fixed, and can be ignored in this step. Using the distribution above gives the following contribution to the prior :where in most experimental studies. If the binding events share a location across conditions, we choose the optimal shared position w by maximizing the expected complete-data log posterior (with terms not affecting the minimization omitted) as follows:Alternatively, the two binding component locations are independent, in which case the two positions are optimized independently:
The decision to use the coupled or uncoupled estimate is based on which scenario yields the higher expected complete-data log posterior probability. Higher values of the ratio S/N encourage the coupling of nearby binding event locations across conditions by increasing with respect to (see Figures S6 & S7). In MultiGPS, we set the ratio S/N to be equal to 0.9, although in practice we observe few differences in the proportion of aligned binding components when varying the ratio in the range 0.5<S/N<0.99. This is because a number of nearby genomic locations give similar probabilities when maximizing μj (Figure S7), allowing the penalties associated with moving the components away from the optimal positions in each condition to be overridden by the positive-valued prior over a range of S/N ratios. Note, however, that MultiGPS will still prefer the uncoupled scenario in situations where the read evidence supports distinct binding locations across conditions. This behavior represents a data-driven joint analysis mode that weighs the statistical confidence given by the reads against prior knowledge of the experimental setup in a probabilistically optimal way. We also note that positional prior terms encapsulating per-condition TF position weight matrix scores can be accounted for in the μ maximization terms above in a manner analogous to that described in the previous section. MultiGPS can therefore account for both motif positional priors and the inter-experiment prior.
Assessing all possible scenarios of coupled and uncoupled binding events during the update of each μj becomes prohibitive when analyzing more than two conditions. Therefore MultiGPS assesses a limited number of scenarios when updating μj in such cases: 1) event j is uncoupled across all conditions; 2) event j is coupled with a corresponding event in one other condition; or 3) event j is coupled with corresponding events in all other conditions. The scenario that yields the best overall likelihood is chosen.
A table containing the replicate-specific read counts associated with each binding event is generated from the MAP-estimated responsibilities γ. MultiGPS uses the edgeR Bioconductor package [6] to detect differential ChIP enrichment between conditions from the read count table. We use edgeR's TMM method to calculate normalization factors, and the glmLRT method to calculate likelihood ratios. In the Cdx2 example described here, we used a fixed overdispersion parameter of 0.15 across all experiments, which results in a stricter definition of significant differential enrichment than the overdispersion parameters estimated by edgeR.
To computationally simulate multi-condition ChIP-seq data, we defined a hypothetical system in which a protein has 20,000 binding events in the mouse genome (version mm9). The relative strengths of each of these binding events was drawn randomly from a distribution of relative read counts observed for Cdx2 binding events in our pMN ChIP-seq experiments. For two hypothetical experimental conditions, A & B, we randomly chose 20% of the binding events to be differentially enriched in condition A with respect to B, and we modify the relative binding event strengths of these sites such that they are 4-fold (or 8-fold in separate simulations) greater in condition A versus B. We similarly chose a non-overlapping 20% of binding events to be differentially enriched in condition B with respect to A. The binding events were placed along the genome in 10 Kbp intervals.
We then generated 20 million read positions for each of two replicates in each of the two conditions. To reflect the typical signal-to-noise ratio observed in real ChIP-seq experiments, 95% of the read positions are spread randomly across the entire genome. The remaining reads (averaging 1 million per replicate) are distributed amongst the binding events according to the relative strength of the event in each relevant condition, and accounting for read sampling noise using a negative binomial distribution with an over-dispersion parameter of 0.1. The MA plot in Figure 2a shows the log2 mean read count and log2 fold difference for each binding event in the simulated experiments. The position of generated reads with respect to the defined binding event location is drawn from a bimodal distribution typical of ChIP-seq binding sites (Figure S5).
We ran the following binding event analysis methods on the simulated data: a) MultiGPS on the entire dataset, using default parameters with the exception of turning off the use of sequence information and the motif prior (since motif information was not used in generating the simulated data); b) MultiGPS without the inter-experiment prior or the motif prior on the entire dataset, which has the effect of calling binding events in each condition independently; c) MACS [18] using default parameters on each condition independently, merging reads across replicates; and d) SISSRs [19] using default parameters (with the exception of using a p-value cutoff of 0.05) on each condition independently, merging reads across replicates. For binding event list comparison approaches, per-condition events were compared with each other using a 200 bp window. In other words, if an event prediction in one condition is located within 200 bp of an event prediction in the other condition, it is treated as being in the intersection of the binding event list comparison, and thus not differentially bound. EdgeR [6] was run either internally in MultiGPS (as described above) or, using the same parameters, on read count tables built by counting reads that overlap the peak regions found by MACS or SISSRs. We also ran DBChIP [9] using default parameters with the exception of an FDR threshold <0.01 and using the MACS peaks as inputs. Sensitivity and specificity in Figures 2 and S1 are defined by comparing predicted binding events to the positions of the simulated differential binding events using a 100 bp window.
Support vector machines were trained using the libSVM [32] interface in Bioconductor (e1071). In all cases, classification accuracy was determined using a randomly selected held-out test set of 100 datapoints, and training of each SVM application was repeated 20 times (using different held-out test sets each time) to calculate average true-positive vs. false-positive AUC values.
To train SVMs using ES chromatin state data, we first gathered 55 mouse ES ChIP-seq and DNaseI-seq experiments from a variety of sources [33]–[42]. We defined positive training sets from the top-most Cdx2 binding events for each condition-specific and condition-independent permutation (up to a maximum of 4,000 binding events), and we also defined a negative training set of 10,000 matches to the Cdx2 cognate binding motif (as defined by UniProbe [43]) that were not bound by Cdx2 in any experiment. Reads were counted in 1,000 bp windows around each of the positive and negative locations for each of the 55 mouse ES experiments, and SVMs were trained on the resulting 55-dimensional vectors without any normalization.
SVMs were trained on k-mer frequencies by enumerating the occurrences of each 4-mer (accounting for reverse-complement redundancies) in 200 bp windows around each of the top-most Cdx2 binding events for each condition-specific and condition-independent permutation (up to a maximum of 4,000 binding events). Similarly, SVMs were also trained on three pieces of information from the same 200 bp windows: the maximum log-likelihood ratio score for the Cdx2 motif in the window; the number of matches to the motif in the window that score more than a 5% FDR threshold; and the probability of binding occupancy in the window [26].
An ES cell line harboring Dox-inducible Flag-tagged Cdx2 was generated as previously described [44]. Anti-Flag ChIP-seq experiments were performed as previously described [44] after 24 hours of Dox-induced expression of Cdx2 in the ES cells or in motor neuron progenitors or endoderm cells that were differentiated from the same ES cell line. Differentiation of the ES cells to pMN and endoderm lineages was also described previously [20], [21]. Mock IP control experiments were performed using the same system. Sequenced ChIP-seq reads were mapped to the mm9 reference genome using Bowtie [45]. ChIP-seq data generated during this study were deposited in GEO under accession numbers GSE39433 and GSE39435.
MultiGPS is available as an open-source Java package, released under the MIT license, from: http://mahonylab.org/software and https://github.com/shaunmahony/seqcode. Simulated multiple condition ChIP-seq datasets are also available from the same webpage.
A summary of this paper appears in the proceedings of the RECOMB 2014 conference, April 2–5 [46].
|
10.1371/journal.ppat.1008031 | A nanobody targeting the translocated intimin receptor inhibits the attachment of enterohemorrhagic E. coli to human colonic mucosa | Enterohemorrhagic E. coli (EHEC) is a human intestinal pathogen that causes hemorrhagic colitis and hemolytic uremic syndrome. No vaccines or specific therapies are currently available to prevent or treat these infections. EHEC tightly attaches to the intestinal epithelium by injecting the intimin receptor Tir into the host cell via a type III secretion system (T3SS). In this project, we identified a camelid single domain antibody (nanobody), named TD4, that recognizes a conserved Tir epitope overlapping the binding site of its natural ligand intimin with high affinity and stability. We show that TD4 inhibits attachment of EHEC to cultured human HeLa cells by preventing Tir clustering by intimin, activation of downstream actin polymerization and pedestal formation. Furthermore, we demonstrate that TD4 significantly reduces EHEC adherence to human colonic mucosa in in vitro organ cultures. Altogether, these results suggest that nanobody-based therapies hold potential in the development of much needed treatment and prevention strategies against EHEC infection.
| Currently, there is no effective treatment or vaccine against enterohemorrhagic E. coli (EHEC), a bacterial pathogen that infects human colon after the ingestion of contaminated food. It thrives in the colon thanks to its ability to attach intimately to the intestinal epithelium. Here, we have identified and characterised a small antibody fragment (nanobody) that recognises Tir, a receptor injected by the bacterium into the host cell to mediate intimate attachment. This nanobody shows higher affinity against Tir than its natural bacterial ligand (intimin) and, most importantly, blocks the intimate attachment of the pathogen to the human colonic tissue. Our results show the potential of this nanobody to prevent and treat EHEC infection.
| Enterohemorrhagic E. coli (EHEC) is a major public health concern in industrial countries with most severe infections linked to serotype O157:H7. In addition to diarrhoea, EHEC can cause hemorrhagic colitis as well as life-threatening hemolytic uremic syndrome (HUS) damaging the kidneys and central nervous system [1–4]. EHEC naturally resides in the intestinal tract of cattle, and most infections are acquired by consumption of undercooked beef products or cross-contaminated vegetables or sprouts [5]. Upon infection, EHEC adheres to the epithelium of the distal ileum and colon by forming attaching and effacing (A/E) lesions, which are characterized by intimate bacterial attachment and effacement of the brush border microvilli [6, 7]. This is mediated by the Locus of Enterocyte Effacement (LEE) [8], a pathogenicity island encoding a filamentous type III secretion system (T3SS) [9, 10], the outer membrane adhesin intimin and the translocated intimin receptor (Tir), and other effector proteins involved in pathogenesis [11, 12].
After formation of the translocation filament consisting of EspA proteins, Tir is injected into intestinal epithelial cells (IECs), where it integrates into the plasma membrane in a hairpin loop topology, presenting an extracellular domain of about 100 residues (TirM) [13, 14] that serves as a binding site for the C-terminal lectin-like domain of intimin [15–17]. Binding of intimin to Tir leads to intimate bacterial attachment, Tir clustering, activation of actin polymerization pathways and subsequent formation of actin pedestals and A/E lesions [7, 18–22].
Other key virulence factors of EHEC are the phage-encoded Shiga toxins (Stx) which are released into the bloodstream and cause the systemic effects associated with HUS [23, 24]. So far, there is no specific treatment for HUS, and application of antibiotics is discouraged as it induces Stx expression and thereby increases the risk of developing HUS [25, 26]. Therefore, there is a need to develop alternative therapies, and the use of antibodies (Abs) has been proposed for treatment of infectious diseases [27]. In particular, members of the family Camelidae (e.g. dromedaries, llamas) produce a class of Abs devoid of light chains [28, 29]. In these heavy-chain-only Abs, the antigen-binding site is formed by a single variable domain termed VHH [30]. The recombinant expression of camelid VHHs yields single domain Ab fragments, which are also referred to as nanobodies (Nbs) [31]. The VHHs have extended complementarity determining regions (CDRs) capable of adopting novel conformations and recognizing epitopes located in otherwise non-accessible clefts or protein cavities, such as active sites of enzymes [32, 33] and inner regions of surface proteins from pathogens [34]. They also show strict monomeric behavior, reversible folding properties, higher resistance to proteolysis and thermal degradation, when compared with the variable domains of conventional antibodies [31, 35, 36]. In addition, the high similarity between VHHs and human VH3 sequences opens their potential use in therapeutic applications [31]. These beneficial properties offer opportunities to use Nbs for the development of therapeutic inhibitors against extracellular pathogens [37, 38].
We have previously isolated a set of Nbs binding to EspA, the C-terminal receptor-binding domain of intimin (Int280) and the TirM domain from a library of VHHs obtained after immunization of a dromedary (Camelus dromedarius). Nanobodies were secreted to the extracellular medium using the hemolysin (Hly) transport system of E. coli and purified from the culture supernatants [39]. Here, we have investigated the ability of the selected Nbs to inhibit EHEC adhesion to HeLa cells and human colonic mucosa. We have identified a Nb clone that binds TirM, named TD4, which reduces the interaction of TirM with Int280 and interferes with actin pedestal formation and the intimate attachment of EHEC to human cells. Importantly, using infection of human in vitro organ cultures (IVOC), we demonstrate that Nb TD4 can also inhibit the attachment of EHEC to human colonic tissue.
To determine if purified Nbs against EspA, Int280 and TirM affected EHEC A/E lesion formation, HeLa cells were infected with EHEC for 3 h in the presence or absence of Nbs. Actin pedestal formation was visualized and quantified by immunofluorescence staining. While EHEC attachment and pedestal formation was not affected by Nb clones recognizing EspA (EC7) or Int280 (IB10) at concentrations 200 nM (Fig 1A), a Nb clone binding TirM (TD4) significantly reduced the accumulation of actin beneath the attached bacteria. As clustering of Tir is necessary for A/E lesion formation, we also evaluated the Tir localization in the presence of Nb TD4 or an unrelated Nb. As shown in Fig 1B, localization of Tir beneath adherent EHEC was evident in samples incubated with the control Nb or non-treated controls, while no Tir accumulation was observed in the presence of TD4. No Tir staining was detected in HeLa cells infected with EHECΔtir. We quantified the effect of TD4 by determining the mean of pedestals formed on infected cells under various concentrations of Nb TD4. This revealed a significant decrease in the number of actin pedestals per infected cell when TD4 Nb was added at concentrations ≥ 100 nM (Fig 1C).
We wanted to rule out the possibility that the lack of Tir clustering beneath the bound bacteria could be due to a block of Tir translocation through the T3SS and not to the direct interaction of the Nb TD4 to the exposed region of Tir upon its translocation and insertion in the plasma membrane of the host cell. We tested this possibility and simultaneously evaluated whether TD4 can interfere with EHEC actin-pedestal formation when added at different times during infection. To this end, we increased the infection rate by halving the volume of the medium and added 200 nM of Nbs TD4 or control (Vamy) simultaneously with the infection, or at 1 or 2 h post-infection. After a total of 3 h of infection, all samples were stained for Tir and the HA-tagged Nbs to test for their co-localization (Fig 2). Due to the higher infection rate, some Tir signal could be observed beneath the bound bacteria with TD4, but the arrangement of Tir staining in the host cells was altered, being distributed in the cytoplasm and showing only weak staining marks at the site of the EHEC adhesion (Fig 2A). Furthermore, we could detect colocalization of the HA-tagged TD4 with Tir, showing that the interaction between TD4 and Tir occurs at the surface of infected host cells, once Tir has been translocated, and suggesting that this interaction is responsible for the observed phenotype. In contrast, infections incubated with the control Nb (Vamy) showed strong Tir signals accumulated beneath EHEC bacteria and no staining of the cytoplasm (Fig 2A). Importantly, this inhibitory effect of TD4 was observed at similar levels independently of the time of addition of the Nb (0, 1 or 2 h post-infection) as determined by quantification of Tir clusters in the infected cells (Fig 2B).
Lastly, we assessed the inhibitory effect of TD4 at longer times of infection. HeLa cells were infected with EHEC for 6 h in the presence or absence of TD4, and stained for bacteria, Tir and F-actin. In this experiment fresh medium and Nbs were added after 3 h of infection. Inspection of these samples revealed the presence of a high number of intimately attached EHEC bacteria and dense clusters of Tir in the absence of TD4 (Fig 3). The high density of EHEC bacteria did not allow us to visualize individual bacteria with Tir clustering for quantification purposes. Nonetheless, we clearly observed that the presence of TD4 dramatically reduced the number of EHEC bound to HeLa cells, as well as the intensity of actin and Tir signals in those bacteria that were bound to the cells (Fig 3). Taken together, the above data showed that Nb TD4 reduces EHEC attachment, Tir clustering and actin polymerization by binding to the extracellular TirM domain exposed after Tir translocation. The Nb TD4 shows this inhibitory activity even when added once infection has begun.
One mechanism by which Nb TD4 could interfere with the attachment of EHEC to human cells is by directly competing with intimin for binding TirM. To investigate this, ELISA plates coated with purified TirM were incubated with biotinylated Int280 in the presence of different Nbs (Fig 4). While Int280:TirM interaction was not affected by the presence of camel pre-immune serum or Nb EC7 binding EspA (control), incubation with camel immune serum or Nb TD4 inhibited the interaction. In addition, Nb IB10 (anti-Int280) also reduced Int280:TirM interaction, but to a lesser extent than TD4. These results show that Nb TD4 is a potent inhibitor of Int280-TirM interaction.
To further characterize the binding of Nb TD4 to TirM and its inhibitory activity, we compared the affinities of Int280 and TD4 for TirM using surface plasmon resonance (SPR). Biotinylated TirM was immobilized onto a chip for SPR, and purified Int280 and TD4 Nb-HlyA fusion were passed over the chip at different concentrations in successive rounds of binding and regeneration. The change in resonance units (RU) with time was recorded as a direct indication of the binding of these proteins to TirM. The sensograms obtained are represented in Fig 5. These experiments revealed a distinct pattern of binding of Int280 and TD4 to TirM. While Int280 quickly bound to and dissociated from TirM after stopping Int280 injection (Fig 5A), TD4 bound to TirM more slowly and the interaction remained stable without any detectable dissociation even >300 sec after the injection stopped (Fig 5B). The kinetic constants of association (kon) and dissociation (koff) of Int280-TirM binding could be calculated directly from the obtained sensograms (Fig 5A). A model 1:1 Langmuir interaction fitted the binding curves, suggesting the formation of a 1:1 complex, as observed by protein crystallography of the EPEC Int280:TirM complex [15]. Using this binding model, we determined a koff of 3.75·10−2 s-1 and a kon of 7.85·105 M-1s-1 for EHEC Int280:TirM interaction. The equilibrium dissociation constant (KD) for EHEC Int280-TirM interaction was calculated from the ratio of these kinetic constants (koff/kon) and determined to be 48.1 nM. In contrast to Int280, the fact that TD4 had no detectable dissociation of TirM during SPR analysis impeded the determination of its kinetic constants kon and koff from the obtained sensograms. In addition, the KD could not be determined from RU values at equilibrium since the steady state was only reached at the highest concentration of TD4 (Fig 5B). Using the RU values closer to an apparent binding plateau at the different concentrations tested, we could estimate an apparent KD ~4.8 nM for the TD4:TirM interaction (Fig 5B). The actual KD for this interaction is likely to be below this estimated value (KD < 4.8 nM) as the actual steady state would be reached with higher RU values. Hence, this quantitative binding analysis indicated an at least 10-fold higher affinity of TD4 for TirM than the one of its natural ligand, Int280.
Using SPR we also investigated whether TD4 recognised an epitope of TirM overlapping the binding site of Int280, taking advantage of the extremely slow dissociation of TD4. We injected 40 nM of TD4 into the TirM-chip until reaching RU values close to steady state followed by 80 nM of Int280 (Fig 5C). We compared the increment of RU values obtained by Int280 injection in this condition (with bound TD4) with those obtained by injecting the same concentration of Int280 to the TirM-chip in the absence of TD4. This experiment showed that the RU values of Int280 binding to TirM were reduced in the presence of TD4, but binding of Int280 occurred simultaneously to TD4, indicating that the binding sites of Int280 and TD4 are not identical, although they could partially overlap. Interestingly, when Int280 injection was stopped, Int280 quickly dissociated whereas TD4 remained bound to TirM and the RU in the assay came back to those of TD4 binding alone. Thus, while Int280 quickly dissociates from TirM, TD4 remains stably bound to it.
To identify the specific binding site of TD4 to TirM, we synthesized 12-mer peptides of EHEC TirM covering its sequence, with a 10 amino acid (aa) overlap between consecutive peptides on a PVDF membrane. After incubation with TD4, bound Nb was subsequently detected with a secondary antibody. This identified two peptides recognized by TD4: VNIDELGNAIPS (aa 296–307) and GVLKDDVVANIE (aa 308–319) (Fig 6A). These consecutive peptides are localized within the interaction interface between Int280 and Tir (Fig 6B) [15–17, 40]. A BLASTP search [41] with non-redundant DNA sequences in databases (S1 Data) allowed us to determine that the 24-mer sequence of these peptides is 100% identical (BLASTP score 77.4 in S1 Data) in Tir proteins from all EHEC strains, including O157, O55, O145 and other relevant non-O157 serotypes [42].
The sequence of TirM is highly conserved among the related A/E pathogens EHEC, enteropathogenic E. coli (EPEC) and Citrobacter rodentium (CR) but the sequence of these Tir peptides is not identical in EPEC and CR strains, so we were therefore interested to determine the affinity of TD4 to purified TirMEPEC and TirMCR. We found that TD4 bound TirMCR with lower affinity (ca. 10-fold) than TirMEHEC. Surprisingly, no binding of TD4 to TirMEPEC was detected (Fig 6C). Comparing the aa sequences of the TD4 binding site in TirMEHEC with corresponding regions in TirMEPEC and TirMCR (Fig 6D) revealed that TirMEHEC differs from both TirMEPEC and TirMCR in residues V309, N317 and E319, suggesting that these changes may affect the affinity of TD4 towards TirM. Moreover, TirMEPEC specifically differs from TirMEHEC in residues E300, L301, V314 and A316, suggesting that these residues may be essential for TirM recognition by TD4.
We tested the effect of TD4 on EHEC binding to human colonic mucosa by employing IVOC. Human colonic biopsy samples were infected with EHEC in the presence or absence of TD4 or control Nb (Vamy). In addition, infection with EHECΔtir was included as a negative control. Immunostaining of biopsy samples showed a significant reduction in the number of adherent EHEC in the presence of TD4 but not of the control Nb (Fig 7). As expected, very few adherent bacteria were observed in biopsy samples infected with EHECΔtir (Fig 7). These results demonstrate that Nb TD4 reduces EHEC binding to human colonic mucosa ex vivo.
EHEC infections are associated with severe diseases such as bloody diarrhoea and HUS [1, 2]. Efficient therapies against EHEC infections are lacking, and current treatment is based on fluid replacement and supportive care [43]. However, increasing knowledge on EHEC virulence factors and infection mechanisms is contributing to the development of new treatment strategies [44], such as inhibition of quorum sensing [45], use of EHEC LPS-specific bacteriocins [46] and inhibition of Stx binding to its host receptor globotriaosylceramide (Gb3) with antibodies [47, 48] or other ligands [49].
In this work, we tested the possibility of using specific Nbs against the EHEC proteins EspA, intimin and Tir as an alternative approach to interfere with EHEC infection. Nb clones binding Int280 (IB10) or EspA (EC7) did not interfere with EHEC infection. Since intimin covers the entire surface of EHEC, binding of Nb IB10 in the concentration used might not be sufficient to mask all the Tir-binding sites, despite some inhibitory activity of this Nb in the in vitro Int280:TirM binding assay. Similarly, binding of Nb EC7 to EspA, which forms the translocation filament, did not affect EHEC infection nor inhibit Tir translocation. In contrast, a Nb binding TirM (TD4) reduced the attachment of EHEC and actin pedestal formation in HeLa cells. As the TirM domain is exposed on the host cell surface after Tir translocation [13, 14], binding of TD4 appears to block intimin binding. The fact that TirM is only presented on infected cells, suggests that a relatively low Nb concentration is needed for inhibition.
Staining of Tir after EHEC infection showed that TD4 hindered the formation of actin pedestals by preventing the characteristic Tir clustering produced at the bacterial:host interface [19, 50], which is achieved even after its addition 2 h post infection and is maintained for 6 h. In vitro protein interaction assays confirmed a strong inhibition of Int280:TirM interaction by the presence of TD4, which prevents Tir clustering by binding to TirM. SPR analysis of this interaction demonstrated an extremely slow dissociation rate of TD4. This analysis also revealed that the affinity for TD4 to TirM (KD ≤ 4 nM) is at least 10 times higher than the affinity of Int280 to TirM (KD ~40 nM). Strikingly, Int280 showed a fast dissociation of TirM, suggesting a dynamic interaction.
SPR experiments also determined that the TirM epitope recognized by TD4 could partially overlap with the binding region of Int280 as the addition of TD4 reduces the binding of Int280 to TirM. We mapped two TirM consecutive non-overlapping peptides bound by TD4: VNIDELGNAIPS (296–307) and GVLKDDVVANIE (308–319). It may be possible that each of these peptides is recognized by different CDRs of the Nb TD4, but this experiment does not exclude that TD4 may recognize a conformational structure of TirM. Its CDRs could still bind to the primary structure of these peptides, albeit with reduced affinity. Importantly, these recognized peptides are fully conserved in all Tir sequences from EHEC strains. We also determined that TD4 does not bind to TirMEPEC and has a weak interaction with TirMCR, which are highly similar but not identical to TirMEHEC. This information helped us to narrow the interaction site of TD4 and TirMEHEC by comparing the TirM sequences of the three pathogens. Differences between TirM of EHEC and CR—i.e. V309, N317 and E319- reduce but do not abolish the interaction with TD4. On the other hand, differences with TirM of EPEC—i.e. E300, L301, V314 and A316—could be critical for the binding of TD4, likely representing energetic hotspots of protein-protein interaction [51, 52].
We could further localize the residues that may participate in the interaction of TD4 with TirM based on the crystal structure of Int280 and Tir of EPEC [15]. Within the TirM sequence, it has been identified the so-called Int280-binding domain (IBD) [13], composed of two long alpha-helices (HA, residues 271–288, and HB, residues 312–331) separated by a ß-hairpin (residues 294–308). The described complex of EPEC reveals that the Int:Tir interaction is primarily mediated by the lectin-like D3 domain of Int280 and the ß-hairpin and the N-terminal part of the HB of Tir IBD, corresponding to residues 294–313 of TirMEHEC. The peptides identified to which TD4 binds (residues 296–319 of TirMEHEC), are enclosed within the IBD of Tir, indicating that TD4 is directly interfering with the Int:Tir interaction.
Importantly, we have shown that TD4 can also block the interaction of EHEC to intestinal human colonic tissue ex vivo [7], as the number of bacteria bound to the epithelium was significantly reduced in the presence of this Nb. This result opens the possibility of testing TD4 protection in humans, which could be administered using a passive immunization strategy. The fact that TD4 shows inhibitory activity once EHEC infection has already begun opens also the possibility of using this Nb as a therapeutic Ab to treat infections.
Nbs can be overproduced in bacteria, yeast, plants and mammalian cells to obtain highly concentrated purified proteins [53–56]. A purified Nb recognising EHEC toxins Stx1 and Stx2 has been administered, in combination with IgG, for the treatment of HUS [57]. However, the use of purified antibodies is a costly strategy for therapy development. To circumvent this problem, some studies describe the production of Abs and Nbs in edible plants and seeds. The production of Abs in edible tissues allows oral passive immunization at the gastric mucosal surface. For instance, a Nb against rotavirus infection has been expressed in rice and shown to protect infant mice from severe diarrhoea [58]. Abs contained in seeds enable long-term storage and the direct use for passive immunization with oral administration, which is particularly advantageous. Interestingly, a Nb against enterotoxigenic E. coli (ETEC) has been fused to the constant region (Fc) of immunoglobulins and produced in seeds. Piglets fed with these seeds were protected against ETEC infection [59, 60].
Alternatively, probiotic strains such as E. coli Nissle 1917 (EcN) [61] could be used for delivery of TD4 to the gastrointestinal tract. EcN is known to compete with EHEC for colonisation of the mouse intestine [62] through specific mechanisms including the secretion of microcins [61, 63]. Hence, secretion of TD4 by EcN could enhance its natural anti-microbial activity and leads to the development of a superior therapeutic strain against EHEC infection. Other probiotic bacteria can be considered for local delivery of Nb TD4. For instance, Gram-positive Lactobacillus strains producing surface-bound or secreted Nbs against rotaviruses have been shown to reduce the severity and duration of rotavirus-induced diarrhoea in mice [64–66].
Overall, this study demonstrates that a Nb recognising Tir reduces intimate attachment of EHEC to human cells and colonic tissue by competing with its natural partner, intimin, thereby preventing colonization of the epithelium. These results open the possibility for passive immunization and therapeutic strategies that could prevent EHEC adhesion to intestinal tissues during infection. This could also be applied to reduce the prevalence of EHEC in its natural bovine host and minimize the risk of EHEC contamination into the food chain.
This study was performed with approval from the University of East Anglia Faculty of Medicine and Health Ethics Committee (ref 2010/11-030). All samples were registered with the Norwich Biorepository which has approval from the National Research Ethics Service (ref 08/h0304/85+5). Biopsy samples from the transverse colon were obtained with informed written consent during colonoscopy of adult patients. All samples were anonymized.
All E. coli strains used in this work are listed in Table 1. Bacteria were grown at 37°C on Lysogeny broth (LB) agar plates (1.5% w/v), in liquid LB or Dulbecco’s Modified Eagle’s Medium (DMEM). Ampicillin (Ap, 150 μg/ml), Chloramphenicol (Cm, 30 μg/ml) and Kanamycin (Km, 50 μg/ml) were added for plasmid selection as required. For infection of HeLa cells, EHEC strains were grown for 8 h at 37°C (200 rpm) in a flask with 10 mL of liquid LB, inoculated in capped Falcon tubes (BD Biosciences) with 5 mL DMEM, and incubated o/n at 37°C in a CO2 incubator (static) for the induction of the T3SS. For infection of biopsy samples, 2 ml of LB media were inoculated with an EHEC colony from an LB-agar plate and grown standing at 37 oC overnight (o/n).
Plasmids used in this study are listed in Table 1. Strain E. coli DH10B-T1R was used as a host for the cloning and propagation of plasmids. TD4-HlyA and Vamy-HlyA DNA fragments were excised with BglII from pEHLYA5-TD4 and pEHLYA5-Vamy, respectively, and cloned into the same site of pVDL9.3 [71]. TirM sequences of EPEC (aa 255–363) and CR (aa 253–360) were amplified by PCR using primers listed in Table 2, cloned after EcoRI-HindIII digestion into the same sites of pET28a plasmid backbone. The TirM constructs in this plasmid are under the T7 promoter and fused to an N-terminal His-tag for purification. PCR reactions were performed with Taq DNA polymerase (Roche, NZyTech) for standard amplifications in screenings. All DNA constructs were fully sequenced (Secugen SL, Madrid, Spain).
Cultures of E. coli BL21(DE3) carrying the corresponding pET28a-derivative were grown at 30°C in 500 ml of LB with Km to an optical density at 600 nm (OD600) ~0.5 and subsequently induced with 1 mM isopropyl-1-thio-β-D-galactoside (IPTG) for 2 h. Bacteria were harvested by centrifugation (10 min, 10,000 x g, 4°C), resuspended in 20 ml of 50 mM NaPO4 pH 7, 300 mM NaCl, DNase (0.1 mg/ml; Roche) and protease inhibitor cocktail (Roche), and lysed by passing through a French-Press at 1200 psi three times. The resultant lysate was ultracentrifuged (60 min, 40000 x g, 4°C) to obtain a cleared lysate supernatant. For purification of the His-tagged Int280EHEC, TirMEHEC, TirMEPEC and TirMCR, lysates were passed through 2 ml of pre-equilibrated Cobalt-containing resin (TALON, Takara) in a chromatography column and washed with 20 mM HEPES pH 7.4, 200 mM NaCl. The bound His-tagged proteins were eluted by adding the same buffer complemented with 150 mM imidazole. The eluted fractions were dialyzed against HEPES-buffer (sterile filtered and degassed) and concentrated 10-fold in a 3-kDa centrifugal filter unit (Amicon Ultra-15). Proteins were loaded onto a gel filtration column (HiLoad 16/600 Superdex 75 preparative grade, GE Healthcare), pre-equilibrated with HEPES-buffer and calibrated with protein markers (Gel Filtration Standards, Bio-Rad) and Blue dextran (for exclusion volume Vo; Sigma). Fractions of 1 ml containing the purified proteins were collected and checked for purity by SDS-PAGE. Protein concentration was estimated using the Bicinchoninic acid protein assay kit (Thermo Scientific).
Cultures of E. coli strain HB2151 carrying pVDL9.3 (hlyB hlyD) and the indicated pEHLYA5-derivative, or pVDL9.3-derivatives with Nb-HlyA fusions (Table 1), were grown o/n at 30°C (170 rpm) in liquid LB with the appropriate antibiotics. Next, bacteria were inoculated in fresh medium (200 ml of liquid media in 1L flask) and grown at 37°C (170 rpm) until OD600 reached 0.4. At this point, bacteria were induced with 1 mM (IPTG and further incubated for 6 h with shaking (100 rpm). The cultures were centrifuged twice (10 min, 10000 g, 4°C) to retrieve the supernatants, which were mildly sonicated (3 pulses of 5 seconds) and filtered (0.2 μm syringe filters). Then, they were loaded in columns for metal affinity chromatography (IMAC) purification. The supernatants were loaded at ca 4 ml/min onto chromatography columns with pre-equilibrated Cobalt-containing resin (TALON, Takara). Columns were washed with Tris pH 7.5 (50 mM) NaCl (150 mM) or HEPES buffer and eluted with a gradient of imidazole reaching 500 mM. A second purification step by gel filtration was performed for His-tagged antigens and Nb-HlyA fusions used in Surface Plasmon Resonance (SPR). Fractions eluted from metal-affinity chromatography were dialysed against HEPES-buffer (sterile filtered and degassed) and concentrated to 2 ml in a 3 kDa centrifugal filter unit (Amicon Ultra-15, Millipore). Next, protein samples were loaded onto a calibrated gel filtration column (HiLoad 16/600 Superdex 75, GE Healthcare), pre-equilibrated with HEPES-buffer. The elution of Nb-HlyA proteins was performed using HEPES buffer and collecting 1 ml fractions. Protein concentration was estimated using the BCA protein assay kit (Thermo Scientific).
For the generation of 12-mer TirMEHEC peptides on a PVDF membrane, a MultiPep RSi synthesizer (Intavis) with SPOT module (Proteomics Service, CNB-CSIC) was used. The resulting membrane was blocked in PBS containing 0.1% Tween 20 (PBST) and 3% (w/v) skimmed milk for 1 h at room temperature (RT) and subsequently incubated in purified TD4-HlyA dissolved in PBST, 3% skimmed milk for 2 h. After washing in PBST, the membrane was sequentially incubated with anti-E tag mAb (Phadia, 1:5000) and secondary rabbit anti-mouse IgG-POD (1:5000, Sigma). Signal detection was performed using the Clarity Western ECL Substrate kit (Bio-Rad) and exposure to X-ray films (Agfa).
ELISA was performed as described previously [39]. Briefly, 96-well immunoplates (Maxisorp, Nunc) were coated for 2 h at RT with 5 μg/ml of purified TirM (from EHEC, EPEC or CR, as indicated) diluted in PBS. Bovine serum albumin (BSA, Roche) was used as a negative control antigen. Nb-HlyA fusions were added at the indicated concentrations for 1 h and plates were subsequently washed three times with PBS. For detection of bound Nb-HlyA fusions, anti-E-tag mAb (1:2000; Phadia) and anti-mouse IgG-POD (1:2000; Sigma), as secondary antibody, were added. The reaction was developed with o-phenylenediamine (Sigma) and H2O2 (Sigma), as previously reported [72], and the OD490 was determined using a microplate reader (iMark ELISA plate reader, Bio-Rad).
For the neutralization assay, 1 mg/ml Int280 was biotinylated using a 20-fold molar excess of Biotinamidocaproate N-hydroxysuccinimide ester (Sigma). After incubation on a gyratory wheel for 1 h at RT, the reaction was stopped by addition of 50 mM Tris-HCl pH 7.5, and placement on ice for 1 h. The reaction mix was subsequently loaded onto a pre-packed column for gel filtration chromatography (Sephadex G25 PD-10; GE Healthcare) and the biotinylated protein was eluted in 500-μl fractions with PBS. Protein concentrations were estimated using the BCA protein assay kit (Thermo Scientific). For the assay, 5 μg/ml non-biotinylated TirM was bound to plastic 96 wells plates for 2 h. The wells were blocked with 3% (w/v) skimmed milk in PBS for 1 h. At the same time, biotinylated Int280 (50 μg/ml) was incubated with a 1:50 dilution of the camel immune or preimmune serums or 1 μM (50 μg/ml) of the corresponding purified Nb-HlyA. These solutions were added to the microtiter wells for 1 h incubation after removing the blocking solution. Then, the wells were developed as a standard ELISA using Streptavidin-POD (Roche, Sigma).
SPR experiments were performed using BiaCore3000 (GE Healthcare). All proteins solutions were dialyzed against HEPES-buffer (sterile filtered and degassed) at 4°C for 2 h. TirMEHEC was biotinylated (as described above) at 0.1 μg/ml and immobilised on a Streptavidin SA chip (GE Healthcare) at 150 response units (RU) at a flow rate of 10 μl/min in HEPES-buffer containing 0.005% (v/v) of the surfactant Polysorbate 20 (P20, GE Healthcare). To determine binding kinetics, dilutions of purified TD4-HlyA or Int280 (as indicated) were run at 30 μl/min in HEPES-buffer and sensograms were generated. Regeneration of TD4-HlyA was performed by sequential injections of 10 μl 10 mM glycine-HCl pH 1.7, 5 μl 5 mM NaOH and 10 μl 10 mM glycine-HCl pH 1.7. No regeneration was needed for Int280. Sensograms with different concentrations of analyte were overlaid, aligned and analysed with BIAevaluation 4.1 software (GE Healthcare) under assumption of the 1:1 Langmuir model and using both the simultaneous kinetics model and the steady-state equilibrium analysis [73].
The human cervix carcinoma cell line HeLa (ATCC, CCL-2) was grown in DMEM supplemented with 10% fetal bovine serum and 2 mM glutamine at 37°C in a 5% CO2 atmosphere. For infection, cells were seeded out on glass coverslips in 24-well plates at a concentration of 105 cells/well. Cells were inoculated with EHEC at a multiplicity of infection (MOI) of 1000 for 3 h at 37°C in a 5% CO2 atmosphere. The purified Nb-HlyA fusions, at the indicated concentrations, were added to the cells simultaneously with EHEC bacteria, or 1 h or 2 h post-infection, as indicated, in a final volume of 0.5 or 1 ml. The infection was stopped by three washes with sterile PBS. In the case of EHEC infections for 6 h, cells were washed with PBS after 3 h of infection, fresh medium and Nbs were then added, and incubation was continued for another 3 h.
Cells were fixed with 4% (w/v) paraformaldehyde in PBS for 20 min at RT and permeabilized in 0.1% (v/v) of saponin (Sigma) in PBS for 10 min. All antibodies were diluted in PBS with 10% goat serum (Sigma), and mouse monoclonal anti-O157 (Abcam, 1:500), mouse monoclonal anti-HA (Cambridge bioscience, 1:200) and rabbit polyclonal anti-TirEHEC (1:200) were used to detect EHEC bacteria, HA-tag and TirEHEC, respectively. After incubation for 1 h at RT, coverslips were washed three times with PBS, and incubated for 45 min with secondary antibodies, Alexa477-conjugated goat anti-mouse IgG or Alexa647-conjugated goat anti-rabbit-IgG (1:500, ThermoFisher Scientific), Tetramethylrhodamine (TRITC)-conjugated phalloidin (1:500, Sigma) and 4',6-Diamidino-2-phenylindole (DAPI) (1:500, Sigma) to label F-actin and DNA, respectively. Coverslips were washed 3 times with PBS after incubation, mounted in of ProLong Gold anti-fade reagent (ThermoFisher Scientific), and analysed with an SP5 confocal microscope (Leica).
Biopsy samples from the transverse colon were taken from macroscopically normal areas, transported to the laboratory in IVOC medium and processed within the next hour. IVOC was performed as described previously [74]. Briefly, biopsies were mounted on foam supports in 12 well plates and incubated with 30 μl EHEC standing overnight culture (approximately 107 bacteria) and 200 nM of TD4 or Nb control (Vamy). Samples were incubated for 8 h on a rocking platform at 37°C in a 5% CO2 atmosphere with medium changes after 4 and 6 h of incubation. At the end of the experiment, tissues were washed in PBS to remove the mucus layer and fixed in 3.7% formaldehyde/PBS for 20 min at RT. Samples were permeabilised with 0.1% Triton X-100/PBS, and blocked with 0.5% BSA/PBS for 20 min. Tissues were incubated with goat polyclonal anti-E. coli (1:400, Abcam) for one hour, followed by incubation in Alexa Fluor 568-conjugated donkey anti-goat IgG (1:400,ThermoFisher Scientific) and DAPI for 30 min to counterstain cell nuclei. Biopsy samples were mounted with Vectashield mounting medium (Vector Labs) and analysed using an Axio Imager M2 motorized fluorescence microscope (Zeiss). EHEC colonisation of colonic biopsies was quantified by counting adherent bacteria in a surface area of 1 mm2.
Means and standard errors of experimental values were calculated using Prism 5.0 (GraphPad software Inc). Statistical analyses comparing multiple groups were performed using one-way ANOVA and Dunnett’s post- test. Statistics for Fig 2 were done using One Way ANOVA analysis doing logarithms for normal distribution. Data was corrected with the Bonferroni test. A value of p<0.05 was considered significant.
|
10.1371/journal.pntd.0005124 | Oral Cholera Vaccination Delivery Cost in Low- and Middle-Income Countries: An Analysis Based on Systematic Review | Use of the oral cholera vaccine (OCV) is a vital short-term strategy to control cholera in endemic areas with poor water and sanitation infrastructure. Identifying, estimating, and categorizing the delivery costs of OCV campaigns are useful in analyzing cost-effectiveness, understanding vaccine affordability, and in planning and decision making by program managers and policy makers.
To review and re-estimate oral cholera vaccination program costs and propose a new standardized categorization that can help in collation, analysis, and comparison of delivery costs across countries.
Peer reviewed publications listed in PubMed database, Google Scholar and World Health Organization (WHO) websites and unpublished data from organizations involved in oral cholera vaccination.
The publications and reports containing oral cholera vaccination delivery costs, conducted in low- and middle-income countries based on World Bank Classification. Limits are humans and publication date before December 31st, 2014.
No participants are involved, only costs are collected.
Oral cholera vaccination and cost estimation.
A systematic review was conducted using pre-defined inclusion and exclusion criteria. Cost items were categorized into four main cost groups: vaccination program preparation, vaccine administration, adverse events following immunization and vaccine procurement; the first three groups constituting the vaccine delivery costs. The costs were re-estimated in 2014 US dollars (US$) and in international dollar (I$).
Ten studies were identified and included in the analysis. The vaccine delivery costs ranged from US$0.36 to US$ 6.32 (in US$2014) which was equivalent to I$ 0.99 to I$ 16.81 (in I$2014). The vaccine procurement costs ranged from US$ 0.29 to US$ 29.70 (in US$2014), which was equivalent to I$ 0.72 to I$ 78.96 (in I$2014). The delivery costs in routine immunization systems were lowest from US$ 0.36 (in US$2014) equivalent to I$ 0.99 (in I$2014).
The reported cost categories are not standardized at collection point and may lead to misclassification. Costs for some OCV campaigns are not available and analysis does not include direct and indirect costs to vaccine recipients.
Vaccine delivery cost estimation is needed for budgeting and economic analysis of vaccination programs. The cost categorization methodology presented in this study is helpful in collecting OCV delivery costs in a standardized manner, comparing delivery costs, planning vaccination campaigns and informing decision-making.
| We reviewed and re-estimated oral cholera vaccine delivery costs in low and middle income countries standardizing cost categories. The cost categorization proposed here can help in collation, analysis, comparison and economic analysis of OCV delivery costs across countries.
| Cholera is transmitted through the fecal-oral route, and humans are the natural host. It is caused by the ingestion of O1 and less commonly O139 serogroups of the Vibrio cholerae bacterium and characterized by severe, potentially life-threatening diarrhea [1]. The disease inflicts a significant health burden on many low-and-middle-income countries (LMICs) in settings where food and water are contaminated with human feces. Infrastructure disruption resulting from natural disasters, civil unrest, and war often precipitates cholera outbreaks, particularly in settings where there is endemic cholera risk. Cholera outbreak risk is further increased when infrastructure disruption is superimposed on the poor sanitation and unsafe drinking water found in parts of Africa, Asia, and South and Central America [2]. While improving water and sanitation infrastructure would greatly enhance the control of cholera in the long-term, the use of preventive vaccines has shown promise in the interim [3–5].
The struggle to develop a safe and effective cholera vaccine that can prevent and control the disease has a long history. Injectable whole-cell cholera vaccines were developed as early as the 19th century and extensively used in the 20th century in the Indian subcontinent and later abandoned due to their limited efficacy and systemic adverse events [6,7]. Subsequently, a new generation of live-attenuated or killed oral cholera vaccines were developed, licensed, and deployed. A killed whole-cell cholera vaccine with recombinant B subunit of cholera toxin (Dukoral) was licensed in 1991 (two-dose regimen for >2 years of age) [6] and used by travelers visiting cholera-endemic regions. This vaccine received World Health Organization (WHO) prequalification in 2001 and has a price of $5 per dose on the public market. Meanwhile, Vietnam developed and deployed a locally manufactured OCV, ORC-Vax [8]. The vaccine was licensed in 1997 in Vietnam and was modified to mORC-Vax in 2009 after improving the production process. Currently, the price of this vaccine is US$1.25 per dose on Vietnam’s public market. At the same time, international efforts were made to reformulate ORC-Vax into a less expensive modified killed whole-cell OCV, which was first licensed in India in 2009 (Shanchol, two-dose regimen for >1 year of age), and later WHO-prequalified in 2011. Currently, the price of this vaccine is $1.85 per dose on the public market worldwide. A WHO OCV stockpile was then created in 2013 to make the vaccine available and affordable in emergency settings [9,10]. These two WHO-prequalified OCVs, Dukoral and Shanchol have been deployed in mass vaccination campaigns across many endemic regions either pre-emptively or reactively; notably in Haiti, Comoros, Indonesia, Uganda, Mozambique, Tanzania, India, Bangladesh, Guinea, South Sudan, Malawi, Thailand, Ethiopia and Nepal [11–21].
A cholera vaccination can be broken down into several small and large activities or actions. Understanding the activities involved in vaccination campaigns and estimating cost of each key activity is vital in planning and deployment of OCVs. When deploying a new vaccine, besides routine recurrent costs, the introduction cost such as initial planning, extra logistics and cold chain, training, social mobilization, sensitization, and other new implementation activities such as management of Adverse Events Following Immunization (AEFI) should be considered [22]. Analysis of cost items helps to identify major cost drivers in mass vaccination programs which are critical elements in planning program implementation. This research intends to assess the costs of the different activities required for OCV delivery in LMICs based on systematic literature search and collection of unpublished data from organizations involved in oral cholera vaccination. We propose to categorize cost items in a standardized method and re-estimate delivery costs. Through this analysis we recommend a standardized cost-collation approach for OCV campaigns that can be used in developing OCV delivery cost-estimation tools and comparing costs across different geographical regions.
A systematic literature review was conducted using search terms (vaccination cost) AND (cholera) in Medline database through PubMed restricting search to humans and dated up to December 31, 2014. Detailed search terms are (("vaccination"[MeSH Terms] OR "vaccination"[All Fields]) AND ("economics"[Subheading] OR "economics"[All Fields] OR "cost"[All Fields] OR "costs and cost analysis"[MeSH Terms] OR ("costs"[All Fields] AND "cost"[All Fields] AND "analysis"[All Fields]) OR "costs and cost analysis"[All Fields])) AND ("cholera"[MeSH Terms] OR "cholera"[All Fields]) AND (("0001/01/01"[PDAT]: "2014/12/31"[PDAT]) AND "humans"[MeSH Terms]). After initial screening on title and abstract, studies using Dukoral, ORC-Vax and Shanchol conducted in LMICs as per the World Bank’s classification [23] that quantified the costing items in cholera vaccination were included. We excluded costing or cost-effectiveness analyses that used simulated or assumed costs, studies that referred to traveler’s vaccination, and studies that considered vaccination in developed countries. The systematic review followed PRISMA guidelines [24] (S1 Checklist). In addition, to find unpublished literature, we searched the Google Scholar and WHO website for OCV mass campaign-related publications and contacted organizations involved in OCV campaigns, including the International Vaccine Institute (IVI), Medecine sans Frontieres (MSF), and the US Centers for Disease Control and Prevention to obtain available reports.
We categorized cost items into four groups with subcategories in each based on the chronological order of implementing OCV campaigns using standardized definitions (Table 1). Vaccination program preparation costs were incurred in field capacity building which includes microplanning, training of personnel, community sensitization, social mobilization and other costs like the storage of vaccines in central warehouses prior to vaccination implementation. Vaccine administration costs included actual vaccine administration costs in the field to individuals as well as transportation of the vaccines from central warehouse to field headquarters and to vaccination field sites. The cost items included are conveyance, per-diem, logistic arrangement, equipment, and location costs for vaccine administration, supervision and monitoring. Finally, all costs related to the AEFI management were included under this category. The last three categories constitute vaccine delivery costs. The vaccine procurement costs included cost of vaccine purchase at preclearance and add-on which comprised costs of freight, insurance, taxes, and customs. This categorization allows comparison of cholera vaccination campaign expenditures across countries that have deployed the vaccines. The financial costs of OCV campaigns were used in our analysis as no opportunity costs were taken into consideration.
After categorizing costs from each paper we summarized the results and presented overall vaccination program costs as the sum of all four cost categories. As vaccines are often donated to countries, we differentiated overall vaccination program costs and vaccine delivery costs by segregating vaccine procurement costs. We estimated program cost and vaccine delivery cost per person for complete vaccination using three methods: 1) in United States Dollars (US$) as reported in the literature for the campaign year, 2) in 2014 US$ after adjusting for country level inflation and current exchange rate, and 3) in 2014 international dollars (I$) after adjusting for country level inflation and current purchasing power parity. The year 2014 was selected as the base year for cost analysis.
In Method 1, we presented costs as recorded by the investigators for the campaign year (campaign year cost) in US$. Costs in local currency units (LCU) were converted to US$ based on the World Bank exchange rate reference database for that year [25]. In Method 2, we adjusted the base year costs to 2014 US$ cost-equivalent by first converting the costs to LCU for the vaccination year using the US$-LCU exchange rate for that year and inflating it to year 2014 based on the country inflation rate (inflation, consumer prices, annual %) using the World Bank inflation data [25]. The adjusted results were presented in US$ 2014 after converting LCU to US$ based on the 2014 exchange rate. In Method 3, we adjusted the campaign year cost to the 2014 I$ cost-equivalent by first converting the costs to LCU for the vaccination year, and then inflating to the year 2014 as described for Method 2. The adjusted results were presented in I$ after converting LCU to I$ for 2014 [26].
We employed three methods of program cost estimation for two reasons. First, costs from different campaign years and different countries are not comparable and therefore need to be adjusted to the same base year in order to eliminate inflation effects [27]. Second, the exchange rate conversion does not always consider the differences in the cost of living between countries [28]. For example, the vaccination personnel costs (e.g., per diem) vary by country, which cannot be adequately captured in US$. Purchasing power parity expressed in I$, defined as the number of units of a country’s currency required to buy the same amounts of goods and services in the domestic market as US$ would buy in the United States [28], allows comparison across countries.
We identified 83 papers on PubMed search, of which eight were included based on the inclusion-exclusion criteria and two papers were obtained from other sources and personal communications (S1 Flowchart). The program costs for Shanchol delivery were available from four countries (five campaigns) that deployed the vaccine in 2011 and 2014 (Table 2) [15–17,21,29]. A publication presented OCV delivery cost summary for a campaign conducted in three internally displaced people (IDP) camps in 2014 in South Sudan without detailed cost categorization [21]. We also obtained more detailed delivery cost for another OCV campaign conducted in South Sudan in 2013 from personal communications [29]. We were aware that in Ethiopia and Malawi, OCV campaigns were conducted in 2015 and delivery costs were estimated [18], but data was unavailable. The program costs for Dukoral were available from four countries that deployed vaccines from 1997 to 2009 [12–14,20]. The data for Indonesia was obtained from WHO website [20]. In reference review we found one paper presenting a brief description of a Dukoral campaign in Darfur in 2014 stating direct cost of vaccination was US$336,527 or US$ 7 per full vaccinated person [30]. We had to exclude this study from further analysis because costs could not be categorized. The program costs for ORC-Vax were available from Vietnam that deployed vaccine in a 1998 campaign [31].
The number of fully vaccinated people per campaign ranged from 23,751 in India to 143,706 in Guinea. The reported price per dose of Shanchol procurement varied from US$ 1.00 to US$ 2.40 while the price of Dukoral ranged from US$ 0 (free) to US$ 5.00.
The cost categorization and presentation was inconsistent across the studies conducted as shown in Table 2, limiting their comparability. AEFI management, micro-planning, training, sensitization and social mobilization were often mentioned as activities, but costs were merged with other categories which we could not de-merge. In India, micro-planning was considered as a management activity for existing staff and costs were excluded. In Bangladesh and South Sudan, expenses on office furniture and office supplies were categorized under micro-planning while Indonesia had purchased a computer, which was classified as costs for micro-planning. Costs related to AEFI management were reported only from India, representing 15.73% of the total OCV delivery costs.
Of the four major cost categories, vaccine procurement was the costliest component in all OCV campaigns. Whereas, the vaccine administration was the costliest item under vaccine delivery costs (Table 2). Staff salary and allowance when reported, material and supplies such as vaccination card followed by plastic cups, water and soap, when used, were the cost drivers for vaccine administration. Staff salary and allowances are among the costliest items ranging from 23.6% to 87.8% of delivery costs in Guinea and India, respectively. The material and supplies costed 43.9%, 42.1% and 37.0% of delivery costs in Tanzania, South Sudan (2013) and Uganda respectively.
Costs re-estimated in US$2014 shows high variability of delivery costs across the sites (Fig 1, Table 2). The variability of delivery costs within the same country was also prominent. Although average delivery cost of OCV campaign in South Sudan IDP camps was $1.72 in US$ 2014 (I$6.57) [21], the costs in three different IDP camps were $1.28(I$4.88), $2.02(I$7.71) and $3.38(I$12.89). When prices were adjusted to I$, the costs of vaccination program increased substantially in all settings enhancing the variability across the sites. For example, the cost per unit of OCV delivery increased from US$ 4.70 to I$ 21.77 in Indonesia.
When there is a cholera outbreak or an impending outbreak, there are three main intervention options besides management of cases and public awareness: Do nothing, water and sanitation improvement, and cholera vaccination. The cost of inaction against cholera outbreak could be substantial. One study reported that the drop in exports alone results in substantial trade loss accounting up to 1% of the countries’ GDP [32]. Besides exports, the economic impact of a cholera outbreak includes tourism revenue loss, treatment expenditures and loss of income for those who are affected because they are unable to work. Water and sanitation improvement remains the choice of intervention for cholera and other diarrheal disease control, but requires large investments and takes long-term except personal level interventions such as provision of soap for hand wash and chlorine for water purification. The investments needed for upgrading water and sanitation system is difficult to measure and estimates widely vary. One study estimated that the access to regulated in-house piped water supply with quality monitoring and in-house sewerage connection with partial treatment of sewage for all would require a total investment of US$136.5 billion per year [33]. Oral cholera vaccination is the interim intervention that is effective against cholera, at least in short-term. Accordingly, many OCV campaigns have taken place in different parts of the world, but the costs from those studies have been categorized and presented differently. A well-defined and limited set of basic categories may be more helpful to investigators, health authorities, policy makers, vaccination planners, and community stakeholders. The categories described herein allow for a clear, comparative understanding of vaccination campaign costs that can better guide decision-making.
The delivery costs of OCV through mass campaigns differed by country and even within the same country and same settings. The delivery cost of Shanchol in US$ 2014 varied from $1.14 in India to $3.05 in South Sudan per fully immunized person. The difference was higher in I$2014, ranging from I$4.08 in India to I$14.39 in South Sudan. Some of those differences could be because of the difference in provisions and activities during vaccination as discussed below, while other factors could be that the costs are collected and reported differently. However, the costs of OCV delivery in US$ 2014 in three different IDP settings in South Sudan ranged from $1.28 to $3.38 per fully immunized person. This cost difference could be partially attributed to the scale of vaccination, lowest costs of $1.28 was at IDP camp that vaccinated 38,200 people compared to the highest costs of $3.38 was at IDP camp that covered only 7,400 people. Once cost collection, categorization and presentations are standardized, the costs in US$ should help donors and financing bodies to decide the comparative resources required for vaccination in various settings. Whereas the costs in I$ will be helpful in comparing delivery costs across countries when in-country resources are used partially or completely as the case in many preemptive vaccinations in endemic settings.
Vaccine delivery costs were generally higher for Dukoral than Shanchol, with the exception of Uganda. The higher cost of Dukoral delivery could be partially because of its buffer requirement which complicates vaccine administration process requiring more materials and supplies. A higher proportion of delivery costs are constituted by materials and supply as reported in Tanzania (43.9%), Uganda (37.0%) and Mozambique (28.3%) where Dukoral was used. In Tanzania the material and supply costs were high because it included domestic vaccine storage and transport costs. The delivery costs in Uganda was lower because it did not include costs for program preparation (micro-planning, training, social mobilization & sensitization and other preparation costs) and cold chain costs as the campaign used existing cold chain system for vaccine storage at operational headquarters (Entebbe) and did not use cold chain at field level (Adjumani). Also, vaccination coverage and acceptability survey, and AEFI data collections were not accounted for in the costs. In Uganda, Mozambique and Indonesia the vaccine was air-delivered to the site of vaccination due to difficulty in transport or security reasons or due to the fear of breaking the cold chain, which added substantially to the costs.
The costs of Shanchol delivery were highest in 2013 campaign in South Sudan compared to the other three countries that deployed the vaccine. In South Sudan 2013 campaign, items such as soap, cup and water was provided to Shanchol recipients, resulting in increased proportion of material and supply costs (42.1%). This was higher than other sites such as Bangladesh (13.6%) where these provisions were a part of a research activity and in Guinea (14.3%) where such provisions were part of outbreak preventive measures. Costs for Shanchol delivery were next highest in Guinea because of the transport costs where vaccination teams were mobile on car or boat. The proportion of staff salary and allowances as a part of vaccine administration cost was high in India (87.8%) and Bangladesh (62.7%) compared to Guinea (23.6%) and South Sudan (25.3%). The proportion was high in India because it included staff training costs and in Bangladesh it included costs for pre-vaccination census and intensive house to house mop-up vaccination.
The overall costs for administering ORC-Vax was relatively cheaper in Vietnam compared to other OCVs partly due to the fact that it was integrated into their routine immunization system [31]. This suggests that the routine administration of OCV through existing immunization systems may reduce the vaccine delivery costs. However, the staff costs as a proportion of vaccine administration costs were relatively high (75.9%).
Vaccine procurement accounts for the highest proportion of the total vaccination program costs, the majority of which is due to the cost of the vaccine itself. Even if a country receives donated OCVs, international transportation of the donated vaccines to its borders as well as the clearance of the vaccines at the point of entry accounts for a sizable proportion of the costs. Besides scaling up vaccine supply through the entry of multiple competitive manufacturers [34], a single dose vaccine strategy, if deployed, particularly in outbreak settings is likely to lower vaccine costs [35]. The vaccine administration cost was the next highest because it involves intensive efforts to reach each individual to be vaccinated that needs lot of human resources, cold chain and materials such as vaccination card, soap, water and cups. As Shanchol does not need buffer, administering vaccine without provision of water and keeping OCV outside the cold chain could reduce some of these costs [36].
There are several limitations in our analysis. First, the studies analyzed included only direct costs. The indirect costs—such as loss of income and transportation costs for those who spend time to visit a vaccination post [37] are not accounted for. Adding these vaccine recipient costs would be valuable for a better understanding of the total costs of vaccination and help in developing plans to reduce vaccine recipient’s costs which may improve vaccine acceptance [38]. Two of the Shanchol costing studies later published OCV delivery costs under societal perspective [37,39]. Second, several of the studies included in our analysis organized the costs using their own methods and costing categories, which made it difficult to reorganize the costs for the purposes of our analysis [12–14,20]. This insufficient and unclear information may have resulted in some misclassification of cost categories. Third, we only could include financial costs, not economic costs as most studies presented financial costs. Inclusion of economic costs in future studies is important to understand all the costs-involved in conducting OCV campaigns and also to conduct cost-effectiveness analysis. Fourth, the presentation of costs in selected papers did not allow us to differentiate between fixed and variable costs. The fixed costs will not be affected by a larger OCV introduction while, the variable costs per unit will be further reduced by scaling up the program. It is important to identify and list fixed and variable costs in future costing studies. Fifth, the scope of this work was confined to vaccine delivery costs and it does not include value for money measures such as cost-effectiveness analysis. The reviews on health economic evaluations around OCV delivery will be useful in informed decision making. Finally, availability of unpublished data from two sites (Malawi and Ethiopia) would have improved the cost estimation.
Understanding the costs of cholera vaccination campaigns is of paramount importance in the economic evaluation as well as in planning future vaccination programs. Currently, there is limited OCV delivery cost data, collected inconsistently and reported capriciously limiting the comparability of costs across settings. Categorizing the costs into easily differentiable categories is useful to the planning process and comparison between campaigns. We recommend that future OCV costing studies include both financial and economic costs and use the cost categories defined in this study for clearer collation, analysis, and comparison of campaign costs.
|
10.1371/journal.pgen.1000426 | Independent S-Locus Mutations Caused Self-Fertility in Arabidopsis thaliana | A common yet poorly understood evolutionary transition among flowering plants is a switch from outbreeding to an inbreeding mode of mating. The model plant Arabidopsis thaliana evolved to an inbreeding state through the loss of self-incompatibility, a pollen-rejection system in which pollen recognition by the stigma is determined by tightly linked and co-evolving alleles of the S-locus receptor kinase (SRK) and its S-locus cysteine-rich ligand (SCR). Transformation of A. thaliana, with a functional AlSRKb-SCRb gene pair from its outcrossing relative A. lyrata, demonstrated that A. thaliana accessions harbor different sets of cryptic self-fertility–promoting mutations, not only in S-locus genes, but also in other loci required for self-incompatibility. However, it is still not known how many times and in what manner the switch to self-fertility occurred in the A. thaliana lineage. Here, we report on our identification of four accessions that are reverted to full self-incompatibility by transformation with AlSRKb-SCRb, bringing to five the number of accessions in which self-fertility is due to, and was likely caused by, S-locus inactivation. Analysis of S-haplotype organization reveals that inter-haplotypic recombination events, rearrangements, and deletions have restructured the S locus and its genes in these accessions. We also perform a Quantitative Trait Loci (QTL) analysis to identify modifier loci associated with self-fertility in the Col-0 reference accession, which cannot be reverted to full self-incompatibility. Our results indicate that the transition to inbreeding occurred by at least two, and possibly more, independent S-locus mutations, and identify a novel unstable modifier locus that contributes to self-fertility in Col-0.
| The mating system adopted by a species has a profound influence on extent of polymorphism, population structure, and evolutionary potential. In flowering plants, the switch from outbreeding to inbreeding has occurred repeatedly, yet little is known about the underlying genetic events. This is true even for the model species A. thaliana, a highly self-fertile member of the crucifer family. In this family, outbreeding is enforced by a self-incompatibility system controlled by the S locus, which involves the recognition of pollen by the stigma to prevent self-fertilization and familial inbreeding. We recently demonstrated that A. thaliana accessions may be reverted to full or partial self-incompatibility by transformation with S-locus genes isolated from its close self-incompatible relative A. lyrata. Despite much recent debate, however, we still do not know how A. thaliana became self-fertile. Here, we use our recently established A. thaliana transgenic self-incompatible experimental model to address these issues. Analysis of the S locus in accessions that can be reverted to full self-incompatibility demonstrates that self-fertility in A. thaliana arose by at least two independent S-locus mutations. Furthermore, analysis of an accession that expresses only partial self-incompatibility shows that self-fertility is associated with an unstable allele at a locus unlinked to the S locus.
| Sexual reproduction may have evolved because it can combine different sequence variants through recombination [1] and because it can remove deleterious mutations linked to advantageous ones [2],[3]. However, approximately 20% of flowering plants are self-fertilizing and engage in sexual reproduction without obtaining either of these benefits [4]. It has been proposed that inbreeding plant lineages represent evolutionary “dead ends” [5] that evolved from outbreeding ancestors [4]–[6]. In this view, mating system switches from an outbreeding to inbreeding mode may have been selected for by pollinator scarcity or population bottlenecks [7], with inbreeding providing the benefits of reproductive assurance and increased potential for colonization, and in some cases possibly representing a survival mechanism used as a last resort to perpetuate a species. Because the outbreeding mode of mating is typically associated with the accumulation of recessive deleterious alleles that cause inbreeding depression, self-fertile taxa can only become established if this genetic load is purged. Theoretical models of the evolution of selfing have shown that inbreeding depression can indeed be overcome and selfing alleles can spread when the advantage of reproductive assurance outweighs the reduction of fitness [8]. However, mechanistic studies of switches from outbreeding to self-fertility have rarely been performed, and the genetic basis of these switches is poorly understood.
In the crucifer (Brassicaceae) family, switches to inbreeding have occurred frequently and entailed loss of self-incompatibility (SI). Self-incompatibility is a barrier to self-fertilization that is determined by variants of a single highly polymorphic locus, called the “S locus”. In self-incompatible plants, pollen is prevented from hydrating, germinating, and producing pollen tubes at the stigma surface if the same “S-locus” variant is expressed in pollen and stigma, whether these structures are located within the same flower or derived from different flowers on the same plant or different plants (for recent review, see [9]). As a result, self-incompatible plants are largely but not completely self-sterile, and autonomous seed set is typically less than 5% that set by self-compatible plants. In all self-incompatible crucifer species investigated to date, the “S locus” is not a single gene, but rather consists of two polymorphic genes, allelic forms of which together constitute a unique S-locus haplotype (hereafter S haplotype) that defines a unique recognition specificity. One gene encodes the S-locus Receptor Kinase (SRK) [10] and the second gene encodes the small S-locus Cysteine-Rich protein (SCR), which is the ligand for SRK. SRK is expressed in stigma epidermal cells, and its product is anchored via a single transmembrane domain in the plasma membrane of these cells. SCR is expressed in the anther tapetum, a cell layer that lines the sacs in which pollen grains develop, from which its SCR product is secreted and becomes incorporated into the outer pollen coat [11]. SCR proteins are delivered to the stigma surface upon pollen-stigma contact, but an SCR will bind to the extracellular domain of SRK and activate its cytoplasmic kinase domain, thus triggering the SI response, only if the SRK and the SCR proteins are encoded by the same S-locus haplotype [12],[13], i.e. when stigmas are pollinated with pollen derived from the same plant or from plants expressing the same S haplotype.
In view of this S haplotype-specific interaction, recombination events that disrupt the genetic linkage of matched SRK and SCR alleles will cause loss of SI. Consequently, there is strong selection for maintaining the tight linkage of these genes. Recombinants between SRK and SCR are rare in self-incompatible plants, either because self-compatible genotypes that might arise do not persist in nature (due to their genetic load) or because recombination is actively suppressed in the S-locus region [14]–[17]. Similar to other genomic regions exhibiting low effective recombination rates [18]–[20], the S haplotypes of self-incompatible Brassica and A. lyrata strains have been shown to accumulate haplotype-specific sequences due to divergent evolutionary trajectories and independent degeneration of non-coding sequences, and these features no doubt limit recombination in the region [14], [17], [21]–[23].
The model dicot plant Arabidopsis thaliana is a highly self-fertile crucifer that is thought to have had a self-incompatible ancestor based upon phylogenetic inference [24] and rescue of the SI trait by transgenic complementation with a functional SRK-SCR allelic pair from its close self-incompatible relative A. lyrata [25],[26]. However, despite several recent studies and much debate [27]–[31], the nature and number of mutational events that caused the switch to self-fertility in the A. thaliana lineage have not been established. Consistent with the expectation that selective pressures for maintaining the integrity of the S locus and its genes would be relaxed subsequent to the switch to self-fertility, all A. thaliana accessions analyzed to date harbor a non-functional S locus, referred to as pseudo-S (ΨS), which carries inactivating mutations in the SRK and/or SCR genes [23]. Analysis of SRK and SCR sequence divergence in various accessions identified three distinct ΨS haplotypes, designated ΨSA, ΨSB, and ΨSC [23],[28],[29],[32]. These three A. thaliana ΨS haplotypes are inferred to be orthologous, respectively, to the S37, S16, and S36 haplotypes of A. lyrata. This conclusion is based on the observation that SRK or SCR sequences in the A. lyrata S37, S16, and S36 haplotypes share much higher sequence similarity with the ΨSRK or ΨSCR sequences of the A. thaliana ΨSA, ΨSB, and ΨSC haplotypes, respectively, than with other A. lyrata S haplotypes [30].
Despite clear evidence for inactivating mutations in the SRK or SCR sequences of many A. thaliana accessions [23],[27],[28], it is not possible to conclude that inactivation of the S locus was the primary cause of the switch to self-fertility in all A. thaliana populations. Indeed, the species also harbors mutations at other genes required for SI, as indicated by differences among accessions in the ability to express SI upon transformation with A. lyrata SRKb-SCRb (AlSRKb-SCRb) genes [26],[33]. Among seven accessions analyzed by inter-specific complementation experiments, only C24 yielded a developmentally-stable SI response identical to that of A. lyrata Sb plants (<5 pollen tubes/self-pollinated stigma at all stages of stigma development), demonstrating unequivocally that a non-functional S locus is the only cause of self-fertility in this accession [26],[27]. By contrast, in other accessions, SI was transient [starting strong (<5 pollen tubes/self-pollinated stigma) in young flower buds, and later breaking down (>100 pollen tubes/self-pollinated stigma) in older flower buds and flowers], weak (25–50 pollen tubes per self-pollinated stigma), or absent (large numbers of pollen tubes/self-pollinated stigma at all stages of stigma development, similar to wild type untransformed A. thaliana). These phenotypes indicate the presence of mutations not only at the S locus, but also at “SI modifier” loci required for SI [26],[33]. Indeed, one such SI modifier was identified in a cross between a C24::AlSRKb-SCRb transformant, which expresses a robust and developmentally-stable SI response, and a plant from the ΨSA-containing RLD accession, which expresses transient SI [33]. Molecular genetic analysis of this cross determined that transient SI is associated with reduced SRK transcript levels in older flowers caused by sequences upstream of the Col-0 allele of PUB8 (Plant U-Box 8), a gene tightly-linked to the S locus [33].
A comprehensive understanding of the switch to self-fertility in A. thaliana requires analysis of the S locus and of SI modifier loci, because any of these loci might have been targets of selection for self-fertility. Accordingly, we used a two-pronged approach to elucidate the genetic events that accompanied the evolution of self-fertility in A. thaliana. Firstly, we transformed several A. thaliana accessions with the AlSRKb-SCRb genes in an attempt to identify accessions like C24, which express a robust and developmentally-stable SI response, and would therefore harbor mutations at the S locus but not at SI modifier loci. We reasoned that only in such accessions might it be possible to determine if the transition from outbreeding to inbreeding in A. thaliana occurred by a single mutational event or by multiple independent events. Secondly, we performed a Quantitative Trait Loci (QTL) analysis of SI modifier loci that differentiate AlSRKb-SCRb transformants of the reference Columbia (Col-0) accession, which express transient SI, from those of the C24 accession.
To identify additional A. thaliana accessions, which, like C24, might express a robust and developmentally-stable SI phenotype, we transformed several previously-untested accessions with AlSRKb-SCRb. In selecting accessions for transformation, we excluded accessions that carry the ΨSA haplotype [27] and its closely-linked PUB8 allele previously associated with transient SI [33], because AlSRKb-SCRb transformants of these accessions are not expected to express stable SI. For each selected accession, independent AlSRKb-SCRb transformants were generated and tested for SI by pollination assays at different stages of stigma development (Table 1). AlSRKb-SCRb transformants of four accessions, Sha, Kas-2, Hodja, and Cvi-0, were found to express a developmentally-stable SI phenotype identical to that observed in C24::AlSRKb-SCRb transformants and in A. lyrata Sb plants [26]: immature floral buds were self-compatible, and strong inhibition of self pollen was first detected in stage-13 buds and persisted in older flowers. In addition, there was very little seed set on these plants, either by open pollination (Table 1) or following manual self-pollination of mature floral buds and flowers. Significantly, these self-incompatible phenotypes are stably transmitted to subsequent transgenic generations, as determined by analysis of pollination phenotype over 20 generations in C24, 10 generations in Sha, and two generations in each of Cvi-0, Kas-2, and Hodja.
Our successful complementation of the Sha, Kas-2, Hodja, and Cvi-0 accessions suggests that self-fertility in these accessions is due to a non-functional S locus, as in the C24 accession. It is therefore of interest to determine if the ΨS-haplotypes in these five accessions are the same or different (i.e. are likely to be derived from the same ancestral mutant ΨS-haplotype or from independently-derived ancestral ΨS-haplotypes).
At present, detailed descriptions are available only for the Col-0, C24, and Cvi-0 ΨS haplotypes. The Col-0 reference accession was shown to harbor a ΨSA haplotype containing aberrant SRK and SCR sequences. Its ΨSRKA allele contains a frameshift mutation that introduces a premature stop codon within the fourth of seven exons found in SRK genes. Its SCR sequences consist of several truncated ΨSCR sequences, the longest of which is designated ΨSCR1 [23]. In contrast, the C24 ΨS haplotype was shown to have been produced by recombination between ΨSA and ΨSC haplotypes [27]: it contains rearranged remnants of ΨSRKA exon 1 [which encodes the SRK extracellular domain (ΨeSRK)], a truncated version of ΨSRKC consisting of exon 7, and two copies of ARK3 (At4g21380), a polymorphic gene located at one flank of the S locus in Arabidopsis species [23]: one copy consists of an ARK3SC allele characteristic of ΨSC haplotypes located at its normal location and an additional chimeric ARK3 copy located between the ΨSRKA and ΨSRKC sequences, which resulted from recombination between an ARK3SC allele and an ARK3SA allele characteristic of ΨSA haplotypes. As for the Cvi-0ΨSB haplotype, its complete DNA sequence (accession number EF637083 [29],[32]) revealed the presence of a ΨSRKB allele containing a splice-site mutation at the end of intron 2 [28] and a convergently-oriented ΨSCRB allele lacking obvious inactivating mutations [29],[32].
Information on the molecular events associated with the transition from out-crossing to selfing in A. thaliana may also be gleaned by genetic analyses of crosses between accessions that differ in expression of SI. In previous studies, genetic analysis of a relatively small C24::AlSRKb-SCRb x Col-0 F2 population [26] had inferred the segregation of two loci affecting pollination phenotype and identified a major modifier causing breakdown of SI in close linkage to the Col-0 ΨS locus [33].
In this study, we raised a larger F2 population of 300 plants derived by selfing an F1 plant, and we performed a cursory analysis to confirm the hypothesis that two loci with dominance of SI-conferring alleles segregated in this cross. Individual plants were classified into four phenotypic groups based on autonomous seed set: plants producing empty fruits with only an occasional fruit containing seed, similar to the C24::AlSRKb-SCRb parent (1 in 80 fruits measured); plants with a full seed set similar to wild-type untransformed plants; plants producing few fruits with seed (1–3 for every 10 fruits measured); and plants producing many fruit with seed (4–8 for every 10 fruits counted). Subsequent manual self-pollination of these plants determined that the number of pollen tubes formed at the stigma surface was consistent with fruit set. Plants in the empty-fruit group exhibited the SI response in all self-pollination assays, while plants with full fruit set exhibited a self-compatible pollination phenotype similar to untransformed plants. Plants that produced few or many fruits with seed exhibited variable pollination phenotypes, in which breakdown of SI occurred in an apparently random fashion in individual flowers, and these plants are classified as being partially self-compatible. In addition, loss of SI was stigma specific as determined by reciprocal pollinations of self-compatible plants with the C24::AlSRKb-SCRb parent. The results of a chi-squared test based on the proportions of the phenotypic categories were consistent with segregation of two loci with dominance of SI-conferring alleles (X2 = 3.38; p = 0.3). A scan of the genome with molecular markers distributed on all five chromosomes confirmed the presence of a Col-0-derived modifier locus with strong effect located on chromosome 4 near the ΨS locus, which corresponds to the previously-identified S-locus-linked modifier on chromosome 4 [33]. It also determined that a second Col-0-derived modifier locus responsible for partial self-compatibility was located on the bottom of chromosome 3.
The strong-effect S–locus-linked modifier [33] can mask the effects of weak-effect modifiers. Therefore, to ensure detection of weak-effect loci, a QTL mapping population was generated that subtracted the genetic effects of this major modifier (see Methods). This population segregated for self-fertility, as expected. Manual self-pollination of a developing series of stigmas from two representative self-compatible plants revealed weakening of SI and some pollen tube growth in the most mature flowers. In contrast, self-pollination of a developing series of stigmas from two representative self-incompatible plants detected no pollen tubes in mature stigmas. Furthermore, reciprocal pollinations of self-compatible plants with C24::AlSRKb-SCRb transformants confirmed that the modifier alleles segregating in this population have stigma-specific effects as in the original C24::AlSRKb-SCRb x Col-0 cross. However, the self-compatible trait exhibited low penetrance in this population. On any given self-compatible plant, some flowers would not develop fruits with seeds, due to the SI response, while other flowers would develop into fruits filled with seeds. There was also great variability as to where on the stem SI would break down, the number of flowers that exhibited breakdown of SI, and the strength of the breakdown for each individual flower.
In view of this variability, manual self-pollinations of a small number of randomly-selected individual flowers, as is usually done in pollination assays, cannot reflect overall plant phenotype. Consequently, standard pollination assays are not useful for phenotypic classification of plants in the QTL mapping population. Therefore, we used the size of mature fruit produced by autonomous self-pollination as a measure of the extent of SI breakdown in individual flowers. We reasoned that fruit size was a valid proxy for pollination phenotype because of the known strong correlations between fruit size and number of seeds per fruit (as described previously [34] and confirmed in our study), and between number of seed in a fruit and strength of SI (as observed in our F2 population).
QTL analysis was performed using a total of 186 individuals (see Methods). For phenotypic classification, it was important to distinguish between empty fruits and fruits with few seeds. Based on dissection of 25 of the smallest fruits in this population, it was determined that a mature fruit containing at least one seed had a width of at least 0.6 mm. Therefore, fruits that were narrower than 0.6 mm were classified as being empty and indicative of a self-incompatible response, while fruits that had a width of 0.6 mm or greater were classified as containing seed and indicative of a breakdown of SI. Similar measurements of mature fruit produced by self-incompatible plants in the QTL mapping population gave an average fruit length of 0.42 cm±0.05 (n = 912, with only one fruit in 25 having a width of 0.6 mm), a value very similar to that of the C24::AlSRKb-SCRb parental strain, in which average mature fruit length was 0.48 cm±0.07 (n = 80, with only one fruit having a width of 0.6 mm). By comparison, the average length of seed-filled mature fruit in the self-compatible parent of the QTL population was 1.33 cm±0.4 (n = 59), while average fruit lengths in untransformed plants of the C24 and Col-0 accessions were 1.54 cm±0.19 (n = 80) and 1.38 cm±0.07 (n = 80), respectively.
As shown in Figure 4, the trait value distribution for the mapping population was continuous and approximately normal, suggesting the involvement of several genes in the control of fruit length. Individual plants were genotyped using 24 markers, microsatellites, and single nucleotide polymorphisms in chromosomal regions that segregated for Col-0-derived sequences. As shown in Figure 5 and Table 3, four QTL underlying the observed differences in fruit length were found: two QTL (QTL3.1 and QTL3.2) on chromosome 3, one QTL (QTL5) on chromosome 5, and one QTL (QTL1) on chromosome 1, which accounted respectively for 25%, 24%, 15%, and 16%, of the observed variation in fruit length. All of the QTL regions were well above the significance threshold, and none corresponded to “minor QTL” with peaks near the threshold line.
Nearly isogenic lines (NIL) were generated for each QTL region. Among these, only one NIL exhibited a breakdown of SI, as determined by manual self-pollination of flowers over the course of stigma development and by observation of autonomous seed set (Table 3). This NIL, NIL3.2, incorporates QTL3.2 and likely corresponds to the chromosome-3 modifier that was associated with partial self-compatibility in the original C24::AlSRKb-SCRb x Col-0 F2 population. Epistasis between QTL1, QTL3.1, and QTL5 was assessed by crossing the corresponding NILs to generate “double NILs”. However, none of the “double NILs” showed a breakdown of SI based on observations of seed set. A possible explanation for this result is that these QTL do contribute to breakdown of SI, but their effect may only be detected when all three are combined with QTL3.2. Another possibility is that QTL1, QTL3.1, and QTL5 control SI-independent variation in fruit length. Although the SI response exerts the major influence on seed number and consequently on fruit size in the populations we analyzed, modifier loci affecting differences in fruit size and seed number per fruit between wild-type Col-0 and C24 may also be segregating, similar to the loci uncovered in a previous analysis of natural variation for various fruit parameters [34]. Interestingly, this earlier study of fruit length differences between the Cvi and Landsberg accessions had identified a QTL in the QTL3.1 region [34], but not in the other QTL regions identified in this study.
In an attempt to fine-map QTL3.2, an F2 mapping population was generated by crossing an NIL3.2 plant with a wild type (untransformed) C24 plant. This population segregated for the 1-megabase Col-0 introgression encompassing QTL3.2 (Table 3). F2 plants exhibiting recombination within the QTL3.2 region were identified by screening 2,016 individual plants, both phenotypically for seed set and genotypically with markers “NGA12” and “intron2” located just inside the introgressed region (Table S1). Three phenotypic groups were observed among recombinant plants: self-incompatible, self-compatible, and surprisingly, partially self-compatible. The occurrence of partially self-compatible plants in the recombinant pool was not expected because the gene underlying QTL3.2 was determined to be recessive in the Col-0 background. Also unexpectedly, these recombinants did not show a tight correlation between genotype and phenotype under the assumption of complete dominance of the SI-conferring C24 allele (Table S2). Nevertheless, F3 families were generated from self-compatible NIL3.2 F2 plants. Analysis of nine such NIL3.2 F3 families failed to identify self-compatible plants in six of those families, indicating that the self-compatibility phenotype can be completely erased from one generation to the next (Table S2). In view of this result, the genotype-to-phenotype correlations inferred for the self-incompatible class of NIL3.2 F2 plants become questionable. Nevertheless, with this caveat in mind and considering only the unambiguous self-compatible NIL3.2 F2 plants, QTL3.2 is tentatively mapped to a region of approximately 105,000 base pairs between genes At3g60440 and At3g60730 (Table S2 and Table S3).
Our results have extended our understanding of the genetic events at the S locus and at modifier loci that accompanied the switch to self-fertility in A. thaliana.
The identification of four accessions, in addition to C24, in which self-fertility may be clearly attributed to a non-functional S locus is significant for several reasons. From a practical point of view, the availability of several strains with diverged genetic backgrounds that do not contribute SI modifier alleles in crosses to laboratory-generated mutants will greatly facilitate the mapping of these mutants and the eventual cloning of genes required for SI. From an evolutionary perspective, the finding demonstrates that rather than being unique, the C24 accession is only one of potentially many accessions whose self-fertile phenotype may be fully reverted to SI by transformation with the AlSRKb-SCRb genes. Interestingly, these accessions are not confined to one geographical region: C24 is a southern-European accession originally isolated in Portugal, whereas Kas-2, Hodja, and Sha are all central Asian accessions from Kashmir (Kas-2) or Tajikistan (Hodja and Sha), and Cvi-0 is restricted to the Cape Verdi Islands. A genome-wide polymorphism study in which 876 loci spread across the genome were surveyed in 96 accessions [35] had indicated that all accessions isolated from Tajikistan are genetically very similar to one another (although Hodja was not included in the study), that Sha and Kas-2 are very closely related to each other, and that both are significantly diverged from C24 and Cvi-0, which in turn are also highly diverged from each other.
Our analysis of the C24, Cvi-0, Kas-2, Hodja, and Sha accessions has illuminated the genetic events that likely caused loss of SI in these accessions and potentially others with similar ΨS-loci, genome-wide polymorphisms, and provenance. Keeping in mind that the ΨSA, ΨSB, and ΨSC haplotypes were derived from distinct ancestral functional S haplotypes, the four haplotypic structures observed in C24, Cvi-0, Kas-2, and the Hodja/Sha group (Figure 2) are consistent with independent origins of these ΨS haplotypes. The Cvi-0 ΨSB haplotype, which lacks ΨSA and ΨSC sequences was clearly independently derived. The Sha and Hodja ΨS haplotypes are highly-decayed versions of the ancestral SA haplotype also found in Col-0, and it is possible that the S haplotypes in these three accessions might have been derived from the same ΨSA haplotype. In contrast, the C24 and Kas-2 ΨS haplotypes are both recombinant haplotypes generated by illegitimate recombination between ancestral SA and SC haplotypes. It is possible that the C24 ΨS haplotype was derived from a Kas-2-like ΨS haplotype via a complex series of restructuring events. Alternatively, based on the extensive genome-wide divergence inferred for the C24 and Kas-2 accessions [35], their recombinant ΨS haplotypes might have arisen independently, as illustrated in Figure 6.
Our data thus demonstrate that the ability to express a developmentally-stable transgenic SI response is not restricted to one group of highly-related accessions or to accessions harboring one ΨS haplotype. Additionally, the divergence of ΨS haplotypes harbored by these accessions provides further evidence for the lack of a single selective sweep at the A. thaliana ΨS locus [27],[29]. Rather, the results support the hypothesis that the switch to self-fertility in this species occurred by recurrent selection of distinct S-locus loss-of-function mutations. Such a process involving selection of adaptive mutations of independent origins has been referred to as a “soft sweep” [36]. Notably, soft sweeps are not restricted to the switch to self-fertility described here, and evidence of their occurrence is suggested by studies of polymorphisms in a variety of systems and organisms ranging from protozoa to human [36]. For example, in three-spine stickleback fish, selection for reduced body-plate armor in isolated European and Japanese populations has apparently resulted in the fixation of different alleles of ectodysplasin, a factor required for epithelial cell morphogenesis [37],[38].
Possible scenarios for the generation of the observed ΨS haplotypes are shown in Figures 7 and 8. It should be noted however, that the exact nature of the inactivating mutation and sequence of events that produced these ΨS haplotypes cannot be inferred from our data. A major difficulty in charting the history of the A. thaliana S locus is distinguishing a primary inactivating mutation from subsequent decay of the non-functional haplotype by further mutation, sequence loss, and rearrangement. For example, it is impossible to know whether the recombination events that produced the C24 and Kas-2 S haplotypes caused S-locus inactivation by disrupting the physical linkage between functional allelic SRK-SCR pairs, or if they occurred between already-mutated SA and/or SC haplotypes. There is also uncertainty as to whether the Kas-2 primary mutation is the same as that of Hodja and Sha. Although all three accessions have closely-related genomes and originate from close geographical locations, their ΨS loci differ in allele content and extent of decay. Furthermore, in contrast to the ΨSA haplotypes and the ΨSB haplotype of Cvi-0, for which both ΨSRK and ΨSCR sequences as well as their A. lyrata orthologues are known, only an incomplete picture of ΨSC haplotypes is available because neither A. thaliana ΨSCRC sequences nor the orthologous A. lyrata SCR36 (AlSCR36) sequences have as yet been isolated. Identification of AlSCR36 is likely to be particularly informative. Just as AlSCR37 sequences allowed a resolution of the Col-0 ΨSCR1 structure in this study, AlSCR36 sequences may be used to investigate the fate of the SCRC allele in A. thaliana and to determine if, and in what form, these sequences were maintained in Kas-2, C24, or other ΨSRKC-carrying accessions.
The structures of the ΨS haplotypes observed for Kas-2 and C24 as well as Nok-3 (Table 2) reveal an important role for recombination in shaping extant S-locus structure in A. thaliana. The ΨSA-ΨSC recombinant haplotypes of these accessions provide clear evidence for the occurrence of inter-haplotype recombination events in geographical areas where the SA and SC haplotypes were both present [27], as in southwestern Europe for the C24 ΨS haplotype and in central Asia for the Kas-2 ΨS haplotype (Figure 6). Only the ΨSB haplotype, which is restricted to the Cape Verdi Islands, did not participate in inter-haplotype recombination (Figure 6). Thus, recombination between S haplotypes that encode different SI specificities can occur, despite the extensive structural heteromorphism and sequence divergence that typically distinguish these S haplotypes. It is possible that DNA crossover might occur in small regions of sequence similarity, such as regions containing the many transposon-like sequences present within the locus [27].
The contrast between the occurrence of inter-haplotype recombination events inferred in this study and the very low effective rate of recombination that typically characterizes the S-locus region in self-incompatible species [15],[17] suggests that purifying selection against recombinants actively maintains low rates of recombination in the region, as previously discussed [17]. The switch to self-fertility is expected to have caused relaxation of this selective pressure, leading to further restructuring of the S-locus region. Thus, it is interesting to consider whether current recombination rates at the ΨS locus of A. thaliana are consistent with this expectation. The potential for recombination certainly exists despite high levels of self-fertility, as gene flow via pollen dissemination has been shown to contribute to genetic variability in local populations of the species [39]. Furthermore, the S-locus region was identified as a recombination hotspot in a cross between the Col-0 and Ler-0 accessions [40]. However, these two accessions harbor highly similar if not identical ΨSA haplotypes [27], and much lower recombination rates are expected in crosses involving structurally-divergent ΨS haplotypes. This expectation was confirmed by a recent analysis of 3,210 plants derived from a cross between C24 and RLD, an accession that carries the same ΨSA haplotype as Col-0 (Figure 2). Using the S-locus flanking markers PUB8 (At4g21350) and ARK3 (At4g21380), which are separated by 34 kilobases in RLD, only 1 recombinant was recovered, and this recombinant was produced by a cross-over event within the promoter region of PUB8, not within the S locus proper [33]. Thus, the likelihood of further S-locus restructuring by recombination between structurally-diverged ΨS haplotypes is low, despite relaxed selection on the locus.
The acquisition of a robust and developmentally-stable SI response by accessions that harbor independently-derived ΨS haplotypes provides the strongest evidence to date that A. thaliana evolved from an obligate out-crosser to a predominantly selfing species through multiple S-locus inactivating mutations in distinct outbreeding individuals. One interpretation of our data is that self-fertility in A. thaliana arose at least twice: once in an SA or SC haplotype (producing the Hodja/Sha, C24, and Kas-2 ΨS haplotypes) and once in an SB haplotype (producing the Cvi-0 ΨSB haplotype). A less conservative interpretation would invoke three origins of self-fertility if the C24 and Kas-2 S haplotypes are assumed to have arisen independently (Figure 6).
When and how frequently mutations at SI modifier loci occurred in A. thaliana must await the molecular cloning of these loci. At least one such SI modifier was uncovered in our QTL analysis of differences in expression of SI between AlSRKb-SCRb transformants of the C24 and Col-0 accessions. This previously-unidentified recessive modifier, defined by QTL3.2, was associated with self-fertility in Col-0 and was mapped to chromosome 3. However, phenotypic instability, low heritability, and erasure of the self-compatibility trait in advanced mapping populations precluded further fine mapping and isolation of the underlying gene(s). The cause of this instability is not known. One intriguing possibility is that it might reflect an epigenetic component in control of the self-compatibility trait in these populations. Indeed, phenotypic instability is a hallmark of epigenetically-controlled traits in various organisms [41]–[44]. Furthermore, examples of naturally-occurring epialleles have been reported in plants [43],[45], and widespread epigenetic natural variation has been noted among accessions of A. thaliana [46]–[48]. Similar to other epialleles that display unpredictable patterns of instability, the instability of QTL3.2 might be due to the loss of an unlinked trans-acting “maintainer” locus through segregation in NIL populations.
In any case, our identification of an unstable modifier of SI has relevance for theoretical modeling and mechanistic studies of switches to self-fertility in A. thaliana and other plant species. Clearly, approaches more suited to the identification of unstable alleles than traditional QTL analysis and association mapping [49] will be required to clone at least some of the genes associated with self-fertility. Future molecular genetic analysis of polymorphisms at SI modifier loci, as well as investigation of S-locus structure in additional accessions that might express developmentally-stable SI upon transformation with the AlSRKb-SCRb genes, will undoubtedly determine if switches to self-fertility occurred exclusively by inactivation of the S locus in the A. thaliana lineage.
A. thaliana plants were typically grown at 22°C and a photoperiod of 16 hours. Plants that were used for transformation by the floral dip method [50] were grown under a 24-hour photoperiod. All accessions used in this study were obtained from the Arabidopsis Biological Resource Center in Columbus, Ohio. The Kashmir (Kas-2; CS22638), Shahkdara (Sha; CS929), and Hodja-Obi-Garm (Hodja; CS6178) accessions were transformed with the p548 plasmid (here designated AlSRKb-SCRb), a previously-described pBIN-PLUS derivative containing the A. lyrata SRKb and SCRb genes [26]. DNA gel blot analysis was used to confirm the independent origin of transformants and to identify transformed lines carrying single integrations of the transgene pair: genomic DNA was isolated from individual plants by the CTAB method [51], digested with EcoR1, transferred to Hybond H+ membrane (Amersham Biosciences, Piscataway, NJ), and hybridized according to the Hybond H+ membrane instruction manual with a probe specific for the Neomycin PhosphoTransferase (NPTII) gene that was labeled with 32P using the Random Priming kit (Roche, Indianapolis, IN). Hybridized membranes were washed at 65°C first in a solution containing 2× SSC and 0.5% SDS and subsequently in a solution containing 0.2× SSC and 0.5% SDS. Blots were exposed to phosphor screens, scanned using a GE Healthcare STORM phosphorimager (Piscataway, NJ), and analyzed with the ImageQuant software package purchased as a bundle with the phosphorimager. In all cases analyzed, each transformant was found to exhibit a unique transgene pattern (data not shown), consistent with independent transgene integration events and demonstrating that each of the analyzed transformants was or independent origin.
Pollination responses were tested on pollen-free stigmas just before anthesis, when the stigmas are receptive to pollen but before the pollen grains are mature and released from the anthers. Using a stereomicroscope, stigmas were manually pollinated with hundreds of pollen grains from the dehisced anthers of mature flowers. Two hours after pollination, flowers were fixed for 10 minutes in a 3∶1 mixture of ethanol and acetic acid at 65°C, softened for 10 minutes in 1N NaOH at 65°C, washed two times in water, stained in decolorized aniline blue, and transferred to a slide for observation by epifluorescence microscopy [52]. Under these conditions, a pollination is scored as strongly incompatible if no or fewer than 5 pollen tubes are observed per pollinated stigma, as fully compatible when more than 50 pollen tubes are observed per pollinated stigma, and as partially self-compatible (or weakly self-incompatible) when intermediate numbers of pollen tubes are observed.
Genomic DNA gel blot analysis with probes derived from different ΨSRKs was used to assess the composition of the S locus in various accessions of A. thaliana. This method is more suitable than amplification by the polymerase chain reaction (PCR) for our study because of the known or expected sequence divergence of the loci under study. Indeed, previous applications of this method to analysis of S-locus polymorphisms in A. thaliana have demonstrated that it can identify homologous sequences that are missed by PCR (27). Under low-stringency hybridization and washing conditions, DNA gel blot analysis can detect sequences that share as little as 50% overall similarity with the probe but not small stretches of sequence similarity or sequences that have decayed to below the 50% sequence similarity threshold. The probes for this analysis were fragments corresponding to the first exon and the seventh or last exon of A. thaliana ΨSRKA (At4g21370) from Columbia (Col-0; CS1092), to the first intron of ΨSRKB and ΨSCRB from the Cape Verdi Islands accession (Cvi-0; CS1096), and to the first intron of ΨSRKC from the Ibel Tazekka accession (Ita-0; CS1244). Fragments were amplified from genomic DNA using specific primers (Table S1), labeled with 32P, and used in sequential hybridizations of EcoRI-digested genomic DNA isolated from various accessions, as described above. An insertion/deletion polymorphism in ARK3 [27], a gene tightly linked to the S locus in Arabidopsis species, was also assessed by PCR using specific primers (Table S1) to distinguish between the ARK3SC allele (characteristic of ΨSC haplotypes), which has the deletion, and the ARK3SA allele (characteristic of ΨSA haplotypes), which lacks the deletion. Accessions used in this analysis included Kashmir (Kas-2; CS1264), Shahkdara (Sha; CS929), Hodja-Obi-Garm (Hodja; CS6178), C24 (CS906), Col-0, Lezoux (Lz-0; CS22615), Noordwijk (Nok-3; CS22643), Randan (Ra-0; CS22632), Ita-0, Monte (Mr-0; CS22640), and Cape Verdi Islands (Cvi-0; CS902 and CS1096). Standard PCR reagents were used with 35 cycles of the following: 94°C for 30 seconds, 55°C for 30 seconds, and 72°C for one minute or longer.
The accessions were also assayed for previously-unidentified AtΨSCR1 exon 2 sequences, which were isolated in this study as follows. A recently-reported partial sequence of the A. lyrata SCR37 (AlSCR37) gene, the ortholog of A. thaliana ΨSCR1 in Col-0 [30], was used as anchor to clone the remainder of AlSCR37 using the “DNA Walking SpeedUp Premix Kit II” (Seegene, Rockville, MD) and gene-specific primers (Table S1). Amplification of AlSCR37 genomic DNA (kindly provided by Dr. Jesper Bechsgaard) was performed according to the manufacturer's directions and amplified products were cloned into pGemT-easy (Promega, Madison, WI). Inserts were sequenced at the Cornell University Life Sciences Core Laboratories Center (Ithaca, NY) using SP6 and T7 universal primers. A BLAST search of the A. thaliana Col-0 genome using the newly-identified A. lyrata SCR37 second exon located the corresponding portion of A. thaliana ΨSCR1, and primers were designed (Table S1) to screen for the presence of an intact AtΨSCR1 second exon in 96 accessions of A. thaliana [35] using A. lyrata S37 DNA as positive control.
SI prevents self pollen from reaching and fertilizing the ovule, and thus precludes fruit expansion. A breakdown or absence of SI allows self pollen to fertilize the ovules, resulting in fruit expansion and elongation. Consequently, for QTL analysis, fruit size was used as a proxy for self-pollination phenotype. Data used to calculate the phenotype values for individual plants were collected by sampling three inflorescence stems, scanning them using a flat-bed scanner, and measuring the length and width of each fruit using ImageJ software (http://rsb.info.nih.gov/ij/). An average of 80 fruits were scanned and measured for each plant, and on average across the population, one-fourth of those fruits contained seeds and were used in the average length calculation. Each of these fruits was the result of autonomous self-pollination, because they were grown in the absence of pollinators. A flower was deemed self-compatible, if the fruit width was greater than 0.6 mm, i.e. the minimal width of one fully-developed seed. Because of variability in fruit development, the trait values reported here were calculated for each plant as the average length of fruits with at least one seed.
The QTL mapping population was generated using a self-fertile F4 plant derived from the C24::AlSRKb-SCRb x Col-0 cross, which was homozygous for the PUB8C24 allele and for the Col-0 allele at the chromosome-3 modifier. The F2 parent of the selected F4 plant displayed a transient SI phenotype as determined by seed set and pollination assays (<5 pollen tubes/stigma in young buds and >50 pollen tubes/stigma in older buds and flowers). The F4 plant also produced abundant seed, although some flowers remained self-incompatible throughout development and did not produce seeds. It was homozygous over most of its genome, with Col-0-derived DNA occurring in large stretches on chromosomes 1, 3, and 5, and in a small region on chromosome 4. This plant was back-crossed to C24, producing F4BC progenies that were self-incompatible, similar to the original C24::AlSRKb-SCRb x Col-0 F1 hybrid. The F4BC was subjected to forced selfing in immature floral buds (i.e. before stigmas acquire the ability to reject self pollen) to generate an F4BCF2 population for QTL analysis, which we refer to as the QTL mapping population.
Since the C24 accession was not completely sequenced when this study was undertaken, a search for markers that showed co-dominant polymorphisms between C24 and Col-0 was done by PCR screening of publicly available microsattelite markers designed for other pairs of accessions and of random amplification of repetitive elements found in the Col-0 genome (www.arabidopsis.org). In addition, a limited number of dominant SNP markers were designed to detect differences as small as one base pair between the two parents. Twenty-four marker loci (Table S1) were found to be polymorphic between the two accessions and were scored on 186 individuals in the QTL mapping population. Markers were amplified using forward primers with M13 adapters to enable large scale genotyping [53]. A linkage map and mapping files containing genotype and phenotype data were produced using MapManager for analysis in MapManager and also exported into WinQTL Cartographer (http://statgen.ncsu.edu/qtlcart/). All recombination distances, measured in centiMorgans (cM), were co-linear with physical distances (data not shown). QTL interval mapping and composite interval mapping methods were applied to the genotype and marker data using both software programs. The various analyses and programs all produced similar results. A 0.05 significance threshold of LOD 2.8 was determined in WinQTL (http://statgen.ncsu.edu/qtlcart/) by creating a random distribution of the data through 1000 permutations.
|
10.1371/journal.pntd.0000175 | Increased Risk for Entamoeba histolytica Infection and Invasive Amebiasis in HIV Seropositive Men Who Have Sex with Men in Taiwan | Incidence of Entamoeba histolytica infection and clinical manifestations and treatment response of invasive amebiasis (IA) in HIV-infected patients have rarely been investigated before.
At the National Taiwan University Hospital, medical records of HIV-infected patients who received a diagnosis of IA between 1994 and 2005 were reviewed. The incidence of amebiasis was investigated in serial blood and stool samples from 670 and 264 HIV-infected patients, respectively, using serological and specific amebic antigen assays. DNA extracted from stool samples containing E. histolytica were analyzed by PCR, sequenced, and compared. Sixty-four (5.8%) of 1,109 HIV-infected patients had 67 episodes of IA, and 89.1% of them were men having sex with men (MSM). The CD4 count at diagnosis of IA was significantly higher than that of the whole cohort (215 cells/µL vs. 96 cells/µL). Forty episodes (59.7%) were liver abscesses, 52 (77.6%) colitis, and 25 (37.3%) both liver abscesses and colitis. Fever resolved after 3.5 days of metronidazole therapy (range, 1–11 days). None of the patients died. The incidence of E. histolytica infection in MSM was higher than that in other risk groups assessed by serological assays (1.99 per 100 person-years [PY] vs. 0 per 100 PY; p<0.0001) and amebic antigen assays (3.16 per 100 PY vs. 0.68 per 100 PY; p = 0.12). In multiple logistic regression analysis, only MSM was significantly associated with acquisition of E. histolytica infection (adjusted odds ratio, 14.809; p = 0.01). Clustering of E. histolytica isolates by sequencing analyses from geographically-unrelated patients suggested person-to-person transmission.
HIV-infected MSM were at significantly higher risk of amebiasis than patients from other risk groups. Despite immunosuppression, amebic liver abscesses and colitis responded favorably to treatment.
| Entamoeba histolytica, morphologically identical to but genetically different from E. dispar and E. moshkovskii, is the causative agent of amebiasis. Recently there have been reports of increased risk for amebiasis among men who have sex with men (MSM) due to oral-anal sexual contact in several developed countries. In this longitudinal follow-up study, the incidence of amebiasis was determined among HIV-infected patients using serological and specific amebic antigen assays. DNA extracted from stool samples containing E. histolytica were analyzed by PCR, sequenced, and compared. Clinical manifestations and treatment response of invasive amebiasis in HIV-infected patients were reviewed. The results demonstrated that HIV-infected MSM were at significantly higher risk of amebiasis than patients from other risk groups. Clustering of E. histolytica isolates by sequencing analyses from geographically unrelated patients suggested person-to-person transmission. Despite immunosuppression, amebic liver abscesses and colitis responded favorably to metronidazole therapy. It is important to investigate in areas of high incidence of both amebiasis and HIV (sub-Saharan Africa) how generalizable these findings are.
| Invasive amebiasis (IA) is the second most common cause of mortality due to parasite infections worldwide, accounting for 40,000 to 100,000 deaths annually. High risk populations for developing IA include infants, pregnant women, and patients who are taking immunosuppressives [1],[2]. Interestingly, IA has not been considered to occur at a higher frequency in HIV-infected patients [3],[4]. In industrialized countries, the rare occurrence of IA in HIV-infected patients or persons at risk for HIV infection is probably attributed to the rare intestinal carriage of E. histolytica [4]–[9]. This is in contrast with the relatively frequent carriage of the non-pathogenic E. dispar among men who have sex with men (MSM) who attend sexually transmitted diseases clinics [10]–[13]. In a retrospective review of medical records of more than 34,000 HIV-infected patients in the US [9], 111 (0.3%) patients were diagnosed as having E. histoytica or E. dispar infection, and only 2 had extra-intestinal amebiasis. Amebiasis was significantly more prevalent among MSM and patients from E. histolytica endemic areas. However, the interpretation of the results of this study is limited by the retrospective study design and failure to differentiate between E. histolytica and E. dispar [14].
In developing countries, studies comparing the prevalence of amebiasis in HIV-infected and HIV-uninfected persons yielded inconsistent results [15]–[22]. The interpretation of these studies, however, is difficult because a majority of the diagnosis of amebiasis was based solely on microscopic examination of stool samples, which is an insensitive test that fails to distinguish E. histolytica from E. dispar [14]. In a cross-sectional study using stool antigen detection and polymerase chain reaction (PCR) from Mexico, where amebiasis is endemic, investigators found that HIV-infected patients appeared to have a higher rate, though not statistically significant, of E. histolytica infection than their sexual partners or close contacts [23]. However, those patients colonized with E. histolytica did not develop invasive diseases over the 12-month follow-up period.
Over the past few years, we and many investigators in Japan, Taiwan, and Korea have found that IA is increasingly diagnosed among HIV-infected MSM [24]–[31]. Of the estimated 500 to 600 reported cases of amebiasis annually in Japan, 80% of them occurred in MSM [32] and a substantial proportion of patients with IA were also co-infected with HIV and syphilis [24],[30]. In Taiwan, an estimated 5–6% of HIV-infected patients developed IA, and in many IA was the presenting disease of HIV infection [31]. Serologic surveys in the US, Italy, Japan, and Taiwan also demonstrated that MSM, regardless of HIV status, were at an increased risk of exposure to E. histolytica [13], [26], [31], [33]–[35]. Recent detection of locally acquired amebiasis among MSM who had no recent travel to endemic areas for E. histolytica has raised concerns in Sydney, Australia [36]. Oral-anal sexual contact has been found to be significantly associated with acquisition of E. histolytica infection [37]. Although IA has been considered an increasingly important parasitic infection in HIV-infected patients in three East Asian countries, the incidence of amebiasis and the clinical spectrum and the response of IA to standard metronidazole therapy have not been well studied.
In this study, we conducted a longitudinal follow-up study to assess the incidence of E. histolytica infection among persons with HIV infection at a referral medical center for HIV care in Taiwan. We also described the clinical spectrum and treatment outcome of IA.
Medical records of 1109 consecutive, non-hemophiliac HIV-infected patients aged 15 years or greater were reviewed to identify cases of IA at the National Taiwan University Hospital from June 1994 to December 2005 with the use of a standardized case record form. Of the 1109 patients, 781 (70.4%) were MSM. During the study period, a standardized protocol was followed to investigate HIV-infected patients who presented with gastrointestinal symptoms [28],[31]. Those investigations included at least two stool specimens for bacterial cultures and microscopy of concentrated wet mount preparations and modified acid-fast staining; indirect hemagglutination (IHA) assay to detect anti-E. histolytica antibodies (Cellognostics, Boehhringer Diagnostics GmbH, Marburg, Germany); endoscopy and biopsy for histopathologic examinations in patients whose stool examinations were non-diagnostic; abdominal sonography followed by computed tomography for patients with abnormal liver function tests, and space-occupying lesions of the liver. Specific Entamoeba antigen assays using commercial test kits (ENTAMOEBA TEST, TechLab, Branchburg, NJ) followed by polymerase chain reactions (PCR) using specific primers for E. histolytica was introduced after 1 January, 2001 [31].
Definite IA was diagnosed when erythrophagocytic trophozoites and/or positive PCR to E. histolytica were identified in clinical specimens from patients with symptoms compatible with IA, such as colitis and liver abscesses [28],[38]. Probable IA was confirmed when a patient with IA symptoms responded to metronidazole monotherapy and the aspirates or blood specimens showed high IHA titers, but microbiological cultures for bacteria, fungi, or histopathological examination of aspirates and biopsy specimens did not reveal any other pathogen. Results of the IHA assay were considered positive if the titer was 128 or greater. AIDS was defined according to the 1993 revised classification system for HIV infection and expanded surveillance case definition for AIDS among adolescents and adults [39]. Highly active anti-retroviral therapy (HAART) was defined as anti-retroviral therapy containing two nucleoside reverse transcriptase inhibitors and a (boosted) protease inhibitor(s) or a non-nucleoside reverse transcriptase inhibitor.
Serum samples from HIV-infected patients at baseline were tested for IHA and patients who remained in clinic follow-up from January 2001 to December 2005 were re-tested to determine the sero-incidence of E. histolytica infection. The interval between the two blood samples was at least one year. In patients with samples that tested positive at the last clinic follow-up or the end date of the study (31 December, 2006), serially stored serum samples were retrospectively tested to determine the seroconversion date. Seroconversion was defined as changes from sero-negative at baseline to IHA titers of 128 or greater at subsequent IHA assay; or increases of IHA titers by four-fold or greater. The seroconversion date was defined as the mid-point between the dates when the last sero-negative sample and the first sero-positive sample were collected.
Sequential stool samples from HIV-infected persons were tested for the presence of stool Entamoeba antigen between 1 January, 2001 and 31 December, 2005. Those patients who were negative for Entamoeba antigen were asked to provide stool samples for follow-up testing using the same method in order to assess the incidence of new acquisition of E. histolytica.The interval between the two stool samples were at least 6 months. The date of new infection was estimated as the mid-point between the date when the last antigen-negative sample and the first antigen-positive sample were collected.
Stool specimens tested positive for E. histolytica/E. dispar antigen were further confirmed by PCR. The primer sets for a multiplex nested PCR were based upon the variable regions between 16S-like rDNAs of E. histolytica (GenBank X56991) and E. dispar (GenBank Z49256) [31]. The procedures to isolate total DNA from the stool samples and the PCR conditions were described previously [31]. Individual E. histolytica isolates were genotyped by PCR amplification and sequencing of the previously described polymorphic loci, locus 1-2, using one set of primers (R1: CTGGTTAGTATCTTCGCCTGT and R2: CTTACACCCCCATTAACAAT) previously described [40],[41]. PCR was carried out in a 50 µl reaction mixture containing 0.1 µg of DNA, a 1.5 µM concentration of each primer, 2.5 mM MgCl2, a 100 µM concentration of each deoxynucleoside triphosphate, and 1.5 U of AmpliTaq Gold DNA Polymerase (Applied Biosystems) with the Taq activation at 95°C for 15 min and 30 cycles of denaturation at 94°C for 30 s, annealing at 45°C for 30 s, and extension at 72°C for 1 min, and then final extension at 72°C for 10 min [40]. The PCR products were fractionated by electrophoresis in 3% NuSieve 3∶1 agarose (Cambrex, East Rutherford, USA), stained by ethidium bromide and visualized under UV illumination. After purification using the QIAquick PCR purification kit (QIAGEN), locus 1-2 PCR products were sequenced twice in the forward and reverse directions. The sequences from representative genotypes chosen to infer the phylogenetic trees of locus 1-2 were were manually edited and aligned by using BioNumerics V. 4.01 software (Applied Maths, Kortrijk, Belgium). The study protocols were approved by the Institutional Review Board of NTUH and patients gave written informed consent.
All statistical analyses were performed using SAS statistical software (Version 8.1, SAS Institute Inc., Cary, NC, U.S.A.). Categorical variables were compared using χ2 or Fisher's exact test and non-categorical variables were compared using Wilcoxon's rank-sum test. The incidence rate of E. histolytica infection or seroconversion was calculated as number of episode per 100 person-years (PY) of observation. Exact 95% confidence intervals (95% CI) for incidence rates were calculated on the basis of the Poisson distribution. The follow-up duration was from the date with the first stool or blood sample that was negative for E. histolytcia antigen or IHA to the date with the sample that was positive, date of death, or on 31 December, 2006, whichever occurred first. Multiple logistic regression analysis was performed between patients who were diagnosed with newly acquired E. histolytica infection by serologies or antigen assays and those who remained uninfected in order to identify the risk factors associated with E. histolytica infection. All tests were two-tailed. A p value <0.05 was considered significant.
During the 11-year study period, 64 (5.8%) HIV-infected patients were diagnosed as having 67 cases of IA (Table 1). All of the 64 patients were males and 57 (89.1%) were MSM. MSM had a higher risk of invasive amebiasis compared with other risk groups: 57/781 vs. 7/328 (risk ratio, 3.42; 95% CI, 1.5777, 7.417). In 29 cases (43.8%), HIV infection was concurrently diagnosed with IA. The CD4 count at diagnosis of IA was significantly higher than that of the whole cohort (215 cells?L vs. 96 cells?L). Fever (72.6%), diarrhea (70.8%), right upper quadrant pain (32.3%), and dysentery (20.6%) were the most common symptoms of IA. Fifty-two (77.6%) of the 67 IA episodes were amebic colitis, 40 (59.7%) episodes were liver abscesses (including 4 multiple abscesses), and 25 (37.3%) were both amebic liver abscesses and colitis (Figures 1 and 2). By IHA assays, 51.6% of the patients with IA had titers≧512 (range, 0–16384). Eight (11.9%) developed serious complications necessitating surgical intervention, which included 3 intestinal perforations and peritonitis, 2 ruptures of the liver abscess, 2 subphrenic abscesses, 1 empyema, and 1 hepatogastric fistula. Metronidazole was administered for 13 days (range, 3–27 days) and the interval from initiation of metronidazole to defervescence was 3.5 days (range, 1–11 days). Thirty-five patients received concurrent antibiotic therapy, mainly ceftriaxone, and the fever resolved after 2 days of antibiotic therapy (range, 0–10 days). Of 21 patients receiving only metronidazole, the fever resolved after 3 days of therapy (range, 1–6 days). Liver aspiration and drainage was performed in 14 (20.9%) patients. Two required a laparotomy and chest tube drainage. Iodoquinol was administered to 42 (62.7%) patients following completion of metronidazole therapy to clear intestinal colonization and prevent relapse. Nobody died of IA after a median observation of 748 days (range, 9–4179 days).
Of 991 patients (89.4%) with available IHA assay results at baseline, 66 patients (6.7%) had IHA titers of 128 or greater. Between January 2001 and December 2005, 670 patients, including 433 (63.6%) MSM, who had more than one blood sample available for follow-up IHA assay, were enrolled in the sero-incidence study (Figure 3). There were no significant differences in demographics and clinical characteristics between the 670 patients who had follow-up IHA assays and the 321 who did not (data not shown). There were no significant differences in median CD4 count and plasma HIV RNA load (PVL) between MSM and patients from other risk groups when the first IHA assays were performed. At baseline, a significantly higher proportion of MSM (7.2%) had IHA titers of 128 or greater than patients from other risk groups (p = 0.006) (Table 2).
The median interval between the two blood samples was 1054 days (interquartile range [IQR], 606–1857 days) (Table 2). Twenty-one (3.1%) of the 670 HIV-infected patients seroconverted with IHA titers from 0 to 128 or greater; the median interval for seroconvertion was 1507 days (IQR, 790–2039 days). MSM were at statistically significantly higher risk for seroconversion for E. histolytica infection. The crude sero-incidence of E. histolytica infection among MSM was 4.9% compared with 0% among other risk groups (p<0.0001). The incidence rate of seroconversion was 1.44 per 100 PY (95% CI, 0.89, 2.20 per 100 PY) among MSM, compared with 0 per 100 PY (95% CI, 0, 0.38 per 100 PY) among patients of other risk groups (p<0.0001). When an increase of IHA titer by 4-fold or greater was included along with changes of IHA titers from 0 to 128 or greater, the incidence rate of seroconversion was 1.99 per 100 PY (95% CI, 1.33, 2.86 per 100 PY) among MSM, compared with 0 (95% CI, 0, 0.38 per 100 PY) among patients of other risk groups (rate ratio, 19.15; 95% CI, 2.61, 140.6) (p<0.0001).
Four hundred and sixty-nine (37.6%) patients, including 303 (64.6%) MSM, provided a total of 732 stool samples (range, 2–6 samples; median, 3 samples) for antigen testing between 2001 and 2005. At baseline, 45 (9.6%) patients, including 36 MSM (80%), had stool samples that tested positive for E. histolytica/E. dispar antigen. Two hundred and sixty-four patients, including 165 (62.5%) MSM, who had no intestinal infection with E. histolytica/E. dispar at baseline by specific stool antigen assays submitted more than one stool sample for repeat antigen assays (Table 3). At the first stool antigen assay, a significantly higher proportion of MSM (8.5%) had IHA titers of 128 or greater than patients from other risk groups (2.0%) (p = 0.03).
After an observation period of 399 PY, 19 (7.2%) patients were found to be infected with E. histolytica/E. dispar; 9 (47.4%) of the 19 isolates were E. histolytica by PCR. The median interval between the negative and positive antigen tests of the 19 patients was 287 days (IQR, 221–393 days). The crude incidence of new E. histolytica/E. dispar infection among MSM was 8.5%, compared with 5.1% among patients from other risk groups (p = 0.37); the incidence rate of new acquisition of E. histolytica/E. dispar infection was 5.56 per 100 PY (95% CI, 3.04, 9..32 per 100 PY) among MSM, compared with 3.40 per 100 PY (95% CI, 1.10, 7.94 per 100 PY) among patients from other risk groups (p = 0.36). Of the 9 E. histolytica isolates, 8 isolates were from 165 MSM, while 1 was from 99 patients from other risk groups (p = 0.16). The incidence rate of E. histolytica infection was 3.16 per 100 PY (95% CI, 1.37, 6.23 per 100 PY) among MSM compared with 0.68 per 100 PY (95% CI, 0.02, 3.79 per 100 PY) among patients from other risk groups with a rate ratio of 4.667 (95% CI, 0.5837, 37.31) (p = 0.12). All of the patients with E. histolytica infection were asymptomatic.
Six of 9 (66.7%) patients, who acquired new E. histolytica infection, seroconverted from sero-negative to sero-positive for anti-E. histolytica antibodies, compared with only 2 of 255 (0.78%) who did not acquire E. histolytica infection (odds ratio, 251 [95% CI, 35.22, 1789]) (p<0.0001).
The results from sequencing extracted DNA from the E. histolytica isolates in this study and other isolates of E. histolytica from our previous prevalence study [31] are shown in Figure 4. Case clustering among isolates from MSM was noted (Locus1-2 allelic genotypes B and D), suggesting a common source or person-to-person transmission. However, geographical unrelatedness among those patients with intestinal E. histolytca infection suggests that person-to-person transmission of E. histolytica might have occurred among MSM.
In univariate analysis, patients acquiring amebiasis were predominantly MSM and had significantly higher CD4 counts than those who remained uninfected (315 vs. 157 cells/L; p<0.001) (data not shown). In multiple logistic regression analysis, we found that MSM was the only risk factor that was associated with new acquisition of E. histolytica infection by serologies or antigen assays followed by PCR, with an adjusted odds ratio of 14.809 (95% CI, 1.824, 120.237; p = 0.01) when compared with heterosexuals or patients with other risk behaviors (data not shown). The adjusted odds ratio for new acquisition of E. histolytica infection for every 50-cells/L CD4 increase or 1−log10 copies/ml plasma HIV RNA load decrease was 1.066 (95% CI, 0.975, 1.167; p = 0.16) and 1.180 (95% CI, 0.813, 1.713; p = 0.38), respectively.
This is the first longitudinal follow-up study to investigate the incidence of E histolytica infection in HIV-infected patients by examining the incidence rate of intestinal E. histolytica infection and seroconversion of anti-E. histolytica antibodies. We found that HIV-infected MSM were at significantly higher risk for acquisition of E. histolytica infection [31],[34]. Despite immunosuppression from HIV infection and the complicated disease course of IA, clinical responses to metronidazole therapy were favorable in terms of rapid defervescence and a low attributable mortality rate.
Exposure to E. histolytica, but not E. dispar, may induce anti-E. histolytica antibody response and development of anti-E. histolytica antibodies may represent either recent or remote exposure to E. histolyica [14],[42], although not every person infected with E. histolytica develops an antibody response. Our analysis also showed that acquisition of E. histolytica infection was significantly associated with seroconversion for anti-E. histolytica antibodies despite immunosuppression from HIV infection. Therefore, such a test may be used as a complimentary tool to understand the epidemiology of E. histolytica among high-risk populations. By using serological surveys, Japanese investigators have found that the seroprevalence of E. histolytica infection in MSM was as high as 13.4–20.4% compared with 1.0% in heterosexuals and 0.8% in prostitutes [24]–[26],[33]. Similarly, we found that the seroprevalence of E. histolytica infection among HIV-infected patients remained significantly higher compared with HIV-uninfected persons with gastrointestinal symptoms who had their sera tested for anti-E. histolytica antibodies [34].
In this study, we further explored the sero-incidence of E. histolytica infection in HIV-infected patients. The results also indicate that MSM are at increased risk of exposure to E. histolytica infection. The findings of higher seroprevalence and sero-incidence of MSM, regardless of HIV status, are caused by higher prevalence and incidence of intestinal infection with E. histolytica. In our previous study, we found that the prevalence of E. histolytica/E. dispar by stool antigen tests was 12.1%, compared with 1.4% healthy controls; and at least 25% of the isolates from HIV-infected persons were confirmed as E. histolytica by PCR [31]. Although the majority of persons infected with E. histolytica are asymptomatic [1],[2], more than 80% of the 600 or greater annually reported cases of amebiasis in Japan occurred in MSM [32]. These findings may reflect a decrease in E. histolytica infection in developed countries where improvement of public hygiene and sanitation has reduced the risk of acquisition of E. histolytica through contaminated water or food.
Sharing the identical transmission route with E. histolytica, the transmission of E. dispar among MSM in developed countries is correlated with oral-anal sexual contact and 20–40% of MSM who visited the sexually transmitted diseases clinic were found to be infected with E. dispar [10]–[12]. Therefore, infection with either E. dispar or E. histolytica is indicative of unsafe oral-anal sexual contact among MSM. In this study, we further demonstrated that HIV-infected MSM were more likely than other risk groups to acquire E. histolytica infection during follow-up, although this finding is not statistically significantly different due to the small sample size. Nearly 4% of HIV-infected MSM acquired E. histolytica, compared with 1% among patients from other risk groups. Furthermore, case clustering that was identified by molecular typing of the isolates occurred probably through person-to-person transmission. These findings highlight the importance of counseling MSM about precautions to prevent acquiring E. histolytica infection through oral-anal sexual contact.
The clinical manifestations of IA in our HIV-infected patients with significant immunosuppression were similar to those previously described in HIV seronegative patients. Amebic colitis and liver abscesses were the two most common presentations [1],[2],[29]. The severity of the diseases was reflected by the high proportion of liver abscesses (59.7%) and complications (11.9%) in our cases. Despite low CD4 counts upon diagnosis with IA, the responses to metronidazole therapy with or without combination with antibiotics were favorable, as shown by rapid defervescence within 2 days of therapy initiation and no death attributable to IA.
There are several limitations of our study. First, the risk for exposure to E. histolytica is low in the general population in Taiwan, as reflected by the low seroprevalence (0.12%) of E. histolytica infection among 2500 healthy controls in a recent survey in northern and southern Taiwan [34]. Therefore, generalizations about our findings to areas of higher endemicity of E. histolytica and HIV infection should be cautious. Second, most patients at the late stage of HIV infection who develop HIV-related complications are referred to this hospital. However, those patients with IA had significantly higher CD4 counts than the patients without IA, suggesting that E. histolyica infection may not be associated with immunosuppression in HIV-infected patients. Rather, it is the risky behavior that increases risk of E. histolyica infection and subsequent development of invasive diseases. Third, our study was limited by the small sample size in assessment of the incidence of E. histolytica infection by stool antigen assays during follow-up. Although the incidence of E. histolytica infection is higher in MSM than in heterosexuals and others, the difference does not reach statistical significance. The shedding of E. histolytica may be intermittent, which may reduce the sensitivity of antigen assays if only one stool sample is tested. However, combinations with IHA assays for E. histolytica infection in our study may compensate for this deficiency by increasing the detection sensitivity. In this study, we chose a high titer of 128 as the cut-off value which decreases the possibility of cross-reactions, and seroconversion was significantly associated with newly acquired E. hisitolytica infection. Last, our genotyping methods [40],[41] may not be as sensitive enough for detection of genetic differences between the isolates as the new genotyping system that uses 6 tRNA-linked short tandem repeats by Ali and colleagues [43].
In conclusion, HIV-infected MSM in Taiwan are at a higher risk of acquisition of E. histolytica infection and IA than other HIV-infected patients. It should also be investigated whether this is the case in other countries. Certainly physicians, treating MSM with or without HIV infection, should be aware of this potential complication, that until recently, in industrialized countries was seen nearly only in travelers returning from E. histolytica endemic regions. |
10.1371/journal.pbio.1002355 | Interpreting the Dependence of Mutation Rates on Age and Time | Mutations can originate from the chance misincorporation of nucleotides during DNA replication or from DNA lesions that arise between replication cycles and are not repaired correctly. We introduce a model that relates the source of mutations to their accumulation with cell divisions, providing a framework for understanding how mutation rates depend on sex, age, and cell division rate. We show that the accrual of mutations should track cell divisions not only when mutations are replicative in origin but also when they are non-replicative and repaired efficiently. One implication is that observations from diverse fields that to date have been interpreted as pointing to a replicative origin of most mutations could instead reflect the accumulation of mutations arising from endogenous reactions or exogenous mutagens. We further find that only mutations that arise from inefficiently repaired lesions will accrue according to absolute time; thus, unless life history traits co-vary, the phylogenetic “molecular clock” should not be expected to run steadily across species.
| We relate how mutations arise to how they accumulate in different sexes, with age and with cell division. This model provides a single framework within which to interpret emerging results from evolutionary biology, human genetics, and cancer genetics. We show that the accrual of mutations should track cell divisions not only when mutations originate during DNA replication but also when they arise through non-replicative mechanisms and are repaired efficiently. This realization means that previous observations of correlations between mutation and cell division rates actually provide little support to the commonly held belief that most germline and somatic mutations arise from replication errors. We further find that only mutations that arise from inefficiently repaired lesions will accrue according to absolute time; thus, without covariation in life history traits, the phylogenetic “molecular clock” should not be expected to run at constant rates across species.
| Because mutations are the ultimate source of all genetic variation, deleterious and advantageous, mutagenesis has been of central interest even before the discovery of DNA as the genetic material (e.g., [1]), and developing a model of mutational heterogeneity along the genome is a major focus of current disease mapping studies [2,3]. From many decades of research into mechanisms of DNA replication, damage, and repair, we know that mutations can arise from errors during replication, such as the incorporation of a non-complementary nucleotide opposite an intact template nucleotide during DNA synthesis [4], or from DNA damage caused by exogenous mutagens or endogenous reactions at any time during normal growth of a cell (Fig 1). If uncorrected by the next round of DNA replication, these lesions will lead to arrested replication and cell death, or to mutations in the descendent cells (either because of incorrect template information or due to lesion bypass by error-prone DNA polymerase) [5].
While the fraction of mutations that is non-replicative in origin remains unknown, the common assumption is that mutations are predominantly replicative [6–9]. The basis for this assumption is a set of observations from disparate fields suggesting that, at least in mammals, mutations seem to track cell divisions. First, in phylogenetic studies, it has been observed repeatedly that species with longer generation times tend to have lower substitution rates, which under neutrality reflects lower mutation rates per unit time (“the generation-time effect”) (e.g., [7,10]). Second, based on comparisons of X, Y chromosomes and autosomes, it has been inferred that substantially more mutations arise in the male than in the female germline (e.g., [6,8,11]). In human genetics, pedigree resequencing studies have confirmed a male bias in mutation of approximately 3:1 at a paternal age of 30, and revealed a linear increase in the number of mutations in the child with the father’s age (e.g., [12,13]). These observations are all qualitatively consistent with mutations arising from the process of copying DNA: all else being equal, organisms with shorter generation times should undergo more germ cell divisions per unit time; in mammals, oocytogenesis is completed by birth whereas spermatogenesis is ongoing since puberty throughout the male lifespan, resulting in more germ cell divisions in males than females (Fig 2A) [14,15].
An informative exception to the “generation time effect” seen in phylogenetic studies is transitions at CpG sites, which represent approximately a fifth of de novo germline mutations [12], and show relatively constant substitution rates across species [16–18]. Their more “clock-like” behavior may reflect their distinct molecular origin [16], as CpG transitions are believed to be due primarily to the spontaneous deamination of the 5-methylcytosine (5mC) [19]. This case demonstrates the potential importance of non-replicative sources in germline mutations and raises the possibility that, despite the usual assumption (e.g., [20,21]), not all non-CpG mutations arise from mistakes in replication.
A third argument for the preponderance of replication errors has been made recently in cancer genetics, on the basis of two observations: (i) that somatic mutations tend to accrue more rapidly in tissues with higher renewal rates [22] and (ii) that, across tissues, the lifetime risk of cancer is associated with the total number of stem cell divisions [9]. Together, these findings were interpreted as indicating that in humans, random errors that occur during DNA replication are the source of most somatic mutations, and hence the main determinant of the odds of developing driver mutations that lead to cancer [9]. However, sequencing of tumor samples also revealed characteristic mutation patterns (“mutational signatures”) that reflect known DNA damage processes by endogenous or exogenous sources [23]. Moreover, environmental mutagens are known to influence the incidence of a subset of cancers, implying a role of mutations of non-replicative origins (e.g., [24,25]). These apparently conflicting observations again point to the importance of understanding how mutations arise in somatic tissues as well as in the germline.
Because, to date, arguments for the replicative origin of mutations have been qualitative and often based on implicit assumptions, we decided to model how the source of mutations relates to their rate of accumulation over cell divisions. For replication-driven mutations, we describe how mutations are expected to accumulate with age, and hence how the generation time relates to the yearly neutral mutation rate. This simple derivation allows us to show that, all else being equal, increases in the generation time will lead to decrease in the mutation rate only under very specific conditions on other parameters. For non-replicative mutations, we relate the mutation rate to rates of DNA damage, repair, and cell division. We show that only when the repair of DNA lesions is highly inefficient will mutations accrue according to absolute time. Otherwise, the accrual of mutations is expected to depend not only on absolute time but also on the rate of cell divisions—a feature previously thought to be specific to replication-driven mutations. By providing explicit expectations for how mutations should accumulate with sex, age, and cell division, these models provide a framework within which to interpret observations from evolutionary biology, human genetics, and cancer genetics.
The mutation rate per generation, i.e., the total number of germline mutations between two consecutive generations, is the sum of mutations inherited from both parents, which arose in the lineages of germ cells that gave rise to the child. If mutations are introduced by replication errors, their accumulation will track rounds of DNA replication. In each developmental stage, the number of replication-driven mutations can then be expressed as the product of the number of cell divisions and the mutation rate per cell division. Although a constant mutation rate per cell division is often assumed, explicitly or implicitly [6,26], this need not hold, especially when the cell lineage goes through different development stages, as do germ cells of multicellular organisms. Thus, we consider a more general case, allowing for variation in per cell division mutation rate (e.g., a higher mutation rate in early embryonic development) [27] and describe the accumulation of replication-driven mutations as a piece-wise linear process (following [18]).
For simplicity, we divide germ cell development from fertilization to reproduction into four stages, separated by the settlement of primordial germ cells in the developing gonads (which almost coincides with sexual differentiation), birth, and onset of puberty, respectively. Let dis and μis be the numbers of cell divisions and replication error rate in the ith stage (i = 1, 2, 3, 4) in sex s (s ϵ{f,m}). Because there is no sex difference in the first stage, d1f = d1m and μ1f = μ1m, and we replace them by d1 and μ1 (see Table 1 for a list of parameters involved in the model). Previous studies in Drosophila melanogaster suggest that the first division of a zygote has an extraordinarily high mutation rate [27,28]. Although the first division in Drosophila is quite distinct from that in mammals, it is possible that it would be more mutagenic in mammals as well, so we consider the first division separately as stage 0, of which the mutation rate is μ0 for both sexes, and re-define stage 1 as from the second post-zygotic division to sex differentiation. The total number of replication-driven autosomal mutations from one parent to the offspring is then:
MRs=(μ0+μ1d1+μ2sd2s+μ3sd3s+μ4sd4s)H,sϵ{f,m}
where H is the total number of base pairs in a haploid set of autosomes.
In mammals, all mitotic divisions of female germ cells are completed by birth of the future mother, so d3f = 0 and d4f = 0, and the total number of replication-driven mutations inherited from mother is (Fig 2B red line):
MRf=(μ0+μ1d1+μ2fd2f)H.
(1)
In contrast, male germ cells undergo divisions in all stages outlined above; furthermore, the number of germ cell divisions after puberty (d4m) is not a fixed number, because after puberty, sperm are continuously produced through asymmetric division of spermatogonial stem cells, at a roughly constant rate. If we assume that males and females have the same ages of onset of puberty and reproduction (denoted by P and G respectively), and that a spermatogonial stem cell undergoes cm divisions each year, the total number of paternal mutations is a function of reproductive age G (Fig 2B blue line):
MRm=[μ0+μ1d1+μ2md2m+μ3md3m+μ4m(cm(G−P−tsg)+dsg)]H,
(2)
where tsg and dsg are the time (in years) and the number of cell divisions needed to complete spermatogenesis from spermatogonial stem cells. The two divisions in meiosis are counted as one here, because only one round of DNA replication takes place in meiosis.
Summing Eqs 1 and 2, the total number of autosomal replication-driven mutations inherited by a diploid offspring from both parents is (Fig 2B purple line):
MR=MRf+MRm=[2μ0+2μ1d1+μ2fd2f+μ2md2m+μ3md3m+μ4m(cm(G−P−tsg)+dsg)]H.
By dividing Eq 2 by Eq 1, we obtain the ratio of male to female replication-driven mutations:
αR=MmMf=μ0+μ1d1+μ2md2m+μ3md3m+μ4mdsgμ0+μ1d1+μ2fd2f+μ4mcmμ0+μ1d1+μ2fd2f⋅(G−P−tsg),
which suggests that, keeping other parameters unchanged, increases in generation time G will lead to a stronger male bias in mutation, as expected intuitively (Fig 2C).
It follows that the average yearly mutation rate (i.e., the substitution rate if all mutations are neutral) is a function of G:
mR,y=mR,gG=2μ0+2μ1d1+μ2fd2f+μ2md2m+μ3md3m+μ4mdsg+μ4mcm(G−P−tsg)2G⋅
(3)
In order to explore the effect of generation time on the average yearly mutation rate, it is useful to reorganize Eq 3 as:
mR,y=μ4mcm2+A*2G,
(4)
where A*=2μ0+2μ1d1+μ2fd2f+μ2md2m+μ3md3m−μ4m(cmP+cmtsg−dsg), which is independent of G.
Eq 4 suggests that if and only if A* = 0 will the yearly mutation rate be independent of G. Otherwise, mR,y will either increase or decrease monotonically with G, depending on the sign of A*. Changes in the timing of puberty (P), in the number of cell divisions (dis) and in the replication error rate per cell division in each stage (μis) will also influence the dependence of mR,y on G.
The relationship between mR,y and G can also be directly read off the curve in Fig 3. The mutation rate per generation increases linearly with G after puberty, but this linear relationship does not apply to the period before puberty. If and only if the extended fitted line passes through the origin will the mutation rate per generation be exactly proportional to the generation time, and the average yearly mutation rate unaffected by G. If the intercept of the extrapolated line at age zero is positive, mR,y decreases with G, consistent with the observed “generation time effect” in primates. Conversely, if the intercept is negative, mR,y increases with G. In fact, the intercept obtained by extrapolation is exactly A* in Eq 3, so interpretation from Fig 3 is equivalent to that suggested by Eq 4.
Although estimates of other parameters exist, little is known about the replication error rate per cell division in germ cells, so it is unclear whether A* is positive or negative. However, it seems highly coincidental that an expression that involves multiple variables would happen to equal zero. Therefore, we argue that there is almost certainly an effect of generation time on yearly mutation rate in humans, although the magnitude of the effect could be small. The magnitude of the paternal age effect in pedigree data suggests that there should be generation-time effect in humans (see S1 Text).
Our model further reveals that, all else being equal, a longer generation time can lead to either an increase or decrease in the average yearly rate at which replicative mutations accrue. Therefore, the general observation that substitution rate in mammals tends to decrease with increasing generation times [7,10,16] is not necessarily expected; in fact, its existence requires very specific conditions on ontogenesis to hold (shown in Fig 3B). Moreover, given the current understanding of germ cell development in humans, the generation-time effect implies a higher mutation rate per cell division in early embryonic development than in spermatogenesis (see S1 Text for a discussion of available data in humans and chimpanzees).
Since mammalian species differ drastically in life history traits as well as development and renewal processes of germ cells [26,29], Eq 4 implies that the yearly mutation rate likely varies among species (even if per cell division mutation rates remain constant). As a result, unless life history traits co-vary in certain ways, we should not expect neutral substitution rates to be constant across mammalian species—or even along single evolutionary lineages. An important implication is that changes in life history among hominins [30] introduce uncertainty about dates in human evolution obtained under the assumption of a molecular clock [31].
DNA is subject to large numbers of damaging events every day as a result of normal cellular metabolism, and more DNA lesions may be generated by exogenous agents [32]. Typical DNA damage includes depurination and deamination due to DNA hydrolysis; alkylation and oxidation of bases induced by chemicals such as ethylmethane sulfonate or reactive oxygen species; pyrimidine dimers caused by ultraviolet radiation; and single- or double-stranded breaks produced by gamma and X-rays. Most single-stranded lesions cannot pair properly with any regular bases (termed “noncoding lesions”) and thus will block DNA replication if unrepaired (Fig 1). However, a few alterations to nucleotides can pair with bases different from the original Watson-Crick partners; such lesions (termed “miscoding lesions”), if unrepaired before replication, will lead to irreversible replacement of a base pair after cell division (Fig 1) [5].
To model the accrual of non-replicative mutations, we start by considering deamination of methylated CpG sites, which is the best understood example of miscoding lesions, and discuss more complex mutagenesis mechanisms in the S2 Text. This modification turns the methylated cytosine (mC) into a thymine (T); if uncorrected before DNA replication, an adenine instead of a guanine will be incorporated into the nascent strand, which results in a mutation in one of the two daughter cells. While DNA replication and cell division are obviously two distinct events, they are tightly coordinated such that DNA is replicated exactly once before each division (other than in meiosis and under a few unusual conditions). In what follows, we therefore do not distinguish between the two events.
We model the proportion of damaged base pairs at the time of cell division by considering the effects of both damage and repair (Fig 4A). For simplicity, we assume that single-strand damage occurs at a constant instantaneous rate μ throughout cell cycle and that the repair machinery recognizes lesions at a constant rate r (Fig 4A). Thus, the proportion of base pairs that carry a lesion at time t after the last cell division, p1(t), is described by a simple differential equation:
dp1dt=μ(1−p1)−rp1,
with the initial condition p1(0) = 0.
The solution to the differential equation is:
p1(t)=μμ+r(1−e−(μ+r)t).
Because each unrepaired single-strand lesion leads to a base pair substitution in one of the two daughter cells, the average mutation rate in one cell division (i.e., the expected fraction of base pairs that differ between a daughter cell and its mother cell) is:
MNR(T)=12p1(T)=μ2(μ+r)(1−e−(μ+r)T),
(5)
where T is the time between two consecutive cell divisions (Fig 4B).
We assume that μ<<1/T for any biologically reasonable value of T, so even in the absence of DNA repair, the absolute mutation rate per base pair per cell division (≈½μT) is very small. In addition, we focus on a single cell lineage and assume an infinite sites model, in which each genomic site can be mutated at most once. Thus, the total mutation rate over many cell divisions is simply the sum of the mutation rates for every division.
A key feature of the result in Eq 5 is that the accumulation of mutations per cell division exhibits two different limiting behaviors, depending on the relative rates of cell division and repair. When the rate at which lesions are repaired is much slower than the rate of cell division (rT<<1), the number of mutations is approximately proportional to time between two rounds of DNA replication:
MNR(T)=μT2.
(6)
The intuition is that, for a cell under this condition, there is almost no time for the repair machinery to correct lesions, so almost all lesions result in mutations. Consequently, mutations accumulate at a constant rate regardless of the rates of cell division and repair (Fig 4B, red box). In other words, non-replicative mutations that are inefficiently repaired will track absolute time.
In contrast, in the other limit where the repair is highly efficient relative to the rate of cell division (rT>>1), the number of mutations approaches an equilibrium level by the time of cell division:
MNR(T)=μ2(μ+r).
(7)
As a result, mutations accumulate at a rate that is roughly proportional to the number of cell divisions, regardless of absolute time (Fig 4B, blue box). Here, the intuition is that when repair is highly efficient, the few lesions that have not been corrected tend to be those that arose right before the cell division, and therefore the time since the last division has little effect. Importantly, under this scenario, the accrual of mutations that arise from lesions mimics what would be expected from replication errors. We note that the existence of such an equilibrium comes from the assumption of no error in repair; however, even when errors in repair are taken into consideration, there exists a phase in which repair and damage roughly balance out, so the mutation rate is proportional to the cell division rate (see S2 Text).
To understand how the mutation rate of non-replicative mutations depends on absolute time and the rate of cell division in general, we derive the mutation rate per unit time as the product of mutation rate per cell division and the cell division rate (c = 1/T>0):
m(c)=cMNR(1c)=c2(1+R)(1−e−(1+R)μc).
(8)
The mutation rate m(c) has two limiting behaviors when c approaches infinity and zero, respectively, which have the same intuitive explanations as Eqs 6 and 7, respectively. Moreover, it can be shown that m(c) is a concave increasing function of c. In other words, in a given period of time, faster dividing cell lineages accumulate more non-replicative mutations than slowly dividing lineages, but the increase in the number of mutations is smaller than the increase in the cell division rate. Therefore, when repair is neither inefficient nor extremely efficient, and given fixed damage and repair rates, faster dividing lineages are expected to accumulate non-replicative mutations at a higher rate per year than more slowly dividing ones (Fig 4C and see Table 2 for a list of parameters involved in the model).
This model can be extended readily to incorporate more features, such as other types of non-replicative mutations as well as to understand phenomena such as the strand bias in mutations associated with transcription (see S2 Text) [33,34]. Although the quantitative results differ, the main conclusion holds: the accumulation of non-replicative mutations depends critically on the repair efficiency in relation to the cell division rate.
These results demonstrate the fundamental importance of repair efficiency in determining the dependence of mutation rates on age, sex, and cell division rate (Fig 5). When DNA repair is inefficient, we should expect a linear accumulation of damage-induced mutations, partially justifying the expectation that neutral substitution rates of non-replicative mutations should not depend on generation time or other life history traits, and hence may be constant across species. However, our model highlights additional conditions for this expectation to be met: in particular, it reveals that the clock-like behavior of CpG transitions in mammals not only requires a non-replicative origin but also implies both relatively low repair efficiency in germ cells and similar damage rates across mammalian species (Fig 5A).
A further implication is that the number of mutations of maternal origin should increase with the mother’s age for CpG transitions and other mutations that arise from inefficiently repaired lesions. In this regard, we speculate that the current lack of a detectable maternal age effect may be due to underpowered sample sizes (notably because of the strong correlation between maternal and paternal ages). In any case, our model predicts that a maternal age effect should be detectable with sufficient data and reliable identification of parental origin of mutations (e.g., by sequencing of a third generation). Conversely, the detection of a maternal age effect on mutation rate would provide prima facie evidence for the existence of non-replicative mutations that are not efficiently repaired (assuming no relationship between the age at which an oocyte is ovulated and the number of cell divisions experienced during oocytogenesis [35]).
Also of note, lesions that have the same damage rate but are recognized by distinct repair mechanisms may differ not only in their absolute mutation rates but also in their time dependencies. Indeed, changes to the repair efficiency (or to the division rate) could alter the sex and time dependence of non-replicative mutations; for example, decreases in repair efficiency could lead mutations that previously tracked cell division rates to depend more on absolute time. Therefore, the phylogenetic molecular clock should not necessarily run at a steady rate even for mutations due to spontaneous DNA damage.
Our modeling results also shed light on studies of somatic mutations. As an illustration, a recent single-cell sequencing study identified mutations in neurons from the cerebral cortex of three healthy individuals [36]. The numbers of mutations in each cell were similar regardless of the donor’s sex and age (ranging from 15 to 42 years, Fig 5C) [37]. The genome-wide distribution of the somatic mutations appeared to be associated with transcription, with most identified mutations being C to T transitions at methylated cytosines. These observations led the authors to conclude that the mutations that they observed were due to non-replicative damage that was poorly repaired [36]. However, if mutations are non-replicative in origin and not repaired, more DNA lesions should accrue in older individuals, even in post-mitotic cells. In light of our model, an explanation is that an equilibrium between DNA damage and repair was reached before adolescence, and thus that the number of mutations does not increase further with age (Fig 5C). If this is the case, then there should be fewer somatic mutations in post-mitotic neurons from younger individuals, in which the equilibrium has not been reached.
Similarly, the model helps to interpret patterns observed in tumor samples, in which the total number of somatic mutations increases with the age of patient at diagnosis and grows at higher rates in fast renewing tissues [22]. Deamination at CpG sites make substantial contribution to mutations in almost all cancer types and accumulate at constant yearly rates that appear to be positively correlated with the turnover rates of the corresponding normal tissues (Fig 5B) [23,38]. As we have shown, all else being equal, a positive correlation is expected even for mutations that arise from DNA damage, so long as lesions are not poorly repaired in all somatic tissues.
Importantly, then, the recently reported correlation between number of stem cell divisions and lifetime risk of cancer across tissues is consistent with mutations of both replicative and non-replicative origins, and does not provide any evidence that most mutations are attributable to replication mistakes in stem cell divisions (what the authors referred to as “bad luck” in [9]). Of course, tumorigenesis is a multistep process that depends not only on the accumulation of mutations but also on tissue architecture as well as the order and consequences of specific mutational events, and gaining insight into its causes will likely require consideration of all these facets. What our model makes apparent is that it will also be important to incorporate a realistic model for the source of mutations.
Similar arguments apply to the male bias in mutation found by resequencing pedigrees and the generation time effect in phylogenetics: neither observation provides evidence for a replication-driven mutational process, as they could also reflect mutations arising from residual lesions left after efficient repair. Given these considerations, it becomes clear that, based on available data, we still do not know if a substantial proportion of human germline and somatic mutations—including those at non-CpG sites—are non-replicative in origin.
In summary, we introduce a model that helps to interpret findings from studies of somatic mutations, human pedigrees, and phylogenies. Although very simple, its behavior appears to be robust. By making explicit the relationship between the genesis of mutations and their accumulation over ontogeny, the model reveals the critical importance of both the source of mutations and the repair efficiency of lesions. Because replicative mutations and non-replicative mutations can display similar properties when repair is efficient, none of the previous observations of correlations between mutation and cell division rates lends strong support to the commonly held belief that most mutations are replicative in origin. Further experimental work is therefore needed to distinguish between different sources of mutation. Notably, fitting models such as this one to growing data from diverse fields should provide a quantitative understanding of how DNA changes accumulate in somatic tissues during a lifetime and in the germline over evolutionary time scales.
|
10.1371/journal.ppat.1003966 | Insights into the Initiation of JC Virus DNA Replication Derived from the Crystal Structure of the T-Antigen Origin Binding Domain | JC virus is a member of the Polyomavirus family of DNA tumor viruses and the causative agent of progressive multifocal leukoencephalopathy (PML). PML is a disease that occurs primarily in people who are immunocompromised and is usually fatal. As with other Polyomavirus family members, the replication of JC virus (JCV) DNA is dependent upon the virally encoded protein T-antigen. To further our understanding of JCV replication, we have determined the crystal structure of the origin-binding domain (OBD) of JCV T-antigen. This structure provides the first molecular understanding of JCV T-ag replication functions; for example, it suggests how the JCV T-ag OBD site-specifically binds to the major groove of GAGGC sequences in the origin. Furthermore, these studies suggest how the JCV OBDs interact during subsequent oligomerization events. We also report that the OBD contains a novel “pocket”; which sequesters the A1 & B2 loops of neighboring molecules. Mutagenesis of a residue in the pocket associated with the JCV T-ag OBD interfered with viral replication. Finally, we report that relative to the SV40 OBD, the surface of the JCV OBD contains one hemisphere that is highly conserved and one that is highly variable.
| Polyomaviruses have been invaluable tools for biomedical research into basic cellular processes. It is becoming increasingly clear, however, that members of this family are also involved in human diseases, particularly among the immunocompromised and the elderly. The subject of this study, the JC virus (JCV), is a member of this family and the causative agent of a brain disease termed Progressive Multifocal Leukoencephalopathy (PML), a disease that is often fatal and for which there is no cure. Herein we present the high-resolution crystal structure of the origin binding domain (OBD) from the JCV initiator protein large T-antigen. Furthermore, we propose a molecular model for the oligomerization of the JCV T-antigen OBD that is based upon the crystal structure. We also report a novel pocket that modeling studies suggest is available when the OBD is site-specifically bound to DNA and therefore may represent a possible starting point for structure-based drug design.
| There are now twelve known human polyomavirus members (e.g., [1], [2]) and particularly for immuno-compromised individuals, there is an increasing association between these viruses and human diseases (reviewed in [3], [4], [5]). For example, JC virus (JCV) is the causative agent of Progressive Multifocal Leukoencephalopathy ((PML); reviewed in [6], [7], [8]); a demyelinating disease of the central nervous system [9], [10]. JCV is also a major opportunistic infection associated with acquired immunodeficiency syndrome [11], occurring in up to 5% of AIDS patients [12]. Further interest in JCV, which is present in approximately 50% of the general population [13], stems from the fact that a promising new treatment of multiple sclerosis (the monoclonal antibody Tysabri) is known to be associated with the induction of PML (reviewed in [8], [14], [15]). Studies have also suggested a possible association between infection with JCV and human brain and non-central nervous system tumors [16], [17]. Unfortunately, there is no specific treatment for JCV.
Central to the JCV life cycle is the replication of its genome. The JCV origin of replication has been the topic of numerous studies (e.g., [18], [19], [20], [21], [22]). The interactions between the origin with the viral initiator, large T-antigen (T-ag), has also been explored (e.g., [23], [24]). The T-antigens encoded by polyomaviruses are multi-domain, multifunctional proteins (reviewed in [25], [26]) that form hexamers and double hexamers at origins of replication (reviewed in [27]). Assays designed to monitor T-ag dependent JCV replication have been reported (e.g., [18], [28]), including a cell free replication system [29]. However, theories regarding how JCV replication takes place are largely based on studies of the replication of Simian Virus 40 (SV40) (reviewed in [26], [30], [31], [32]). For example, an in depth understanding of the enzymology of SV40 DNA replication was obtained following many elegant studies (reviewed in [32], [33], [34]). Related studies have focused on the roles played by the T-ag during SV40 replication (reviewed in [25], [27], [35], [36]). Our laboratories have focused on the multiple roles played by the central origin-binding domain (OBD) of the SV40 T-ag during viral replication (reviewed in [37], [38]). Functions of the OBD include site-specific binding to GAGGC sequences in the origin ([39], [40]), promoting oligomerization of T-ag (e.g., via the B3 motif [41], [42]), melting of the central region of the core origin [43], binding to ssDNA at replication forks [37], [44], [45] and recruiting cellular initiation factors (e.g., [46]).
Structural studies of T-ag have provided critical insights into how this single domain can engage in so many activities (reviewed in [37]). For example, structures of the SV40 T-ag OBD established how the A1 & B2 loops in the OBD bind site-specifically to the GAGGC repeats in the central region of the viral origin (i.e. Site II) [42], [47], [48], [49]. They also established how the same A1 & B2 loops engage other DNA structures (e.g., duplex DNA in a non-sequence specific manner [49] and ssDNA [44], [45]). Crystallography studies also established that the SV40 T-ag OBDs can bind to a fork like DNA structure [45]. The latter observation was one reason for suggesting that the SV40 T-ag OBD is eventually positioned at the replication forks (reviewed in [37]). The structures of additional domains of T-ag have provided many additional insights into the interactions needed to initiate viral DNA replication (e.g., [47], [50], [51], [52]). For example, structures of the C-terminal helicase domain have greatly increased our understanding of how hexameric helicases catalyze DNA replication (reviewed in [53]) and how the helicase and OBDs work together to interact with ds DNA [47].
The initiation of JCV DNA replication is a central event during the viral life cycle [24]. The shared nucleotide sequence identity between the T-ag genes of SV40 and JCV is 71% [54]. Therefore, it is perhaps not surprising that SV40 T-ag recognizes and binds the JCV origin both in vivo and in vitro [24], [55], [56], [57]. Studies show that the converse is not true; that is JCV T-ag is inefficient at promoting replication of an SV40 origin-containing plasmid [24]. Thus, while T-ag and the other proteins encoded by these viruses are highly homologous, they likely contain subtle but important structural differences. To examine these issues, we pursued structural and biophysical studies of the JCV T-ag OBD. The results from these studies suggest how the JCV T-ag OBD binds to the viral origin and its subsequent roles in oligomerization events. They also demonstrate that the JCV OBD contains a pocket that has not been described in previous structures of the polyomavirus OBDs. Collectively, these findings provide a preliminary molecular understanding of the initiation of JC virus replication.
A luciferase based assay for studies of polyomavirus DNA replication was previously reported [75]. We developed a similar assay for measuring levels of JCV replication (unpublished) using the pCMV JC T-ag plasmid and a second plasmid containing the JCV origin of replication that was termed pJCV ori. Additional replication reactions were conducted with JCV T-ags containing point mutations introduced at selected residues using the QuikChange Kit ((Agilent); with oligonucleotides containing the desired mutation. Western blots, conducted with the Pab 416 antibody against T-ag (Santa Cruz Biotechnology), were used to determine whether a given point mutation disrupted T-ag's stability.
The JCV OBD (residues 132–261 (Fig. 1A)) crystallized in three different forms that were termed form 1, form 2 and form 3 (Table 1). Form 1 has two molecules in the asymmetric unit cell, and together the three crystals provide four independent structures of the JCV OBD. The four structures are very similar; a superposition of the four JCV OBD structures revealed root mean-squared deviations (RMSDs) of less than 0.5 Å. Form 3, the highest resolution structure (1.32 Angstroms), is shown in Fig. 1B. The topology of the JCV OBD is a five-stranded antiparallel β-sheet sandwiched between two helices on either side (Figs. 1 A and B). A superposition of the four JCV OBDs structures onto the DNA-free SV40 OBD structure (the only other polyomavirus OBD to be solved in the absence of DNA [63]) revealed an additional low RMSD (between 0.85–0.88 Å over 121 Cα atoms).
JCV OBD region B3 (residues 216–220 [40]) is poorly ordered in crystal forms 1 and 2, but well ordered in form 3 (Fig. 1C). This loop is also poorly ordered in several of the SV40 OBD structures [45], [76], [77]. B3 is well ordered in form 3 because tartrate (a component in the crystallization mixture) modified lysine 168 in a manner analogous to lysine acetylation. The carboxyl groups of the tartrate stabilized the B3 residues via a series of backbone hydrogen bonds. There are no previous reports indicating that JCV T-ag Lys168 is acetylated and further studies are necessary to determine if the observed modification of Lys168 is functionally important.
Phylogenetic studies have established that the amino acid sequence for JCV T-ag is very similar to that of SV40 T-ag (e.g., [37]). Indeed, the amino acid sequence identity between the JCV and SV40 OBDs is 81.5% (106 amino acids identical/130 amino acids (Fig. 2A)). Given that the structures of the JCV and SV40 T-ag OBDs have both been determined, it was of interest to analyze these molecules in terms of the distribution of the identical, conserved and non-conserved residues (Fig. 2B; identical (blue), conserved (pale pink), non-conserved (magenta)). As might be predicted, the interior of the molecule is highly conserved as are the A1 and B2 motifs involved in both DNA binding and interface formation (discussed below) (Fig. 2B; right side). The non-conserved residues map primarily to the hemisphere that is opposite to the one containing the A1 and B2 loops (Fig. 2B; left side. Certain of the conserved and non-conserved residues are indicated).
The JCV origin of replication contains multiple high affinity GAGGC sequences that serve as binding sites for the JCV T-ag OBD [22]. The GAGGC binding sites are arranged as palindromic repeats in Site II and as direct repeats in Site I (Figure 3A).
It has been proposed that in the context of a full-length T-ag hexamer, the high local concentration of OBDs promotes their association [42], [84]. To better understand how the JCV T-ag OBDs may assemble during oligomerization, we examined the interactions among the OBDs within the three crystal forms. As described in this section, the largest interface between adjacent molecules is the same in all three crystals. This was unexpected because the three forms belong to different space groups and have different cell dimensions (Table 1).
In light of the findings derived from our structural studies, it was of interest to determine if particular residues in JCV T-ag are needed for replication. A luciferase-based assay for measuring levels of SV40 and HPV31 DNA replication was previously described [75]. This assay has been adapted for studies of JCV replication using plasmids containing JCV T-ag and the JCV origin of replication (materials and methods). Initially, we used this assay to determine whether residues in the JCV OBD pocket are critical for DNA replication. Inspection of Fig. 9A establishes that a T-ag molecule containing a pocket mutation (i.e., F258L: its location in the pocket is shown in Fig. 7A) does not support DNA replication. Moreover, it is clear from Fig. 9B that the F258L mutation does not cause destabilization of JCV T-ag. (In contrast, two additional mutations in the JCV associated pocket (i.e., L199N and L199R) did cause destabilization (data not shown)). In addition, we initiated studies designed to address whether certain “non-conserved” surface residues (Fig. 2) play a role in JCV replication. Therefore, additional replication assays were conducted with T-ag molecules having the Q240A mutation. Inspection of Fig. 9A establishes that relative to wt JCV T-ag, T-ag molecules containing the Q240A mutation are greatly compromised in terms of their ability to support DNA replication. It is also apparent from Fig. 9B that the decreased ability of the Q240A mutant to support replication is not due to T-ag destabilization. We also analyzed the ability of the F190A mutant to support replication. Surprisingly, this mutant consistently supported higher levels of replication than wild type T-ag (Fig. 9A); a result that is not explained by increased expression of JCV T-ag (Fig. 9B). Finally, no replication of the JCV origin containing plasmid was detected in the control reaction conducted in the absence of T-ag.
The full-length T-ags encoded by both SV40 [86], [87], [88] and JCV [89] form hexamers and double hexamers on their respective origins of replication. Based on previous biochemical and structural studies, we proposed a model for SV40 T-ag's dynamic interactions with the viral origin and its subsequent oligomerization to form double hexamers [37]. One feature of this model is the proposal that following site-specific binding to the GAGGC sequences in the core origin, the OBD domains within SV40 T-antigen rearrange to form hexameric spirals (e.g., [63], [76], [77]). Spiral formation is also a feature of many of the other initiators that have been used as models for studies of the initiation of DNA replication (e.g., [90], [91], [92], [93]). Therefore, spiral formation by replication initiators may be a general phenomenon (reviewed in [94]).
In view of the structures presented herein, we propose that the JCV T-ag OBD undergoes interactions with the JCV origin that are similar to those of the SV40 T-ag OBD (reviewed in [37]). Regarding the initial binding of the OBD to the GAGGC sequences, our analysis of the JCV T-ag OBD structure indicates that the A1 & B2 loops mediate site-specific binding via a mechanism that is similar to that used by the SV40 OBD ([42], [47], [49]; reviewed in [37]). Nevertheless, the ITC studies indicate that there are differences in the interactions between the JCV and SV40 T-ag OBDs and origin sequences. For example, the binding of the JCV T-ag OBD to an oligonucleotide containing the JCV Site II is weaker than the SV40 OBD/Site II interaction [79] (298.6 nM verses 93.5 nM). Related ITC studies demonstrate that the JCV T-ag OBD binds to the GAGGC containing Site I regulatory region with a much higher affinity than Site II (Kds of 18.3 nM and 298.6 nM; respectively. The SV40 T-ag OBD also preferentially bound to Site I [76]). Why the JCV and SV40 T-ag OBDs have different affinities for Site II, and such a wide range of affinities for different GAGGC containing substrates, is not known. Of interest, the B2 regions in the OBDs encoded by JCV and SV40 are identical [40] and there is only one amino acid difference in the A1 regions (H148 in the SV40 OBD is Q149 in the JCV OBD). Therefore, pronounced sequence differences between the A1 & B2 motifs do not explain the observed differences in affinity; however, subtle structural differences in DNA, the OBDs, or both may play a role. Previous SV40 based studies have also established that sequences flanking the individual GAGGC sites play a significant role in modulating OBD binding affinities [95]. Thus, additional studies, including the co-structures of the JCV OBD with oligonucleotides derived from Site II and Site I, are needed to explain the observed differences in OBD affinities for origin sub-fragments. Finally, the full-length T-ag's from JCV & SV40 also have different affinities for Site II [23], [24]. The ITC studies suggest that the differences in the affinities are, at least in part, a function of the OBDs.
The ITC experiments also indicate that four JCV OBDs bind simultaneously to the four GAGGC sequences in Site II. However, in the context of full-length T-ag it is unlikely that all four pentanucleotides are initially bound by the OBDs. This conclusion is based on previous biochemical experiments with SV40 T-ag [96], [97] and structural studies that indicate that once the helicase domain has oligomerized, the shortness of the spacer that links the helicase domain to the OBD restricts OBD binding to only the most proximate pentanucleotide [47]. The subsequent stage(s) during the initiation process at which the initially unbound pentanucleotides are bound by the SV40, and presumably JCV, OBDs remain to be determined. Moreover, studies of both murine [98] and Merkel [79] polyomaviruses have established that in those systems only three pentanucleotide repeats are necessary for DNA replication; further evidence that the interactions of polyomavirus T-ags with the pentanucleotides in Site II are complex.
How polyomavirus T-ags transition from their sequence specific binding mode to fully assembled hexamers and double hexamers is not understood. While the OBDs are monomeric in solution (e.g. [48]), it has been proposed that in the context of T-ag hexamers and double hexamers, the high local concentration of OBDs will promote their association ([42], [84]); reviewed in [37]). Consistent with this possibility, our previous structures of the SV40 T-ag OBD established that it forms a hexameric spiral within the crystal [63], [77]. Therefore, it is of interest that our current studies have established that the JCV T-ag OBD also forms a spiral in the crystal. As in the SV40 T-ag OBD spiral [63], the monomers in the JCV T-ag OBD spiral are arranged in a head-to-tail manner, and the A1 loops are in the DNA-free or “retracted conformation” (reviewed in [37]). An additional common feature of the JCV and SV40 spirals is that they contain a very positively charged central channel that could interact with DNA in a non-sequence specific manner (data not shown). Nevertheless, the spirals formed by the JCV and SV40 T-ag OBDs are not identical. For example, the JCV “spiral” contains 4 OBDs per turn while the SV40 OBD spiral has 6 OBDs/turn (diagrammed in Fig. 8C). In addition, the JCV T-ag OBD forms a right-handed spiral, whereas the SV40 forms a left-handed one. These observations raise the question, “how can different spirals form from T-ag OBDs utilizing very similar interfaces?”
Comparison of the existing spiral structures for the SV40 and JCV T-ag OBDs suggest a common “interface based” model for formation of the observed higher order structures. According to this model, the interface acts like a joint or pivot point and differences in the rotational and translational components of the interface promote the formation of the structures observed to date. For example, in the crystallographic spirals, the angles between the interfaces in the JCV and SV40 T-ag OBDs are very different (i.e., ∼90° and 60°; respectively). In addition, for a spiral to occur, instead of a flat ring structure, there is a requisite translational component (“rise”) to the interface (the SV40 spiral has a rise of ∼6 Å [63], while the rise in the JCV OBD spiral is ∼9 Å). The direction of the translation component relative to the principal rotational axis (i.e., up or down) results in either a left or right-handed spiral (Fig. 8C; legend). Furthermore, in the context of T-ag hexamers and double hexamers, the interactions between the OBDs are likely to be highly dynamic. Support for this postulate includes the relatively small size of the interfaces observed in the crystal structures and previous EM based studies showing multiple orientations of the SV40 T-ag OBDs [85]. In summary, plasticity in the OBD/OBD interface may contribute to the multiple higher-order conformations adopted by the OBD. Nevertheless, it is not known whether the tetrameric JCV T-ag OBD spiral forms in vivo or whether it can rearrange into a hexameric OBD spiral that is analogous to the one formed by the SV40 T-ag OBD. However, given the dynamic nature of the domains within T-ag, it is possible that under certain conditions (e.g., following assembly of the hexameric helicase domain), the tetrameric JCV T-ag OBD spiral rearranges to accommodate two additional OBDs.
The C-terminus of the JCV T-ag OBD contains a pocket into which the A1 and B2 residues are inserted. Furthermore, our studies have established that pocket residue F258 is necessary for JCV replication. However, whether this pocket is a general feature of polyomavirus OBDs is not known. The T-ag OBD-DNA co-structures derived from Merkel (PDB entry 3QFQ [79]) and murine polyomavirus (PDB entry 4FB3 [98]) did not contain suitable electron density for tracing of the residues in the C-termini of the OBDs. Therefore no clear pocket was observed in these structures and it is concluded that there is some flexibility in the C-terminal OBD residues. Analyses of SV40 OBD structures revealed that they contain a groove in the same location, but it is not as pronounced as the one in the JCV OBD structure. Regarding evidence for the OBD pocket in larger T-ag structures; a co-structure of a SV40 T-ag dimer, containing both the OBD and the helicase domain (PDB entry 4GDF) interacting with DNA, was recently reported [47]. This structure revealed two completely different orientations of the linker region connecting the two domains. In the structure in which the OBD is bound to pentanucleotide 1, the linker points away from the OBD and the relatively shallower groove is observed. In the second or “hidden site”, the linker bisects the putative pocket. Together, these observations indicate that the “pocket” in SV40 T-ag may be part of a dynamic structure. However, additional structural studies are needed to further characterize the pocket in the SV40 and JCV T-ag OBDs.
Previous studies have also established that the SV40 T-ag OBD serves as a module for binding cellular proteins (reviewed in [38]). For example, the RPA 70AB domain was reported to bind to the T-ag OBD via interactions that include those with R154 [46]. Furthermore, the Nbs1 subunit of the MRN complex binds to the OBD [99]. Given the central roles played by the OBDs during viral DNA replication (reviewed in [37]), the surfaces on the OBDs that interact with these and related cellular replication factors have likely been conserved. Therefore, it is of interest that the JCV and SV40 T-ag OBDs contain one surface that is highly conserved. This surface contains the DNA binding A1 and B2 loops, but also many additional conserved residues that may be involved in binding to cellular proteins (e.g., R154 associated with RPA recruitment). However, it is also apparent that the opposite hemisphere contains the majority of the non-identical residues and certain of these residues (i.e., Q240) are required for JCV replication. These variable regions may simply reflect genetic drift. Alternatively, they may be binding surfaces for cellular proteins encountered in the very different cell types in which these viruses replicate (i.e., monkey kidney cells needed for SV40 replication versus human glial cells needed for JCV replication). Finally, the F190A mutation leads to higher levels of JCV DNA replication. The biochemical basis for this increase is unknown and subsequent studies are needed to address this issue. Nevertheless, a sequence comparison of JCV, SV40 and BK reveals that while JCV T-ag has a bulky aromatic amino acid at position F190, the T-ags from SV40 and BK contain less bulky residues at comparable positions (SV40: S189; BKV: C191). The alanine substitution at JCV T-ag residue F190 introduces an amino acid that requires less space than a phenylalanine. Therefore, the F190A T-ag mutant is more analogous to the SV40 and BKV T-ags and this may be related to the observed increase in DNA replication.
The initiation of JCV DNA replication, and the regulation of this process, is a complicated process. It is apparent that many additional structures will have to be determined before a molecular understanding of the initiation of JCV replication is obtained. Nevertheless, the individual structures of the proteins involved will provide considerable useful information, including potential targets for drug design, such as the pocket within the JCV T-ag OBD described herein.
|
10.1371/journal.pgen.1004850 | A Multi-Megabase Copy Number Gain Causes Maternal Transmission Ratio Distortion on Mouse Chromosome 2 | Significant departures from expected Mendelian inheritance ratios (transmission ratio distortion, TRD) are frequently observed in both experimental crosses and natural populations. TRD on mouse Chromosome (Chr) 2 has been reported in multiple experimental crosses, including the Collaborative Cross (CC). Among the eight CC founder inbred strains, we found that Chr 2 TRD was exclusive to females that were heterozygous for the WSB/EiJ allele within a 9.3 Mb region (Chr 2 76.9 – 86.2 Mb). A copy number gain of a 127 kb-long DNA segment (designated as responder to drive, R2d) emerged as the strongest candidate for the causative allele. We mapped R2d sequences to two loci within the candidate interval. R2d1 is located near the proximal boundary, and contains a single copy of R2d in all strains tested. R2d2 maps to a 900 kb interval, and the number of R2d copies varies from zero in classical strains (including the mouse reference genome) to more than 30 in wild-derived strains. Using real-time PCR assays for the copy number, we identified a mutation (R2d2WSBdel1) that eliminates the majority of the R2d2WSB copies without apparent alterations of the surrounding WSB/EiJ haplotype. In a three-generation pedigree segregating for R2d2WSBdel1, the mutation is transmitted to the progeny and Mendelian segregation is restored in females heterozygous for R2d2WSBdel1, thus providing direct evidence that the copy number gain is causal for maternal TRD. We found that transmission ratios in R2d2WSB heterozygous females vary between Mendelian segregation and complete distortion depending on the genetic background, and that TRD is under genetic control of unlinked distorter loci. Although the R2d2WSB transmission ratio was inversely correlated with average litter size, several independent lines of evidence support the contention that female meiotic drive is the cause of the distortion. We discuss the implications and potential applications of this novel meiotic drive system.
| One of the strongest expectations in genetics is that chromosomes segregate randomly during meiosis. However, genetic loci that exhibit transmission ratio distortion (TRD) are sometimes observed in offspring of F1 hybrids. Meiotic drive is a type of non-Mendelian inheritance in which a “selfish” genetic element exploits asymmetric female meiotic cell division to promote its preferential inclusion in ova. We previously reported TRD on Chr 2 in the CC, a mouse recombinant inbred panel with contributions from three Mus musculus subspecies. Here we show that maternal TRD consistent with a novel meiotic drive system is caused by a copy number gain. This mutation is similar in size and structure to other known meiotic drive responders, such as the knobs of maize. A deletion of most of the copies is sufficient to restore Mendelian segregation, proving that the copy number variant is causative of the observed TRD. In the CC, and also the related DO population, the transmission frequency of the favored allele varies dependent on genetic background, demonstrating that this system is under genetic control. In conclusion, we describe a novel wild-derived meiotic drive locus on mouse Chr 2 that exploits female meiosis asymmetry to violate the Laws of Mendelian inheritance.
| Mendel’s Laws provide the theoretical foundation of transmission genetics and explain many of the inheritance patterns of biological traits in sexually reproducing organisms. The Laws state that each gamete receives a random collection of alleles—exactly one per pair of homologous loci—and that gametes unite at random. However, reports of exceptions to Mendelian inheritance date back almost to the rediscovery of Mendel’s Laws, and have been instrumental in elucidating the mechanisms of genetic inheritance [1–4]. Transmission ratio distortion (TRD) is defined as a significant and reproducible violation of the inheritance ratios expected under Mendel’s Laws [1,5–7].
Most observations of TRD are due to selection acting upon the products of meiosis (gamete selection) or fertilization (differential pre- or post-natal survival) [5–8]. The latter is a relatively common occurrence in experimental crosses in many types of organisms including plants and animals [8,9], and is routinely used to classify the essentiality of genes and alleles [9–14]. However, a small but increasing number of observations of TRD can be ascribed to the differential segregation of alleles during meiosis, a process called meiotic drive [1,10–14]. To qualify as such, meiotic drive systems must exhibit three characteristics: 1) asymmetry in the meiotic division(s) with respect to cell fate; 2) functional asymmetry of the meiotic spindle poles; and 3) functional heterozygosity at a locus that mediates attachment of a chromosome or a chromatid to the meiotic spindle [1,15,16]. Meiotic drive is an evolutionary force thought to contribute to karyotypic evolution [15–17] and maintenance of non-essential “B chromosomes” in multiple clades [17,18]. The incidence of meiotic drive is unknown, but given that it is a relatively strong evolutionary force that can lead to the rapid fixation of a selfish allele, it should be rare to observe in action [18–20].
In most plant and animal species meiotic drive is restricted to females, which undergo asymmetric meiosis. At the locus where TRD is observed, an allele that is subject to preferential segregation is termed a responder [19–21]. There are examples in many species of meiotic drive responder alleles that, when in heterozygosity, succeed in being transmitted to the functional product of the asymmetric meiosis more than half of the time (S1 Fig.). Responders in known meiotic drive systems typically involve multi-megabase, highly repetitive and heterochromatic sequences, such as the D locus in monkeyflower [11,21], knobs in maize [11,22], homogenously staining regions (HSRs) in wild mice [12,17,22,23] and centromeres and B chromosomes in multiple species [12,17,23,24]. Those systems have mostly proven intractable to molecular characterization, and thus the mechanism(s) by which they gain a segregation advantage are largely unknown. Meiotic drive may be promoted or suppressed by distorter loci (alternately referred to in some publications as effectors, modifiers or drivers).
It is rare for TRD at any single locus to be observed in multiple independent genetic backgrounds. An exception is TRD on mouse Chromosome (Chr) 2, which was reported first in interspecific backcrosses between C57BL/6J (a classical inbred strain, primarily of Mus musculus domesticus origin [24–28]) and SPRET/EiJ (a Mus spretus wild-derived inbred strain) [25–28]. In offspring from two different (C57BL/6JxSPRET/EiJ)xC57BL/6J backcrosses, the SPRET/EiJ allele was overrepresented across a 40 cM region on Chr 2 [25,28] and a ~140 Mb region on Chr 2 with a maximum transmission frequency of 0.66 [25,29]. TRD in Chr 2 was also reported in an F2 cross between two body weight selection lines, one of which (high body weight; M16i) was derived from the Hsd:ICR outbred stock (also known as CD-1) [29,30]. Additionally, in an advanced intercross between the Hsd:ICR-derived high-running selection line HR8 [30–32] and C57BL/6J, TRD in Chr 2 was present in the primary data but not reported in the corresponding manuscripts [31–35]. And recently, we reported TRD in Chr 2 in the Collaborative Cross (CC) [33–35]. The CC is a mouse recombinant inbred panel derived from eight genetically diverse inbred strains: the classical strains A/J, C57BL/6J, 129S1/SvImJ, NOD/ShiLtJ and NZO/HlLtJ and the wild-derived strains PWK/PhJ (M. m. musculus origin), CAST/EiJ (M. m. castaneus) and WSB/EiJ (M. m. domesticus) [35]. We reported TRD in favor of the WSB/EiJ allele across a ~50 Mb region in the middle of Chr 2 in three largely independent sets of CC lines. In the largest sample, involving 350 genetically independent CC lines, the WSB/EiJ allele was present on 22% of Chr 2 [35,36], a significant over-representation compared to the expected frequency of 12.5% (1/8).
Here we report our extensive genetic characterization of Chr 2 TRD in the CC and in the Diversity Outbred (DO), an outbred population initiated from 144 incompletely inbred CC lines and specifically tailored for high resolution mapping of complex traits [33–36]. Using a combination of classical genetics, whole genome sequence analysis and bioinformatics, we demonstrate conclusively that maternal transmission distortion is caused by a large copy number gain of a 127 kb DNA segment containing a single gene, Cwc22. We also provide compelling evidence that meiotic drive is required to explain the TRD in the progeny of heterozygous dams. Finally, we show that there exist several genetically determined levels of TRD controlled by unlinked genetic variation, which, to our knowledge, is unique among meiotic drive systems.
To test whether TRD of the WSB/EiJ allele in Chr 2 is present in the DO, we analyzed 1,175 animals from DO generation 8 (G8) that were genotyped using two related genotyping arrays (MUGA or MegaMUGA, see Materials and Methods). We sampled the genotypes of each individual at 1 Mb intervals along Chr 2 and then computed the overall frequencies of the eight founder alleles at each position. The WSB/EiJ allele was over-represented relative to the other seven founder alleles across a roughly 100 Mb region in the middle of Chr 2 (S2 Fig.). However, there was a striking difference in the level of distortion observed in the CC and the DO, with the WSB/EiJ allele frequency reaching a maximum of 0.22 in the CC compared to 0.55 in the DO. This result indicates that the additional outcrossing in the DO is associated with higher levels of TRD. We conclude that TRD favoring the WSB/EiJ allele is a general feature of crosses in the CC genetic background; however, the level of TRD may vary widely depending on the number of generations of outbreeding.
To determine the parental origin of the TRD, we analyzed 5,499 offspring from 18 experimental crosses in which exactly one parent was heterozygous for the WSB/EiJ allele in an interval spanning the region of maximum distortion on Chr 2 (75–90 Mb) [33–35,37,38]. In all cases the heterozygous parent was an F1 hybrid derived either from an intercross between the WSB/EiJ inbred strain and one of eight other inbred strains (the seven founder strains of the CC or PWD/PhJ), or from two CC strains, of which one was homozygous for the WSB/EiJ allele on Chr 2 and the other was homozygous for a non-WSB/EiJ allele. F1 hybrids were mated to either C57BL/6J or FVB/NJ mice, and their progenies were euthanized at birth and genotyped using genetic markers located in the region of maximum distortion. For each cross, we computed the TR of the WSB/EiJ allele and the non-WSB/EiJ allele using the aggregate genotypes across all litters from parents with identical genotypes (Table 1).
TRs in six paternally segregating crosses (rows 1–6 in Table 1) were as expected under the null hypothesis of Mendelian segregation (range 0.482–0.524, p ≥ 0.37). In contrast, the mean TR in maternally segregating crosses (rows 7–18 in Table 1) was 0.666 and deviated significantly from the null hypothesis (p = 3.4x10–89). We conclude that, in the genetic backgrounds tested, TRD in favor of the WSB/EiJ allele on Chr 2 is restricted to the progeny of heterozygous dams.
The TRs among maternally segregating crosses were significantly different (p = 2.4x10–90), demonstrating that TRD depends on genetic background (i.e., TRD is under genetic control). The 12 crosses using F1 hybrid dams can be divided into three classes based on the observed TR (S3 Fig.). F1 hybrid dams derived from crosses between WSB/EiJ and CAST/EiJ or PWD/PhJ showed no distortion (crosses 7–10 in Table 1; aggregate TR = 0.485, 95% CI = 0.46–0.51, p = 0.23). Moderate but significant distortion was present in F1 hybrid dams derived from crosses between WSB/EiJ and A/J, 129S1/SvImJ, NZO/HILtJ or NOD/ShiLtJ; and in (CC042/GeniUncxCC001/Unc)F1 hybrid dams (crosses 11–15 in Table 1; aggregate TR = 0.645, 95% CI = 0.61–0.68, p = 8.3x10–19). Finally, extreme distortion was observed in reciprocal (WSB/EiJxC57BL/6J)F1 hybrid dams and in (CC001/UncxCC039/Unc)F1 hybrid dams (crosses 16–18 in Table 1; aggregate TR = 0.943, 95% CI = 0.93–0.96, p = 9.6x10–193). We conclude that heterozygosity for the WSB/EiJ allele in the central region of Chr 2 is necessary but not sufficient to observe TRD, because TR was consistent with Mendelian inheritance in some dams that met that criterion.
We also conclude that the grandparental origin of the WSB/EiJ allele has no influence on TRD because the TR levels were not significantly different between three pairs of reciprocal F1 dams (compare crosses 7 and 8, 9 and 10 and 17 and 18 in Table 1; p = 0.53, 0.11 and 0.59, respectively).
To define the boundaries of the locus subject to TRD, we screened 61 CC lines and 378 DO mice that had been genotyped with MegaMUGA for recombinations involving the WSB/EiJ haplotype in the 75–90 Mb interval of Chr 2. We identified five DO females (DO-600, DO-681, DO-732, DO-832 and DO-OCA45) and two CC strains (CC039/Unc and CC042/GeniUnc) that each had at least one informative recombination (Fig. 1). Next, we mated four of the DO females (all except DO-OCA45 that was already heterozygous) and the two CC strains to one of two additional CC lines (CC001/Unc and CC005/TauUnc) that had no contribution from WSB/EiJ on Chr 2, to obtain heterozygous G1 hybrid females. Each hybrid female was genotyped with MegaMUGA and mated to FVB/NJ males (total of 35 crosses; S1 Table).
We found that dams carrying eight of the ten recombinant chromosomes exhibited significant TRD in the Chr 2 interval (TR range 0.69–1.0, p ≤ 2.1x10–5; Fig. 1 A), but dams carrying two other recombinant chromosomes did not (TR = 0.48 and 0.37, p ≥ 0.72; Fig. 1 B). These results are consistent with our conclusion that heterozygosity on Chr 2 is required but not sufficient for TRD; therefore, dams with Mendelian transmission ratios were not used for mapping the locus subject to TRD. Dams with TRD in favor of the WSB/EiJ allele were all heterozygous for a 9.3 Mb interval (the candidate interval; boxed in Fig. 1 A). The proximal boundary of the candidate interval is defined by the recombination found in the CC strain CC039/Unc (i.e., the most distal SNP inconsistent with a WSB/EiJ haplotype). The distal boundary of the candidate interval is defined by the recombination found in DO-732 and DO-832 females (i.e., the most proximal SNP inconsistent with a WSB/EiJ haplotype). Those SNPs define the maximum boundaries of the locus subject to TRD, Chr 2 76,860,362–86,117,205 (all positions from NCBI/37 unless otherwise noted).
Among the eight CC founder strains, the candidate interval has 5,018 SNPs, 1,286 small insertions/deletions and 35 structural variants that are private to the WSB/EiJ strain [37–39]. Although this very large number of variants would typically make it difficult to confidently identify and prioritize candidates, one large structural variant has several unique features that made it a strong candidate causative allele for the TRD phenotype. That structural variant is a copy number gain of a 127 kb-long genomic DNA segment (herein referred as R2d for responder to drive). In the reference genome, R2d is composed of nine non-contiguous sections that, in total, span 158 kb (see R2d1 locus; Chr 2 77,707,014–77,865,265; Fig. 2 A; S2 Table).
We used the normalized per-base read depth from whole-genome sequence alignments generated by the Sanger Mouse Genomes Project [31,32,37,39] and the HR8 selection line to estimate the number of copies of R2d in 18 inbred strains (see Materials and Methods). Similar to C57BL/6J, 15 of the 18 strains, including 5 additional CC founder strains (A/J, 129S1/SvImJ, NOD/ShiLtJ, NZO/HlLtJ and PWK/PhJ) were copy number one (i.e., a single haploid copy), and CAST/EiJ was copy number two. In contrast, WSB/EiJ had an estimated copy number of 34, and SPRET/EiJ had an estimated copy number of 36, resulting in ~4.4 Mb of additional DNA in those strains (Fig. 2 A). We sequenced 10 individuals from the HR8 selection line (for which Chr 2 TRD was also observed when mated to C57BL/6J [31,32,40]) to a total depth of 125x and aligned the reads to the reference genome. All 10 individuals had evidence of a copy number gain with the same boundaries as in WSB/EiJ and SPRET/EiJ (Fig. 2 A; mean copy number 24.5 +/- 1.4, equating to ~3 Mb of additional DNA).
We used two additional methods to assay the copy number of R2d. First, we identified sets of probes on two different genotyping arrays for which the sum hybridization intensity was highly correlated with the copy numbers estimated from sequencing read depth (34 probes in MDA and 3 probes in MegaMUGA; S3 and S4 Tables, respectively). Second, we used real-time quantitative PCR to estimate the R2d copy number (Fig. 2 B) using TaqMan assays internal to exons of the single protein-coding gene within R2d, Cwc22 (Fig. 2 C). Using that gene as a proxy for the copy number gain, we found that the copy number estimates from all three methods were highly concordant for the 28 sequenced strains/individuals.
Using the TaqMan assay, we also found that the M16i inbred strain has a high number of copies of R2d (Fig. 2 B). We conclude that a large increase (> 20-fold) in R2d copy number is found exclusively in strains with TRD (WSB/EiJ, SPRET/EiJ, HR8 and M16i) and that TRD consistently favors the transmission of the allele with the copy number gain.
Many structural variants identified from whole-genome sequencing reads have uncertain genomic positions due to the challenge of mapping large variants that are absent from the reference genome. To determine the position of the copy number gain associated with R2d, we mapped the WSB/EiJ and CAST/EiJ alleles using segregating populations that have been genotyped at medium (MegaMUGA) or high (Mouse Diversity Array, MDA) density [26,40]. In the CC founder strains, probes located in R2d have hybridization intensities correlated with the number of copies estimated from aligned read depth and TaqMan CNV assays (Fig. 2 A, B). The MDA provides robust discrimination between the reference (one copy), CAST/EiJ (two copies) and WSB/EiJ alleles (34 copies; Fig. 3 A). MegaMUGA is able to identify mice carrying the WSB/EiJ allele with little ambiguity (Fig. 3 B). Using the sum intensities of the informative probes as a quantitative trait, we mapped the WSB/EiJ and CAST/EiJ copy number gains in two independent populations and platforms. A genome scan identified a single, broad, highly significant peak on Chr 2 in each population, and those peaks overlap with each other and with the initial candidate interval for TRD (Fig. 3 C-E). We conclude that the copy number gain is closely linked to R2d1. This location is consistent with the large copy number gain being the causative allele. Note that both genome scans (Fig. 3 C, D) demonstrate that all the extra R2d copies found in WSB/EiJ are located in this interval because no other significant peak is observed in either scan. QTL mapping using TaqMan readout as the phenotype confirmed this result (Fig. 3 D, E).
Analysis of individual mice with recombinant chromosomes in the candidate interval revealed that the copy number gain maps to a 900 kb interval (the R2d2 locus; Chr 2 83,631,096–84,541,308; Fig. 2; Fig. 3 A, B). Specifically, the CAST/EiJ copy number gain (R2d2CAST; one additional copy of R2d) is located distal to the transition from the CAST/EiJ to the NZO/HILtJ haplotypes found in mice OR3172m10 and OR3172f9 because both mice have low hybridization intensity consistent with a single copy, hence they lack R2d2CAST (Fig. 3 A; S4A Fig.). Similarly, the WSB/EiJ copy number gain (R2d2WSB; 33 additional copies of R2d) is located proximal to the transition from the WSB/EiJ to the CAST/EiJ haplotype found on DO mouse DP2–446, because it had high hybridization intensity consistent with the presence of R2d2WSB (Fig. 3 B; S4B Fig.). These results demonstrate that R2d2 is not located immediately adjacent to R2d1 but approximately 6 Mb distal to it. The distal location of the copy number gain is confirmed by the analysis of the sum intensity of the three MegaMUGA probes that track R2d in two backcrosses involving the SPRET/EiJ inbred strain [26,41] (S4C Fig.).
We used the TaqMan assay to confirm R2d copy number in all heterozygous females tested for TRD (S1 Table; S5 Fig.). We identified a dam (DO-G13–44) that was homozygous for the WSB/EiJ haplotype across the entire candidate interval but produced offspring that were segregating for the copy number gain (Fig. 4 A). This was confirmed by estimating R2d copy number in each of 27 G3 females and 16 G4 progeny that were heterozygous for a WSB/EiJ haplotype (Fig. 4 B; S5 Fig.). We determined the TR in 825 progeny of G3 dams mated to FVB/NJ sires. The TRs among the 27 G3 dams were significantly different (p = 4.9x10–12). In the progeny of the 15 G3 dams with high copy number there was significant TRD in favor of the WSB/EiJ allele (TR = 0.78, p = 2x10–30; Fig. 4 C). In contrast, we found absence of TRD in the 12 G3 dams that inherited the low-copy allele (TR = 0.53, p = 0.234). A genome scan for TRD as a binary trait demonstrated that presence or absence of TRD in this pedigree maps uniquely to the candidate interval (Fig. 4 D, E).
We were also able to estimate that G3 dams with the low-copy allele had a copy number of ~11. We conclude that the loss of ~22 copies of R2d was sufficient to rescue Mendelian transmission, thus demonstrating that the copy number gain is causative of TRD.
The results presented above demonstrate that TRD at R2d2 is only observed in the progeny of heterozygous dams. This restricts the plausible causes of TRD to meiotic drive, genotype-dependent embryonic lethality (including genotype-dependent competition between embryos) or a combination of both. To identify the cause of TRD, we first determined whether TR levels (S6 Fig.; S1 Table) were correlated with litter size in 127 DO dams (these 56 DO-G13 and 71 DO-G16 females are a random sample from an outbred population). We observed a strong inverse correlation between average litter size and TR at R2d2 (r = -0.65, p = 7.2x10–8 and r = -0.40, p = 5x10–4 in the DO-G13 and DO-G16 dams, respectively; Fig. 5 A, B). We conclude that the presence and the strength of TRD are significantly associated with reduced litter sizes and thus with some type of embryonic lethality. We determined the relationship between TRD and litter size under the assumption of TRD caused exclusively by embryonic lethality [40,41] (S7 Fig.). Under this scenario, in both the DO-G13 and DO-G16 samples the observed average litter size is significantly greater than predicted based on TR (p = 0.021 and 6.0x10–5 for DO-G13 and DO-G16 dams, respectively; S7 Fig.). We conclude that embryonic death alone could only account for a fraction of the “missing” progeny inheriting a non-WSB/EiJ (R2d2NotWSB) allele. We determined directly the levels of embryonic lethality in DO-G13 dams at mid-gestation (see Materials and Methods). We observed that dams with TRD had slightly, but not significantly, higher numbers of resorbed embryos present in utero than did dams with Mendelian segregation (1.3 ± 1.5 and 1.1 ± 1.2 resorbed embryos, respectively, p = 0.66; N = 29 and 19 dams, respectively; S8 Fig.). We conclude that embryonic lethality alone is insufficient to explain TRD at R2d2.
Although embryonic lethality can change the proportion of progeny inheriting alternative alleles at R2d2, only meiotic drive can lead to an increase in the absolute number of progeny inheriting the R2d2WSB allele per litter in dams with TRD compared to dams with Mendelian segregation. To test whether meiotic drive was responsible for TRD, we determined the average absolute number of offspring per litter that inherited the R2d2WSB and R2d2NotWSB alleles in the progenies of the DO-G13 and DO-G16 DO dams with either TRD or Mendelian segregation. In dams with Mendelian segregation, the average numbers of offspring per litter that inherited either allele were not different (3.80 R2d2WSB versus 3.96 R2d2NotWSB, p = 0.73 in DO-G13 dams; 4.13 R2d2WSB versus 4.03 R2d2NotWS, p = 0.29 in DO-G16 dams; Fig. 5 A, B). In contrast, in the progenies of dams with TRD the average number of offspring per litter that inherited the R2d2WSB allele (4.51 and 4.89 in the DO-G13 and DO-G16 dams, respectively) was significantly greater than the absolute number of either allele in the offspring of dams without distortion (p = 0.006 and 0.049 for the R2d2WSB and R2d2NotWSB alleles in DO-G13; p = 0.005 and 4x10–4 for the R2d2WSB and R2d2NotWSB alleles in DO-G16; Fig. 5 A, B). The same result holds true for live embryos at mid-gestation: the average numbers of offspring that inherited R2d2WSB and R2d2NotWSB alleles were 5.0 ± 2.2 and 1.6 ± 1.8 for dams with TRD versus 4.3 ± 1.6 and 3.4 ± 1.8 for dams without TRD. Based on the consistent and significant excess average absolute number of R2d2WSB alleles in the litters of dams with TRD, we conclude again that meiotic drive is required to explain TRD at R2d2.
Further support for meiotic drive was provided by the analysis of the DO-G13–44 pedigree (Fig. 5 C) and crosses between (NZO/HILtJxWSB/EiJ)F1 dams and FVB/NJ sires (cross 15 in Table 1; Fig. 5 D). The average litter size of DO-G13–44 G3 dams inheriting the mutant R2d2WSB allele (R2d2WSBdel1) was larger than in dams inheriting the standard R2d2WSB allele (9.4 ± 2.9 and 6.8 ± 1.6, respectively), but the observed average litter size in dams with TRD is significantly greater than predicted based on TR (p = 0.02; S7 Fig.). Similarly, in the (NZO/HILtJxWSB/EiJ)F1 crosses the average litter size (7.7 ± 2.4; Fig. 5 D) was comparable to DO-G13 and DO-G16 dams without TRD, and was greater than predicted based on TR (p = 0.09; Fig. 5). There was little direct evidence of embryonic lethality at mid-gestation (1.8 ± 1.6 and 0.4 ± 0.5 resorbed embryos, respectively; S8 Fig.). Furthermore, DO-G13–44 G3 dams with different R2d2 alleles differed significantly in the average absolute number of offspring per litter inheriting the R2d2WSB allele (in dams with TRD) compared to the R2d2WSBdel1 allele (in dams with Mendelian segregation; 5.3 ± 2.0 and 4.64 ± 2.4, respectively, p = 0.07; Fig. 5 C). Similar results are observed when comparing the absolute number of offspring per litter that inherited the R2d2WSB allele in the (NZO/HILtJxWSB/EiJ)F1 crosses to the DO dams without TRD (5.1 ± 1.0 and 4.1 ± 1.1, respectively, p = 0.03; Fig. 5 D). In summary, all data from four independent experimental populations were consistent with an explanation of Chr 2 TRD that requires the joint presence meiotic drive and low-level embryonic lethality.
After demonstrating that TRD occurs only through the germline of F1 female mice, we were faced with two major obstacles in our efforts to map the causative locus. First, although heterozygosity for the WSB/EiJ allele is required, it is not sufficient for meiotic drive (Table 1; S1 Table). Therefore, we initially mapped the responder by determining the minimum region of overlap for the WSB/EiJ haplotype only in dams with TRD (Fig. 1). This yielded a 9.3 Mb candidate interval. Second, the candidate interval spans a recombination-cold region [37,40,42], and the frequency of recombination is three-fold lower than expected in the CC (Fig. 2 D). Although this likely contributes to the overall deficit in recombinant chromosomes (none observed versus an expected 23 in the 378 DO females and 4 in 61 CC lines), the complete lack of recombinants involving the WSB/EiJ haplotype is striking, and, for the purposes of this study, a major impediment to the precise mapping the responder.
Within the candidate interval, a single variant (R2d2) stands out as the most likely cause of TRD. R2d2 consists of one or more copies of a 127 kb sequence (R2d). High copy number (≥ 24) is present in all four strains with reported TRD and low copy number (≤ 2) is present in all eight strains without TRD (Fig. 2 A, C). The expansion in copy number leads to an increase of at least 3 Mb in DNA content within the allele favored by maternal TRD. Among CC founders, only WSB/EiJ has a high copy number allele.
As the reference genome is based on a single classical inbred strain, C57BL/6J, copy number gains in other strains or wild mice may be located in a different physical location. Fortunately, the presence of a third allele in CAST/EiJ (which exhibited a twofold enrichment of sequencing reads) combined with the fact that recombinations involving the CAST/EiJ haplotype are not suppressed within the 9.3 Mb candidate interval, enabled us to map the physical location of R2d2 to a 900 kb region located 6 Mb distal to R2d1, the locus where the sequencing reads mapped in the reference genome (Fig. 3). Importantly, the mapping of R2d2 was enabled by the availability of deep sequence data for each of the strains used in our experiments [25–27,37,42,43 and this study] and by combining the results of experiments completed 20 years apart [25–27,43,44].
We determined the number and spatial distribution of SNPs in the 9.3 Mb candidate interval that partition the ten inbred strains with whole genome sequence in a pattern consistent with the TRD phenotype (three strains with TRD: WSB/EiJ, SPRET/EiJ and HR8; and seven strains without TRD: A/J, C57BL6/J, 129S1/SvImJ, NOD/ShiLtJ, NZO/HILtJ, CAST/EiJ and PWK/PhJ). Compared to a genome-wide mean of 1 consistent SNP every ~3.2 kb, within the 900 kb region where we mapped R2d2 there was a mean of 1 consistent SNP every 883 bp (p < 1.0x10–4, one-sided Student’s t-test; Fig. 2 E). This reduction in diversity is not due to undercalling of SNPs in the R2d2 candidate interval (Fig. 2 F). The fact that consistent SNPs are rare in most of the genome but are common within the 900 Kb region in which R2d2 maps supports the hypothesis that R2d2 is the causative allele for TRD.
Most importantly, we identified a DO female (DO-G13–44) that was homozygous for the WSB/EiJ haplotype across the entire R2d candidate interval but was heterozygous for R2d2 alleles with different copy numbers (Fig. 4). We generated a three-generation pedigree and analyzed the R2d copy number, the Chr 2 haplotype and TR in the progeny of heterozygous dams with different copy numbers. This analysis revealed perfect correlations between the inheritance of R2d2WSBdel1 and complete absence of TRD in favor of the WSB/EiJ allele, and between the inheritance of R2d2WSB and presence of TRD. This experiment demonstrates that the reduction in copy number from 33 to 11 is sufficient to restore Mendelian segregation, and that R2d2 is the causative allele for maternal TRD.
Further evidence that TRD requires an R2d2 allele with copy number of above 11 is provided by the NU/J inbred strain. This strain has intermediate copy number (7, estimated by TaqMan) but no TRD in the progeny of (NU/JxC57BL/6J)F1 female hybrids (0.55, p = 0.55; S12 Fig.).
The presence of R2d sequences at two distinct locations (Fig. 2 G) indicates an initial duplication of this segment in the ancestor of CAST/Eij, WSB/EiJ, SPRET/EiJ and Hsd:ICR. R2d spans a highly expressed protein coding gene (Cwc22; Fig. 2 C) that is implicated in RNA splicing [38,44], a predicted gene of unknown function that overlaps with the last exon of Cwc22 (Gm13727) and a pseudogene (Gm13726). DNA copy number variation for Cwc22 has been described previously [38,45]. Cwc22 is highly expressed in mouse oocytes and fertilized eggs [45,46]. The Cwc22 gene is a known eQTL in mouse: allele-specific RNA-seq of brain tissue from reciprocal crosses between WSB/EiJ, PWK/PhJ and CAST/EiJ showed extreme differential expression, with the WSB/EiJ allele more highly expressed than the other two [46,47].
Apart from its size and repetitive nature, an important feature of the R2d2 locus is its remarkable uniformity between three divergent genetic backgrounds that are separated by ~1 million years of evolution: WSB/EiJ, SPRET/EiJ and HR8 [47–49]. For example in WSB/EiJ and SPRET/EiJ the genome-wide mean is 1 SNP every ~60 bp [37] and the mean SNP frequency within R2d is significantly reduced to 1 SNP every 1,342 bp (t-test, p = 3.9x10–58). Further analysis will be required to determine the respective ages of the duplication and the copy number change(s), and whether interspecific introgression [48–51] is required to explain the unlikely degree of sequence conservation between M. m. domesticus and M. spretus.
We note that, while unlikely given the results of our QTL mapping (Fig. 3), it is possible that there have been additional duplication events that have also inserted R2d in other chromosomes. Additionally, the causal allele may incorporate additional DNA sequences, including some that may be absent in the reference genome (similar to the origin of the sequence on maize chromosome Ab10 that causes meiotic drive in that species). If that is the case, the causal allele may be much larger than 4.4 Mb. For example, HSR alleles as large as 200 Mb have been described [50–52].
A second focus of our study was to discriminate among the many mechanisms [29,52] that could give rise to TRD at R2d2, and to rule out as many as possible. First, the fact that TRD is only observed through the maternal germline rules out both spermatogenesis-mediated processes and sperm competition. Second, the presence of TRD at birth rules out differential survival of offspring. Third, the fact that distortion was independent of the maternal granddam precludes cytoplasmic effects. The remaining plausible explanations are differential fertilization based on the oocyte genotype, embryonic lethality and/or meiotic drive. The first two mechanisms should reduce the average litter size proportionally to TR (black line in S7 Fig.), while the average absolute number of offspring inheriting the favored genotype (R2d2WSB) per litter remains constant. The number of resorbed embryos observed in pregnant females could distinguish the two mechanisms because it should be greater in the second than in the first scenario. In contrast, if meiotic drive is solely responsible for TRD then the following should be true: 1) average litter size is independent of TRD, 2) the average absolute number of offspring inheriting the favored genotype (R2d2WSB) per litter is higher in dams with TRD than in dams with Mendelian segregation, and 3) the level of embryonic lethality is independent of the presence and level of distortion. The data shown in the Results section are most consistent with the combined action of embryonic lethality and meiotic drive. Specifically, meiotic drive is required to explain both the fact that the observed average litter size in the DO-G13 and DO-G16 dams, in the DO-G13–44 pedigree and in the (NZO/HILtJxWSB/EiJ)F1 dams is greater than predicted based on TR (S7 Fig.), and that the average absolute number of offspring inheriting the R2d2WSB genotype per litter is greater in dams with TRD (Fig. 5). Note that some p-values in comparisons involving (NZO/HILtJxWSB/EiJ)F1 crosses failed to reach statistical significance due to the small sample size, but the trends were always consistent with those in DO dams with TRD.
An alternative explanation that does not involve meiotic drive would require the combined presence of increased ovulation in dams with TRD and pre- or post-implantation genotype-dependent competition between embryos favoring the allele with the high copy number at R2d2. Genotyping at R2d2 and re-analysis of 159 F2 females from the M16ixL6 intercross [29] confirms an overdominant effect of the R2d2 genotype in the number of live and dead embryos at day 16 of gestation, as predicted under the meiotic drive and embryo competition scenarios, but shows no effect of the R2d2 locus on ovulation rates (S9 Fig.). This result is not due to a lack of power, as we have 80% power (at a = 0.5) to detect a difference in the mean ovulation rate du to an effect of the R2d2 genotype and QTLs for ovulation rate were identified in the original study [29,53–55]. In summary, the effect of the R2d2 genotype on reproductive phenotypes is most consistent with the meiotic drive hypothesis. However, the possibility remains that the genotype-associated difference in number of live embryos may be due to differential fertilization or implantation. Additional breeding experiments and genotyping of pre-implantation embryos will resolve the remaining questions concerning the mechanisms involved in TRD at R2d2.
It is interesting to speculate about the types of embryonic lethality that are consistent with our data and with previous reports of TRD on Chr 2. Lethality is associated with distortion at R2d2, and thus the simplest explanation is preferential death of embryos inheriting maternal R2d2NotWSB alleles. However, such a scenario would require parent-of-origin-dependent death of embryos with maternal C57BL/6J, 129S1/SvImJ, NOD/ShiLtJ and NZO/HILtJ R2d2 alleles in crosses involving F1 females (Table 1) and CAST/EiJ, PWK/PhJ and A/J R2d2 alleles in the CC/DO females (S10 Fig.). The lack of evidence of TRD and parent-of-origin lethality in dozens of crosses involving these alleles [53–55], combined with the lack of evidence for imprinted genes in the central region of Chr 2 [2–4,46,56], appears to rule out this explanation. Specifically, the Cwc22 gene present in R2d is not imprinted in brain, kidney, lung and liver in crosses involving the WSB/EiJ, PWK/PhJ and CAST/EiJ strains [46]. A more likely explanation for the joint and correlated presence of meiotic drive and lethality is that the unequal segregation of chromosomes and/or chromatids that leads to TRD in euploid embryos may also lead to increased Chr 2 aneuploidy, and thus to embryonic death (all autosomal aneuploidy is embryonic-lethal in the mouse). This would also explain the slight increase in the number of resorbed embryos observed at mid-gestation (S8 Fig.; S1 Table). This hypothesis makes the testable prediction that Chr 2 should be especially affected by aneuploidy in some dams with TRD.
Importantly, co-segregation of a deletion allele of R2d2 and increased litter size in the DO-G13–44 pedigree demonstrates that lethality is mediated by an element within the R2d repeat.
Overall, we assessed TR at R2d2 in hundreds of females carrying a single WSB/EiJ allele in at least nine distinct genetic backgrounds (Table 1; S1 Table). The presence of significantly different TR levels among F1 hybrid dams, combined with the fact that we observe both extreme TRD and no distortion in the progeny of females with A/J, C57BL/6J, 129S1/SvImJ, NOD/ShiLtJ, CAST/EiJ and PWK/PhJ alleles in trans at R2d2 (S10 Fig.), demonstrates that TRD is under genetic control of at least one additional locus (i.e., there is at least one unlinked distorter locus that is genetically variable in the CC and DO mice). Furthermore, the presence of at least two significantly different levels of distortion among F1 hybrid dams (Table 1; S3 Fig.) indicates either that more than one distorter locus is involved or that an allelic series exists at a single distorter locus.
Further evidence that TRD is under control of one or more unlinked distorters was provided by 15 female DO-G13–44 G1 offspring that inherited the high-copy allele. Those dams had significantly different levels of TRD (p = 9.8x10–5). Note that there was no correlation between the presence or level of TRD and the paternally inherited allele (one-way ANOVA, F = 2.21 on 1 and 23 df, p = 0.15; Fig. 3).
In the DO-G13–44 pedigree, females that inherited the R2d2WSBdel1 allele had copy number 11 (S11 Fig.), indicating a partial rather than complete deletion of the expansion. Using the TaqMan assay, we identified two additional DO females (DO-G13–49 and DO-G16–107; S4 Fig.; S11 Fig.) that had results consistent with a copy number loss in the WSB/EiJ haplotype. The presence of the deletion in the respective germlines was confirmed by the TaqMan assay in their progenies (S4 Fig.). Importantly, each one of the three deletions appears to be independent because these females are not closely related, their WSB/EiJ haplotypes in Chr 2 are different and the copy number present in each female is also different (S12 Fig.). The deletions appear to be internal to R2d2 based on the analysis of the MegaMUGA genotypes and intensities [57] at all surrounding markers. The repeated observation of independent deletions indicates that R2d2 is rather unstable and may explain the fact that, despite its presence in laboratory strains and wild mice, it has not led (yet) to a complete selective sweep.
Known meiotic drive systems (S1 Fig.) consist of one or more responder loci (a locus subject to preferential segregation during meiosis) and a single distorter (the effector locus required for drive at the responder). In meiotic drive systems that are stable in natural populations, responder and distorter loci are tightly linked and are typically protected from decoupling by factors that inhibit recombination, such as structural variation [7,11,14]. Although R2d2 resides within a recombination-cold region, the distorter is not closely linked to R2d2 based on the TR observed and the diplotypes present in F1 hybrid and DO dams (Fig. 1; S8 Fig.). Therefore, at least one unlinked distorter is required to explain the observed variability in TRD.
These observations indicate that the maternal TRD phenotype has a complex genetic architecture. Specifically, a minimum number of copies of R2d are required in heterozygosity at R2d2 for TRD to be observed. Therefore, it can be classified as overdominant, restricted to the female germline and caused by structural variation. Similar characteristics have been recently reported for the Xce locus that controls X-inactivation choice; notably, characterization of Xce relied on the analysis of a genetically diverse set of F1 hybrid mice [58]. In addition, multiple alleles at unlinked loci interact to determine whether distortion occurs at R2d2, and to what extent. This is unique among meiotic drive systems (S1 Fig.) and has important implications for the natural history of the system and for the ease of genetic dissection. We hypothesize that variation in TR levels at R2d2 results from the interaction of alleles originating from multiple taxa, and thus the use of inter-specifc and inter-subspecific mouse populations was key to the characterization of this system. Wild-derived strains and wild-caught mice have enabled important biological discoveries [4,59], and we echo previous encouragements of a more prominent role for these resources in biological and biomedical research [60,61].
Centromeres (i.e., the site of kinetochore formation) are remarkable loci that control, in cis, proper segregation of chromosomes during mitosis and meiosis. It is easy to envision how a responder at, or tightly linked to, a centromere can influence chromosome segregation. Recent evidence shows that kinetochore protein levels and microtubule binding are positively correlated with preferential segregation to the oocyte in mice that are heterozygous for Robertsonian fusions [62], indicating that differences in centromere “strength” lead to meiotic drive. Responders located far away from centromeres are thought to influence their own segregation in cis by becoming “neocentromeres” and taking advantage of the inherited functional polarity of the female meiotic spindle [63]. We hypothesize that R2d2 may act as a neocentromere after epigenetic activation mediated by C57BL/6J, NZO/ShiLtJ, 129S1/SvImJ, and NOD/HILtJ alleles at the distorter(s).
The discovery of multiple R2d2 alleles with different copy numbers demonstrates that the presence of the distal insertion of R2d is not sufficient for meiotic drive; rather, some minimum copy number (> 11) is required for TRD. This raises the possibility that meiotic drive at R2d2 is dosage-dependent, such that fine-scale control over the level of TRD is possible by adjusting the number of copies of R2d. If R2d2 is acting as a neocentromere, this may also indicate that some minimum size and/or number of repeats is required for recognition and activation by the epigenetic machinery. The Ab10 system of maize provides examples of responders that function as neocentromeres and for which the level of meiotic drive depends on the size of the responder (i.e., knob size) [11].
The effect on the Chr 2 centromere of activating an ectopic neocentromere at R2d2 is unknown, but it might explain the moderate levels of lethality caused by aneuploidy and suggests that some coordination between the two loci is required to achieve chromosome segregation. Meiosis involving chromosomes with neocentromeres may lead to an increased rate of non-disjunction and a reduced rate of recombination.
The conclusion that a genetically complex meiotic drive system is responsible for TRD favoring the WSB/EiJ allele at R2d2 is fully consistent with the initial observations of TRD in the CC, with our prediction that positive selection of the WSB/EiJ allele occurred during outcrossing or in early inbreeding generations [35], with the presence of similar levels of TRD in extinct and extant CC lines at intermediate generations of the CC (S5 Table) and with the fact that C57BL/6J, 129S1/SvImJ, NOD/ShiLtJ and NZO/HILtJ haplotypes at R2d2 are not underrepresented among the currently completed CC strains (http://csbio.unc.edu/CCstatus/index.py). The observed levels of TRD in crosses that use DO females are consistent with presence of different alleles at the distorter(s) (S7 Fig.; S1 Table).
Although the discovery and identification of TRD that emerged from the DO pseudo-randomized mating scheme offered the opportunity to characterize a novel meiotic drive responder, the existence of such a locus could negatively impact the utility of this population for genetic studies. Fortunately, the locus was discovered before complete fixation of the R2d2WSB allele. Although the candidate interval spans 900 kb, TRD affects a much larger region in the DO because the strength of selection in favor of the WSB/EiJ allele is outpacing the rate at which recombination can degrade linkage disequilibrium in the region. Ultimately, this region would become an actual or statistical ‘blind-spot’ in the DO, such that the non-WSB/EiJ allele frequencies would become too small to detect allelic effects on phenotypic variation. Efforts are underway to purge the WSB/EiJ allele from the DO breeding population at this locus or to select for mice carrying a WSB/EiJ haplotype with a low copy number for R2d2, rather than allow the region to become fixed. Using marker-assisted selection, progeny of heterozygous WSB/EiJ carrier crosses are excluded from subsequent generations. Allele frequencies and random segregation on all other chromosomes are being preserved (EJC unpublished).
The SPRET/EiJ and WSB/EiJ strains and the Hsd:ICR outbred stocks are among the most extensively characterized and utilized mouse populations. Resources involving those populations include whole-genome sequencing and genotyping [24,37], development of linkage maps of the mouse [40,64,65], creation of genetic reference populations [35,36,66], experimental crosses to map a diverse collection of biomedical and evolutionary traits [33,48,53,61] and selection lines derived from Hsd:ICR (such as M16i and HR) that have been widely used for genetic analyses [30–32,67–69]. The potential for distorted allele frequencies in crosses involving those populations may affect the interpretation of results from a wide range of genetic, behavioral and physiological studies.
The R2d2 system has attributes that make its genetic and mechanistic characterization a tractable problem. Identification of several distorters would allow assembling the pathway(s) responsible for centromere function and spindle polarity. This may open the way to explore at the molecular and mechanistic levels an evolutionary force (meiotic drive) thought to be responsible for karyotype evolution in mammals and in many other organisms [15]. With the advent of genome engineering tools such as CRISPR/Cas9 [70], we also anticipate practical applications of a strong, modulable meiotic drive system with only modest levels of lethality. For example, meiotic drive could be used to increase the efficacy of gene drives for introducing new genes into experimental or natural populations [71].
All animal work was performed according to one of the following protocols: 1) the Guide for the Care and Use of Laboratory Animals under approved IACUC animal use protocols within the AAALAC accredited program at the University of North Carolina at Chapel Hill (Animal Welfare Assurance Number: A-3410–01); 2) the requirements of The Jackson Laboratory Animal Ethics Committees under approved protocol #JAX10001; 3) an animal protocol approved by the North Carolina State University Institutional Animal Care and Use Committee (09–0133-B); or 4) an animal study protocol approved by the NCI Animal Care and Use Committee (ASP# LCBG-013). All animals were euthanized according to the regulations of the governing protocol.
The G2:F1 population has been previously reported and was genotyped on the Mouse Diversity Array [72] (MDA). A population of 96 (FVB/NJx(WSB/EiJxPWK/PhJ)F1)G2 mice was previously reported and was genotyped on the MegaMUGA array [40,53]. DNAs from selected progeny from previously published (C57BL/6JxSPRET/EiJ)xC57BL/6J and (A/JxSPRET/EiJ)xA/J backcrosses [26,43] were regenotyped on the MegaMUGA array. The SPRET/EiJ strain designation had not yet been assigned to the inbred strain at the time the backcross was performed [26]. Finally, DNA from multiple samples from the (M16ixL6)F2 intercrosses and from generations 4 and 10 of the (HR8xC57BL/6J) advanced intercross line [29,31,32] were genotyped at markers closely linked to R2d2.
Crosses 1–2, 7–10 and 16–17 (Table 1). WSB/EiJ and C57BL/6J were used in reciprocal combinations. Male F1 hybrids were backcrossed to C57BL/6J to produce the progeny of crosses 1 and 2. Female F1 hybrids were backcrossed to C57BL/6J to produce the progeny of crosses 16 and 17. The progeny of crosses 7–10 was produced in a similar way to crosses 16 and 17, except that female F1 of reciprocal matings of WSB/EiJ and CAST/EiJ were used for crosses 7 and 8, and female F1 of reciprocal matings of WSB/EiJ and PWD/PhJ were used for crosses 9 and 10. All breeding was done at the Jackson Laboratory (Bar Harbor, ME).
All other crosses. DO mice and standard mouse inbred strains (129S1/SvImJ, A/J, C57BL/6J, CAST/EiJ, FVB/NJ, NU/J, NOD/ShiLtJ, NZO/H1LtJ, PWK/PhJ and WSB/EiJ) were obtained from The Jackson Laboratory (Bar Harbor, ME). CC mice were obtained from the Systems Genetics Core Facility colony at UNC Chapel Hill [73] (http://csbio.unc.edu/CCstatus/index.py). Those mice were used to generate the following number and types of hybrid mice: nine (129S1/SvImJxWSB/EiJ)F1 females; two (A/JxWSB/EiJ)F1 females; seven (NOD/ShiLtJxWSB/EiJ)F1 females; six (NZO/HILtJxWSB/EiJ)F1 females; 10 (CC042/GeniUncxCC001/Unc)F1 females; three (CC001/UncxCC039/Unc)F1; nine (DOxCC001/Unc)F1 females, 13 (DOxCC005/Tau Unc)F1 females and five (NU/JxC57BL/6J). F1 females were mated to FVB/NJ males and cages were surveyed three to five times per week. Litter sizes were recorded and pups were sacrificed at birth, and tissue was collected for DNA isolation. The same breeding schema was followed with 127 DO R2d heterozygous females used to determine the origin of maternal TRD. All breeding was done at UNC Chapel Hill (Chapel Hill, NC).
A single G13 DO female (DO-G13–44) was mated to a male that was the result of an intercross between four CC lines (CC013/GeniUnc, CC053/Unc, CC065/Unc and CC008Geni/Unc; Fig. 4). G3 female progeny were weaned, single housed and mated to FVB/NJ males. Cages were surveyed three to five times per week. Litter sizes were recorded and G4 pups were sacrificed at birth, and tissue was collected for DNA isolation.
TR was measured in G3 dams as described above. Each dam was classified as having TRD (p < 0.05 for 1-df Χ2 test of null hypothesis TR = 0.5) or not having TRD (p ≥ 0.05). Both G2 parents and G3 dams were genotyped on MegaMUGA and phased haplotypes at R2d2 were inferred by manual inspection of haplotype reconstructions. In order to isolate the contribution of maternal and paternal alleles to TRD, MegaMUGA markers called as H in the G2 dam and homozygous in the G2 sire were retained for mapping, and presence of TRD was mapped as a binary phenotype using a logistic regression analog to the Haley-Knott method. The procedure was repeated using only markers called as H in the father of the G3 dams and homozygous in the mother. Significance thresholds for LOD scores were obtained by unrestricted permutation.
Crosses 1–2, 7–10 and 16–17 (Table 1). DNA was prepared from spleens of 21-day old mice. DNA extraction and SNP genotyping were carried out as described previously [74].
All other samples. DNA for PCR-based genotyping was performed on crude whole genomic DNA extracted by heating tissue in 100ul of 25mM NaOH/0.2mM EDTA at 95°C for 60 minutes followed by the addition of 100ul of 40mM Tris-HCl. The samples were then spun at 2000 rpm for 10 minutes and the supernatant collected for use as PCR template. All primers (S6 Table) used in this study were designed using PrimerQuest software (https://www.idtdna.com/Primerquest). PCR reactions contained 1.5–2 mM MgCl2, 0.2–0.25 mM dNTPs, 0.2–1.8 μM of each primer and 0.5–1 units of GoTaq polymerase (Promega) in a final volume of 10–50 μL. Cycling conditions were 95°C, 2 min, 35 cycles at 95°, 55° and 72°C for 30 sec each, with a final extension at 72°C, 7 min. PCR products were loaded into a 2% agarose gel and run at 200 V for 40–120 minutes (depending on the marker). Genotypes were scored and recorded.
DNA for MegaMUGA genotyping was isolated as described previously [40,53]. Briefly, ~2 mm of mouse tail (5 mg) was harvested, flash-frozen on dry ice and digested with proteinase K overnight at 65°C. The following day, DNA was extracted using the QIAGEN Puregene Gentra kit (kit no. 158389; QIAGEN GmbH, Hilden Germany). Genotyping was performed with the MegaMUGA genotyping microarray (Neogen/GeneSeek, Lincoln, NE), a 78,000-probe array based on the Illumina Infinium platform.
Genotyping by TaqMan. After R2d2 was established as the causal variant for TRD, a subset of DO-G16 progeny and all (M16i x L6)F2 intercross progeny were genotyped using TaqMan real-time PCR assays for Cwc22. Samples heterozygous for a high-copy allele at R2d2 can be readily distinguished from samples homozygous for a low-copy allele based on the normalized cycle threshold value estimated from the assay (see section “Copy-number validation” below).
Deviation from Mendelian transmission. TR is reported as the ratio of the WSB/EiJ genotype to the total number of genotypes: WSB / (WSB + nonWSB). P values for aggregate data were calculated using a Χ2 goodness-of-fit test of the observed number of WSB/EiJ genotypes compared to the number of WSB/EiJ genotypes expected under the null hypothesis of equal transmission:
X2=(WSB-WSB+nonWSB2)2WSB+nonWSB2
For individual dams, the small sample sizes (typically fewer than 50 total offspring) would lead to type II error; therefore, p-values were calculated using an exact binomial test. Confidence intervals for TRs were calculated using the binom R package (http://cran.r-project.org/web/packages/binom/).
Average litter size. Average litter size was calculated as the mean number of offspring counted soon after birth per litter per dam (± standard deviation), including the number of viable embryos counted in utero in mid-gestation DO dams (unless otherwise noted).
The expected average litter size (ALS) of a dam under a model in which lethality is the sole explanation for TRD is:
ALSObs=ALSExp(1-2TR-12TR)
,
where ALSExp is the mean ALS in dams with no TRD [41]. Significance of the deviation of ALSObs from ALSExp was determined using a Wilcox signed rank test.
Inheritance of R2d2 alleles. Similarly, the average absolute number of offspring inheriting each R2d2 allele was calculated as the mean number of offspring per litter per dam having each of the possible genotypes. Significance was determined using a one-tailed Student t-test.
DO and F1 dams were euthanized by CO2 asphyxiation 12–18 days after delivery of the previous litter and the uterus was dissected. The number of live embryos and reabsorbed (dead) embryos was recorded. Each live embryo was dissected to isolate DNA for genotyping. Tissue from each live embryo was harvested for DNA extraction and genotyping.
All MDA arrays were genotyped using MouseDivGeno [57], and all MegaMUGA arrays were genotyped using Illumina BeadStudio. We plotted number of H and N calls (as a fraction of the total number of genotypes) for each group of similar samples and excluded outliers from further analysis. For CC lines, DO animals, CCxCC F1 females and DOxCC F1 females, we inferred haplotypes using probabilistic methods [40,75]. As an additional QC step, we grouped DO samples by generation and plotted the number of recombinations (counted as unique transitions in haplotype reconstructions) and removed outliers.
CAST/EiJ allele in the CC G2:F1. Thirty-four MDA SNP probe sets were identified within R2d in the GRCm38 reference sequence (S3 Table). We ensured that these probes were unique using BLAT [76] to map them to the reference genome. In order to map the expansion allele present in the CAST/EiJ strain, phenotypes and genotypes were coded as follows. First, we applied a CCS transform [77] to the mean intensity of all probes in each probe set using MouseDivGeno [57] and summed the values for each sample to obtain the final phenotype value. Next, the genome was divided into a set of disjoint intervals whose boundaries were defined by the 21,933 unique recombination events inferred in the population [40], so that no individual would be recombinant within any of the resulting intervals. Then, using haplotype reconstructions, individuals were coded as either heterozygous (CAST/not-CAST) or homozygous (not-CAST/not-CAST) within each interval (there are no CAST homozygous individuals in this population). Of 474 individuals, 144 with a WSB/EiJ allele in the middle of chromosome 2 were excluded to yield a final sample size of 330. A single-locus QTL scan was then performed via Haley-Knott regression [78], treating the population as a backcross.
WSB/EiJ allele in an intercross population. Three MegaMUGA SNP probes were identified within R2d in the GRCm38 reference (S4 Table). Again, uniqueness was verified using BLAT. In order to map the expansion allele in WSB/EiJ, the sum intensity of these probes was used as a phenotype and genotypes were coded as follows. First the genome was divided into a grid of 1,000 disjoint intervals of approximately equal size, and one MegaMUGA SNP marker segregating between WSB/EiJ and PWK/PhJ was selected per interval. Individuals were coded as heterozygous (WSB/not-WSB) or homozygous (not-WSB/not-WSB) at each marker. A single-locus QTL scan was then performed using Haley-Knott regression as implemented in R/qtl [79], treating the population as a backcross.
In order to refine the location of R2d2, we identified individual mice with recombinant chromosomes within the candidate interval defined by linkage mapping. These critical recombinants define the proximal and distal boundaries of the refined candidate interval.
CAST/EiJ allele. We partitioned the 330 G2:F1 individuals without a WSB/EiJ allele in the R2d locus into two groups according to MDA sum-intensity values. From those with sum-intensity consistent with a non-CAST/EiJ expansion allele, we selected the most distal recombinants from CAST/EiJ to another haplotype. From those with sum-intensity consistent with the CAST/EiJ expansion allele, we selected the most distal recombinant from another haplotype to CAST/EiJ. Together these recombinants define the proximal boundary of the candidate interval in CAST/EiJ. Similarly, in order to define the distal boundary of the candidate interval, we selected the most proximal recombinants from CAST/EiJ to another haplotype that still had sum-intensity consistent with the CAST/EiJ expansion allele.
WSB/EiJ allele. The boundaries of the WSB/EiJ candidate interval were mapped in the same fashion using 229 individuals spanning generations 10 through 14 of the DO, all of which have been genotyped on MegaMUGA and are recombinant for WSB/EiJ in the initial candidate interval. We first excluded individuals homozygous for WSB/EiJ over any interval with in the interval. Then we selected the most distal recombinants from another haplotype to WSB/EiJ, which also had MegaMUGA sum-intensity values consistent with a non-WSB/EiJ expansion allele. These recombinants define the distal boundary of the candidate interval. We mapped the proximal boundary similarly.
SPRET/EiJ allele. (C57BL/6JxSPRET/EiJ)xC57BL/6J (n = 12) and (A/JxSPRET/EiJ)xA/J progeny (n = 17) [26,43] genotyped on the MegaMUGA array were used to refine the candidate interval for the expansion allele in SPRET/EiJ. Haplotypes in the relevant region of Chr 2 were inferred by manual inspection of genotype calls. Samples were partitioned according to sum-intensity at the three MegaMUGA SNP probes tracking the expansion allele. Among individuals with sum-intensity consistent with the expansion allele, the most proximal recombinant from SPRET/EiJ to another haplotype defines the distal boundary of the candidate interval. Likewise the most distal recombinant from a non-SPRET/EiJ haplotype to SPRET/EiJ defines the proximal boundary of the candidate interval.
Ten individuals from the HR8 selection line were selected for whole-genome sequencing. Five micrograms of high-molecular-weight DNA were used to construct TruSeq Illumina libraries, using 0.5 μg starting material, with 300- to 400- and 400- to 500-bp fragment sizes. Each library was sequenced on one lane of an Illumina HiSeq2000 flowcell, as paired-end reads, with 100-bp read lengths. We aligned the sequences to the University of California at Santa Cruz Mouse Build mm9. HR8 sequenced reads were aligned to the mouse genome (mm9) using bowtie 2.2.3 [80] with default options. We removed PCR duplicates and filtered low-quality SNPs using samtools 0.1.19 [81] and Picard 1.88 (http://picard.sourceforge.net/).
We retrieved BAM files of aligned reads (Oct 2012 release) from the Sanger Mouse Genomes Project FTP site (ftp-mouse.sanger.ac.uk). We used the mpileup function of samtools [81] to call sequence variants on the HR8 and Sanger BAM files jointly and to output the read depth at each base. We counted a SNP as private to WSB/EiJ, SPRET/EiJ and the 10 HR8 individuals if those samples all shared a genotype that was different from the seven other CC founder strains. We defined the boundaries of the copy number expansion by identifying consecutive 100bp windows in which the average read depth was at least twice the genome-wide average read depth. We estimated the number of copies of the expansion as the modal per-base read depth.
We used commercially-available TaqMan assays for Cwc22 to estimate the copy number of R2d2. We used two copy number assays (Life Technologies catalog numbers Mm00644079_cn, Mm00053048_cn) to target the number of Cwc22 copies (proximal and distal). We also used two reference assays (Tfrc, cat. no. 4458366, for target Mm00053048_cn; Tert, cat. no. 4458368, for target Mm00644079_cn), for genes known to exist in a single haploid copy in the mouse, to calibrate the amplification curve. Assays were performed according to the manufacturer’s protocol on an ABI StepOne Plus Real-Time PCR System (Life Technologies, Carlsbad, CA). Cycle thresholds (Ct) for each assay were determined using the ABI CopyCaller v2.0 software with default settings. For each target-reference pair, relative cycle threshold (ΔCt) was calculated as
The ΔCt value is proportional to copy-number of the target gene on the log scale but is subject to batch effects. In order to account such effects, normalized ΔCt values for each sample were calculated as follows. A standard set of control samples (from C57BL/6J, WSB/EiJ, CAST/EiJ and (WSB/EiJxC57BL/6J)F1 mice), spanning the expected copy-number range for Cwc22, were included in duplicate or triplicate in every assay batch. A linear mixed model was fit to raw ΔCt values for these control samples, with target-reference pair and batch as random effects, using the lme4 package (http://lme4.r-forge.r-project.org/) for R (http://www.R-project.org/). Predicted values (best linear unbiased predictors, BLUPs) from this model capture technical variation orthogonal to variation due to genotype. BLUPs calculated from control samples were subtracted from raw ΔCt values for all samples, and the residual was used as the normalized ΔCt for copy-number estimation.
In this manuscript we chose in most cases to present ΔCt, rather than extrapolated absolute copy number, because ΔCt is the natural scale of the data (i.e., the log scale). Constant variance (with respect to mean) on the log scale grows exponentially on the linear scale so that estimates of absolute copy number become increasingly uncertain as copy-number grows.
The use of TaqMan assays for Cwc22 as a proxy for copy number at R2d2 was validated by mapping normalized ΔCt for target Mm00644079_cn as a quantitative phenotype in 64 members of the (FVB/NJx(WSB/EiJxPWK/PhJ)F1)G2 intercross population described above. The marker selection and mapping procedure were the same as described above for mapping MegaMUGA sum-intensity values.
Chr 2 genotypes and whole-genome sequence that have not been published elsewhere are available at http://csbio.unc.edu/r2d2.
|
10.1371/journal.ppat.1001285 | Atypical/Nor98 Scrapie Infectivity in Sheep Peripheral Tissues | Atypical/Nor98 scrapie was first identified in 1998 in Norway. It is now considered as a worldwide disease of small ruminants and currently represents a significant part of the detected transmissible spongiform encephalopathies (TSE) cases in Europe. Atypical/Nor98 scrapie cases were reported in ARR/ARR sheep, which are highly resistant to BSE and other small ruminants TSE agents. The biology and pathogenesis of the Atypical/Nor98 scrapie agent in its natural host is still poorly understood. However, based on the absence of detectable abnormal PrP in peripheral tissues of affected individuals, human and animal exposure risk to this specific TSE agent has been considered low. In this study we demonstrate that infectivity can accumulate, even if no abnormal PrP is detectable, in lymphoid tissues, nerves, and muscles from natural and/or experimental Atypical/Nor98 scrapie cases. Evidence is provided that, in comparison to other TSE agents, samples containing Atypical/Nor98 scrapie infectivity could remain PrPSc negative. This feature will impact detection of Atypical/Nor98 scrapie cases in the field, and highlights the need to review current evaluations of the disease prevalence and potential transmissibility. Finally, an estimate is made of the infectivity loads accumulating in peripheral tissues in both Atypical/Nor98 and classical scrapie cases that currently enter the food chain. The results obtained indicate that dietary exposure risk to small ruminants TSE agents may be higher than commonly believed.
| Following the bovine spongiform encephalopathy (BSE) crisis and the identification of its zoonotic properties, a sanitary policy has been implemented based on both eradication of transmissible spongiform encephalopathies (TSE) in food-producing animals and exclusion of known infectious materials from the food chain. Atypical/Nor98 scrapie is a prion disease of small ruminants identified worldwide. Currently it represents a significant part of the TSE cases detected in Europe. The restricted tissue distribution of Atypical/Nor98 scrapie agent in its natural host and the low detected prevalence of secondary cases in affected flocks meant that it is believed to be a poorly transmissible disease. This has led to the view that Atypical/Nor98 scrapie is a spontaneous disorder for which human and animal exposure risk remains low. In this study we demonstrate that in affected individuals, Atypical/Nor98 scrapie agent can disseminate in lymphoid tissues, nerves, and muscles, challenging the idea that it is a brain-restricted infectious agent. Evidence for the deficiencies in the current methods applied for monitoring Atypical/Nor98 scrapie is provided that would indicate an underestimation in the prevalence in the general population and in the affected flocks. These elements challenge the hypothesis on the biology of this recently identified TSE agent.
| Transmissible spongiform encephalopathies (TSE), or prion diseases, are fatal neurodegenerative disorders occurring in sheep (scrapie), cattle (bovine spongiform encephalopathy - BSE), or humans (Creutzfeldt-Jakob disease - CJD). The key event in TSE is the conversion of a normal cellular protein (PrPc) into an abnormal isoform (PrPSc) which accumulates in tissues from infected individuals [1]. PrPSc is currently considered to be the only TSE biochemical marker. According to the prion concept, abnormal PrP would be the causative agent of TSE [2].
Following the BSE crisis and the identification of its zoonotic properties [3], [4], the control of human and animal exposure to TSE agents has become a priority. A sanitary policy has been implemented based on both eradication of TSE in food producing animals and exclusion of known infectious materials from the food chain.
In 1998 an Atypical/Nor98 Scrapie was identified in Norwegian sheep; the PrPSc signature was partially PK resistant and displayed a multi-band pattern as showed by Western Blot (WB) that contrasted with those normally observed in small ruminants TSE cases [5]. After 2001 and the implementation of active TSE surveillance plans, a number of similar cases were identified in most EU members states as well in other countries, like Canada, USA and New Zealand [6]. The transmissibility of Atypical/Nor98 agent has been demonstrated in both rodent models (transgenic animals expressing the ovine Prnp gene) [7] and sheep [8], [9].
Currently Atypical/Nor98 Scrapie represents a significant part of the TSE cases identified in the EU small ruminant population, where its prevalence was estimated to range between 5 to 8 positive small ruminants per 10,000 tested per year [10].
Atypical/Nor98 scrapie cases have different biological features from those observed in other small ruminants TSE [5], [6]. Sheep susceptibility to TSE is strongly controlled by polymorphisms on the gene (Prnp) encoding for PrP protein [11], [12]. The homozygous and heterozygous ARR sheep are considered to be strongly resistant to both the classical scrapie [11], [12] and the cattle BSE agents [13]. This resistance has been the basis of a large scale genetic selection policy aiming at the control of TSE diseases by increasing the frequency of the ARR allele in general population and restocking affected flocks with ARR animals. In Atypical/Nor98 scrapie the sheep genetic susceptibility is significantly different from what is observed in classical TSE forms, with homozygous and heterozygous ARR allele carriers being susceptible to the disease [14], [15], [16], [17], [18].
Information about the tissue distribution of Atypical/Nor98 scrapie agent in the host species is limited [5], [19], [20], [21] but research findings indicate that no detectable abnormal PrP has been found in peripheral tissues and that the infectious agent could be restricted to the central nervous system. This key feature led to consider that dietary exposure risk to Atypical/Nor98 scrapie is low. The apparent limited spreading of Atypical/Nor98 scrapie in the organism of affected individuals is also an argument supporting the hypothesis that this agent has restricted abilities to spread into the environment or between individuals. PrPSc detection generally correlates with the presence of infectivity [1], [22] but infectivity has been reported in the absence of detectable PK resistant PrP [23]. In this study, we investigate the potential presence of TSE infectivity in peripheral tissues (lymphoid organs, striated muscles and nerves) of Atypical/Nor98 scrapie from naturally and experimentally infected sheep. We then compared the relative infectivity level present in peripheral tissues with those estimated in similar tissues from classical scrapie affected sheep.
Atypical/Nor98 scrapie field cases (n = 7) collected in three different countries (Portugal, Norway and France) were investigated for the presence of PrPSc and infectivity in lymphoid tissues and central nervous system (Table 1). These sheep were of various Prnp genotypes including those associated with high susceptibility to Atypical/Nor98 scrapie (homozygous or heterozygous A136F141R154Q171– AHQ) or resistance (homozygous and heterozygous ARR) to classical scrapie or BSE [11], [13]. Amongst the seven cases, five were identified by the active surveillance program either at rendering plant or slaughter house and two through the passive surveillance network (clinical suspects). Three of these cases were identified in fallen stock animals collected in three independent flocks where an Atypical/Nor98 scrapie case had previously been identified (Portugal).
In each of these natural Atypical/Nor98 cases, PrPSc accumulation could be detected in different brain areas by WB and/or Immunohistochemistry (IHC). Conversely, no abnormal PrP deposits were evidenced in any of the investigated lymphoid organs (Table 1).
Bioassay of brain homogenates prepared from these seven sheep into transgenic mice that over-express the VRQ allele of ovine PrP (tg338) [7] were positive for TSE. Surprisingly, despite the absence of detectable PrPSc, the lymphoid tissue homogenates from five out of the seven cases were positive for TSE in tg338 mice. The attack rate in mice challenged with lymphoid tissues was lower and the clinical onset was delayed compared to mice inoculated using CNS homogenate (Table 1). Brains collected in clinically affected mice inoculated either with lymphoid tissues or brain homogenates displayed a similar PrPSc WB pattern (Figure 1: lanes 5, 6, 7 – Figure S2), PrPSc deposits distribution and vacuolar lesion profile (Figure 2A, 2C, 2E- – Figure S2). All these features were identical to those previously reported in tg338 mice inoculated with a panel of Norwegian, French [7] and UK [24] Atypical/Nor98 scrapie isolates. These phenotypic features were clearly different from those associated to two distinct classical scrapie isolates (Figure 1: lanes 1–4 and Figure 2B, 2F, 2H). Infectivity was demonstrated in lymphoid tissues from one out of the two investigated ARR/ARR Atypical/Nor98 scrapie cases.
Samples collected in two experimental Atypical/Nor98 scrapie cases were also analyzed; these were from an AHQ/AHQ (case 9) and an AFRQ/ARQ (case 8) sheep which had been intra-cerebrally challenged with an AFRQ/AFRQ field Atypical/Nor98 scrapie isolate (case 1, Figure 1: lane 5). These sheep had an incubation period of 964 and 2240 days respectively. In both animals' CNS, PrPSc displayed a WB banding pattern that was characteristic [14] of Atypical/Nor98 scrapie (Figure 1: lane 6 – Figure S2). In none of the investigated peripheral tissues PrPSc could be detected (Table 2). In the tg338 bioassay, infectivity was shown in some but not all tested lymphoid tissues, and in striated muscle and peripheral nerves from both cases (Table 2 – Figure 1: lanes 8–9 – Figure S2). Incubation periods were prolonged in comparison to brain homogenates and the attack rate was lower than 100%. The phenotypes of the propagated prions in tg338 mice (WB banding pattern, vacuolar lesion profile and PET Blot PrPSc distribution in brain) were identical to the one observed with natural Atypical/Nor98 scrapie isolates (Figure 1: lanes 5, 7, 9- Figure 2C, 2E, 2G).
A similar experiment was performed using peripheral and CNS tissues from natural or experimental classical scrapie cases at clinical stage of the disease. This study involved two distinct classical scrapie agents (Langlade and PG127), which can be distinguished on the basis of their lesion profile in tg338 mice (Figure 2F). The inoculation of peripheral tissues homogenates from animals infected with those classical scrapie agents into tg338 resulted in a 100% attack rate transmission, but with prolonged incubation period by comparison to mice inoculated with CNS samples (Table 2). For both isolates, PrPSc WB banding pattern (Figure 1: lanes 1–4 – Figure S2), vacuolar lesion (Figure 3F, 3H – Figure S2) observed in mice inoculated with CNS and peripheral tissues were identical. As previously described and conversely to Atypical/Nor98 scrapie, PrPSc could be detected in the investigated lymphoid organs, and striated muscle [25], [26], [27], [28] from all the four classical scrapie affected animals involved in this study (Table 2).
In order to determine the cause of our incapacity to detect abnormal PrP in the peripheral tissues of Atypical/Nor98 scrapie cases that contain infectivity, classical scrapie (cases 10 and 12) and Atypical/Nor98 scrapie (cases 1, 8, 9) brain homogenates dilution series were prepared and processed for a OIE registered PrPSc detection WB (TeSeE WB Kit – BIORAD), a PrPSc ELISA detection assay (TeSeE Sheep and Goat - BIORAD) (Figure 3 and Figure S1) and bioassays in tg338 mice (Table 3).
According to the endpoint titration, the infectious titre in the Langlade and PG127 classical scrapie isolates were estimated to be respectively 106.8 ID50 IC tg338 per gram and 106.6 ID50 IC tg338 per gram (Table 3- Figure 4A). All the titrated Atypical/Nor98 scrapie cases (n = 5, see Table 3) that we investigated displayed substantially higher infectious titres, ranging between 108.7 and 109.5 ID50 IC tg338/g (Table 3-Figure 4 B).
For the two classical scrapie isolates, WB detected PrPSc to a dilution of 103 (positive detection on 25 µg of brain equivalent material) which corresponded to 102.2 (Langlade isolate) and 102 (PG127 isolate) ID50 IC tg338 (Figure 3). For the AHQ/AHQ Atypical/Nor98 scrapie isolate (case 9), PrPSc WB detection limit was the 1/80 dilution (312 µg of brain starting material) (Figure 3A) which corresponded to 106 IC ID50 in tg338 (Figure 4E). Similar results were obtained with the two other atypical scrapie isolates (cases 1 and 8) for which the detection limits of PrPSc assays were respectively equivalent to 105.6 and 105.5 ID50 IC tg338 (Table 3 and Figure S1). These results indicate that PrPSc detection assays currently used for field TSE testing could have a dramatically lower intrinsic sensitivity for identifying Atypical/Nor98 scrapie agent than classical scrapie agent.
As previously described [29], [30], [31], [32], [33], [34], the end point infectivity titration data that was generated with the two different types of classical scrapie agents (case 10 and case 12) and the Atypical/Nor98 scrapie cases (cases 14 and 15) were used to fit the best logistic regression models correlating the incubation periods in tg338 mice with the infectious dose (Figure 4). These models were then applied to estimate the infectious content in the different tissues using the incubation periods in tg338 mice (Tables 1 and 2).
By this approach, the infectious titre of three atypical scrapie samples (cases 1: cerebral cortex - case 8: cerebellum – case 9: cerebral cortex) were estimated respectively 108.7, 108.7, and 108.3 ID50 IC tg338/gram (Tables 1 and 2). The measured infectious titre in the same three samples, by endpoint titration in tg338, were respectively 109.1, 108.7 and 10 9.5 ID50 IC tg338/gram (Table 3).
In the investigated Atypical/Nor98 scrapie cases, the infectious load in muscle and lymphoid tissues samples from sheep affected were close to the sensitivity of our bioassay (102,7 ID50 IC per gram of tissue); ie about 106 fold lower than the infectivity level measured (by endpoint titration) or estimated (on the basis of the incubation periods) in the same amount of brain prepared from clinically affected sheep (Tables 1 and 2).
In sheep affected by the two different classical scrapie agents the incubation period recorded in tg338 inoculated with lymphoid tissues and striated muscle were consistent with infectious titre about 10 fold lower than the one measured in brains from those terminally affected animals (Table 2).
Bioassay endpoint titration is considered as the most accurate method for determining the TSE infectivity titre in tissues. Although regarded as less accurate, dose-response relationships have been used as a method for infectivity estimation when endpoint titration data are not available; in such an approach, the incubation period observed in the inoculated mice is used to estimate the infectious titre of the samples tested [29], [30], [31], [32], [33], [34]. In this study, the dose-response approach was used to estimate the infectious titre in various peripheral tissues from Atypical scrapie/Nor98 and classical scrapie affected sheep. For mice inoculated with peripheral tissue homogenates, the standard curve established using reference CNS homogenates was used, but it was established that the lesion profile was identical in the mice inoculated with the peripheral tissue and the CNS. Although it could be hypothesized that the nature of the peripheral tissue inoculated would impact on the observed incubation length in mice (matrix effect) and consequently on the estimated infectious titre, Dickinson et al. [29], [30] demonstrated that in conventional mice the dose-response obtained with the spleen and the brain from ME7 infected mice are similar.
Together these elements indicate, that even if the peripheral tissues infectious titre reported are ‘estimates’, they provide a good guide to the relative infectivity levels that are present in the Atypical scrapie/Nor98 and classical scrapie cases' brain and peripheral tissues.
The presence of PrPSc and infectivity in small ruminant's peripheral tissues affected with natural classical scrapie or experimental BSE is well established [25], [27], [35], [36], [37]. It is generally considered that peripheral tissues like lymphoid tissues and striated muscle contain much lower levels of prion than CNS from terminally affected animals. This concept is the basis of the statutory measures aiming at limiting the entry of small ruminants TSE agents into the food chain. Tissues considered to be the most infectious (named Specific Risk Material) are systematically discarded from consumption, but tissues that would potentially contain only a low level of infectivity might enter the food chain due to the feasibility/practicality of removing them.
In this study, the estimated infectivity level in skeletal muscle and lymphoid tissues from animals (n = 4) affected with two different classical scrapie isolates did reach up to 1/10 (weight/weight) of the infectivity found in the CNS from terminally affected sheep. These values are higher than those expected from previous work. This could be explained by the fact that previously available data on prion quantities in peripheral tissues of small ruminants (in particular those related to striated muscle) relied on biochemical measurement of PrPSc amount [26] and the cell types accumulating PrPSc and the composition of these tissues may have impact on the PrPSc recovery yield. Also, if in some classical scrapie cases a 3–4 log10 infectivity difference was reported between CNS and some lymphoid tissues using bioassay in conventional mice, in other classical scrapie cases, the same study reported that infectivity in lymphoid tissue was only 1 to 10 fold lower than in CNS [27].
The classical scrapie cases that were investigated in this work cannot be assumed to be representative of all field diversity as only four animal cases of highly susceptible genotypes were used. However, the results indicate that exposure risk to such TSE agents through the unrestricted entry in the food chain of potentially infectious tissues would be significantly higher than previously thought.
In most countries, the identification of Atypical/Nor98 scrapie was a consequence of the implementation of an active surveillance for TSE consisting in random testing for PrPSc presence in brainstem of a fraction of fallen or healthy culled small ruminants [10]. In Atypical/Nor98 scrapie cases, the sensitivity of PrPSc detection tests that are used for initial field screening or confirmation of TSE cases is debated. Several authors reported failure to detect PrPSc in some CNS areas like the obex area [5], [6], [20] from known affected animals or discrepancies in results when applying different diagnostic tests to a same sample [6], [10].
The results obtained in this study by comparing the analytical sensitivity of biochemical PrPSc detection (using an OIE registered WB method and a validated rapid screening test for TSE detection, in small ruminants) and bioassay indicated that CNS samples that would contain up to 107.4.–107.7 ID50/g of Atypical/Nor98 scrapie (according to tg338 IC bioassay) could remain negative for PrPSc detection. In field, Atypical/Nor98 scrapie cases (Table 1) PrPSc positive WB was observed in CNS samples in which infectious titre was estimated (on the basis of incubation period) to be higher than 105.8 ID50/g IC in tg338. Such discrepancies might reflect an individual variability of the PrPSc WB detection limits between atypical scrapie cases. It might alternatively be the consequence of a relative imprecision in estimating the titre of low infectious doses by the incubation period bioassay method.
In contrast to Atypical/Nor98 scrapie cases, using two different classical agents the WB PrPSc detection sensitivity limit was about 102 ID50 IC in tg338 (ie a tissue with a titre of 103.7 ID50/g IC in tg338). These differences strongly support the contention that diagnostic assays based on PrPSc detection have lower performance for identifying Atypical/Nor98 scrapie cases than classical scrapie cases. It is consequently highly probable that a significant number of Atypical/Nor98 cases remain undetected by field testing, leading to an underestimation of Atypical/Nor98 scrapie prevalence in the small ruminant population. It is however not possible on the sole basis of this study to evaluate the importance of such underestimation.
The under detection of Atypical/Nor98 scrapie in the field due to the sensitivity of the current PrPSc based approach would also impact on understanding of the biology of this TSE agent.
While under natural conditions, classical scrapie is known to transmit between individuals, the analysis of data collected through the active TSE surveillance program seemed to indicate that Atypical/Nor98 scrapie could be poorly or not transmissible at all. This is based on the lack of statistical difference of the observed Atypical/Nor98 frequencies between the general population and the flocks where a positive case had been identified [38], [39]. The lower ability to detect Atypical Scrapie incubating animals using the PrPSc based methodologies means that this conclusion should be considered with caution.
Atypical/Nor98 cases are identified in older animals in comparison to classical scrapie [6], [40]. The lack of PrPSc detection in peripheral tissues of reported cases suggested that Atypical/Nor98 scrapie agent could be restricted to CNS. This is supportive of the hypothesis that Atypical/Nor98 scrapie could be a spontaneous disorder of PrP folding and metabolism occurring in aged animals without external cause [6], [38].
However, this hypothesis is questioned by the evidence reported here that a negative PrPSc testing result could be observed in animals harbouring high infectious titre in their brain and that the infectious agent can be present in peripheral tissues of Atypical/Nor98 scrapie incubating sheep. TSE are considered to be transmitted following oral exposure; initial uptake is followed by a peripheral replication phase which is generally associated with a dissemination of the agent in the lymphoid system and the deposition of large amounts of PrPSc. This peripheral replication phase is later followed by the entry of the infectious agent into the CNS through the autonomic nervous system [25], [27], [35], [36]. However, in several situations, like BSE in cattle [41], [42], [43] or classical scrapie in ARR heterozygote sheep [44], [45], the involvement of secondary lymphoid system is marginal, which does not preclude central neuro-invasion through the autonomic nervous system [46]. It could be proposed that Atypical Scrapie/Nor98 might occur following oral exposure to a TSE agent, which would spread marginally in lymphoid tissues before neuro-invasion. The slow propagation of Atypical Scrapie/Nor98 in its host (long incubation period) and the impaired detection sensitivity level of PrPSc based assays would explain the apparent old age of detected cases.
The results presented here are insufficient to rule out the hypothesis of a spontaneous/non contagious disorder or to consider this alternative scenario as a plausible hypothesis. Indeed, the presence of Atypical scrapie/Nor98 infectivity in peripheral tissues could be alternatively due to the centripetal spreading of the agent from the CNS. However, our findings point out that further clarifications on Atypical/Nor98 scrapie agent biology are needed before accepting that this TSE is a spontaneous and non contagious disorder of small ruminants. Assessing Atypical/Nor98 scrapie transmissibility through oral route in natural host and presence in placenta and in colostrum/milk (which are considered as major sources for TSE transmission between small ruminants) [28], [32] will provide crucial data.
The presence of infectivity in peripheral tissues that enter the food chain clearly indicates that the risk of dietary exposure to Atypical/Nor98 scrapie cannot be disregarded. However, according to our observations, in comparison to the brain, the infectious titres in the peripheral tissues were five log10 lower in Atypical/Nor98 scrapie than in classical scrapie. Therefore, the reduction of the relative exposure risk following SRM removal (CNS, head, spleen and ileum) is probably significantly higher in Atypical/Nor98 scrapie cases than in classical scrapie cases. However, considering the currently estimated prevalence of Atypical/Nor98 scrapie in healthy slaughtered EU population [10], it is probable that atypical scrapie infectivity enters in the food chain despite the prevention measures in force.
Finally, the capacity of Atypical/Nor98 scrapie agent (and more generally of small ruminants TSE agents) to cross species barrier that naturally limits the transmission risk is insufficiently documented. Recently, the transmission of an Atypical/Nor98 scrapie isolate was reported into transgenic mice over-expressing the porcine PrP [47]. Such results cannot directly be extrapolated to natural exposure conditions and natural hosts. However, they underline the urgent need for further investigations on the potential capacity of Atypical/Nor98 scrapie to propagate in other species than small ruminants.
All animal experiments were performed in compliance with our institutional and national guidelines, in accordance with the European Community Council Directive 86/609/EEC. The experimental protocols were approved by the INRA Toulouse/ENVT and by the Norwegian ethics committees.
The natural classical scrapie case (case 10) included in this experiment was a Romanov sheep born and bred in the Langlade flock where a natural scrapie epidemic has been occurring at a high incidence since 1993 [11].
Natural atypical scrapie cases were identified though active or passive surveillance programs in France, Norway and Portugal (Table 1). The Portuguese cases were identified in three independent flocks where an atypical case had already been identified in the past (additional cases). In all cases, PrP genotype was obtained by sequencing the Exon 3 of the Prnp gene as previously described [14]. In each case, the polymorphisms at codons 136 (A/V), 154 (H/R) and 171 (R/Q), which have been demonstrated to strongly influence the susceptibility to TSE in sheep are indicated [48]. Additionally the presence of a phenylalanine at codon 141 (F/L), which has been shown to impact on the susceptibility to atypical/Nor98 scrapie, was indicated [14], [49].
Two sheep (one 12 months old AFRQ/ARQ (case 8) and one 14 months old AHQ/AHQ (case 9)) selected in a field flock were IC challenged with French AFRQ/AFRQ Atypical Scrapie (case 1) (Table 2). The animals were euthanized when showing clear clinical signs at respectively 2224 and 964 days post inoculation.
TSE-free Poll-Dorset sheep (VLA- Weybridge- UK) were used for intracerebral inoculation with Langlade isolate (case 10, inoculum derived from a VRQ/VRQ natural isolate), or PG127 isolate (inoculum derived from a VRQ/VRQ experimental case). Animals were killed when displaying evident clinical signs at respectively 380 days and 160 days post inoculation. Oral challenge was performed in 6–10 months old TSE-free New Zealand cheviot sheep. Animals were dosed with 5 g equivalent of brain material (1% brain homogenate in glucose) derived from an experimentally VRQ/VRQ affected sheep (PG127 isolate). Animals were culled at clinical stage of the disease (200 days post inoculation).
All tissues were collected using disposable equipment (forceps and scalpels). The different field and experimental cases were sampled on different dates and/or places. Different instrument sets and containers were used for collecting, transporting and storing each sample. Finally, in all cases, peripheral tissues were collected before CNS to further reduce the risk of cross contamination. In natural Atypical/Nor98 cases, the nature of the tissues collected under TSE sterile conditions might have varied according to the country and date of collection. In all cases, CNS and at least one lymph node were available. In both Atypical/Nor98 and classical scrapie experimental cases, a large panel of tissues (including Central Nervous System, Peripheral Nervous System, digestive tract wall, muscle) was collected under TSE sterile conditions.
From the available samples, tissues homogenates (20% stock material) were prepared in Norway (Norwegian cases) or in France (French natural and experimental cases and Portuguese cases). The list of processed samples is given in Tables 1 and 2. In each case disposable equipment was used to manipulate the tissues. 20% tissues homogenates were prepared using single use grinding microtubes (Precess 48 - BioRad) and filtered through a 25 gauge needle (single use syringe). The tissue homogenates were then aliquoted (in 2 ml and 5 ml tubes) and stored at −80°C. Peripheral tissues homogenates and CNS homogenates were prepared separately.
This method was performed as previously described [50]. PrPSc IHC detection was first performed using 8G8 antibody raised against human recombinant PrP protein and specifically recognising the 95–108 amino acid sequence (SQWNKP) of the PrP protein.
For each sample a negative serum control was included, in which the primary antibody was either omitted or replaced by purified mouse IgG2a serum.
An OIE registered Western blot kit (TeSeE Western Blot, BioRad) was used following the manufacturer's recommendations. For each sample, 250 µl of 10% brain homogenate were submitted to PrPSc extraction. The obtained pellet was denaturated in Laemmli's buffer (15 µl) before being loaded neat or diluted (Figure 4) on a 12% acrylamide gel, and submitted to electrophoresis and blotting. Immunodetection was performed using SHa31 which recognizes the 145–152 sequence of PrP (YEDRYYRE). Peroxidase activity was revealed using ECL substrate (Pierce) [26].
A commercially available TSE detection test (TeSeE Sheep and Goat - BioRad) was used according to manufacturer's recommendations. In summary, five hundred µL of the 20% homogenate were incubated for 10 min at 37°C with 500 µL of buffer A containing proteinase K. PrPsc was recovered as a pellet after addition of 500 µL of buffer B and centrifugation for 5 min at 20 000 g at room temperature. Supernatant was discarded and tubes dried. Finally, the pellet was denatured in buffer C (5 min at 100°C) and 1:6 diluted in R6 reagent before distribution into the wells [28], [51].
PET blots were performed using a method previously described [52], [53]. Immunodetection was carried out using SHa31 monoclonal antibody (4 µg/mL), followed by application of an alkaline phosphatase labeled secondary antibody (Dako reference D0314 – 1/500 diluted). Enzymatic activity was revealed using NBT/BCIP substrate chromogen.
Bioassay experiments were carried out in ovine VRQ PrP transgenic mice (tg338), which are considered to be highly efficient for the detection of sheep scrapie infectivity [54]. At least six mice were intra-cerebrally inoculated with each sample (20 µL).
Prior to inoculation, homogenates were diluted (final concentration 10% or 12.5%) in 5% glucose sterile solution. Each homogenate was then tested for bacteria presence (blood gelose overnight 37°C culture) and non sterile homogenates were submitted to a heat treatment (60°C – 10 min). Heat treated samples are identified in Table 1. The impact of such heat treatment on atypical scrapie infectivity is currently unknown.
Portuguese and French cases' inoculations were carried out in UMR INRA ENVT 1225 (Toulouse, France) facilities while Norwegian cases were inoculated at the NVI (Oslo, Norway). Peripheral tissues and CNS homogenates were inoculated on different days in order to avoid any risk of cross contamination. In some cases, tissues autolysis resulted in the death of some animals inoculated which explain the low number of mice for some isolates. Mice were monitored daily until the occurrence clinical signs of TSE. Mice were culled when they started to show locomotor disorders and any impairment in their capacity to feed. CNS samples were individually collected. A part of the brain (cerebral cortex) was frozen for PrPSc Western blot testing (TeSeE WB kit- BioRad) and the other part of the brain was formalin fixed for vacuolar brain lesion profiling [55] and PrPSc PET-Blotting.
Five different isolates were endpoint titrated in tg338 mice, by inoculating intra-cerebrally (20 µl) successive 1/10 dilutions of CNS homogenate in groups of tg338 mice (6 or 12 mice). The material used for the titration was 10% brain homogenate except for the Langlade isolate (12.5% homogenate).
Two classical scrapie inocula used for sheep inoculation (Langlade: case 10, posterior brain stem - PG127: case 12, posterior brain stem) were titrated in UMR INRA ENVT 1225. The titration of the Langlade material was already published in a previous study [32].
Two confirmed atypical scrapie isolates (one from an ARQ/ARQ Norwegian sheep: case 14, cerebellum and one from a French ARR/ARR sheep: case 15, cerebellum) were titrated in INRA Jouy-en-Josas. These isolates correspond to two atypical cases originally described in the Le Dur et al. study in which they were respectively identified as Lindos and DS8 [7].
Three additional atypical scrapie isolates (AFRQ/AFRQ: case 1, cerebral cortex- AHQ/AHQ: case 9, cerebral cortex– AFRQ/ARQ: case 8, cerebellum) were titrated in UMR INRA ENVT 1225.
The infectious titre (Infectious Dose 50) of the brain homogenates was determined by the Spearman-Kärber's method [56].
For each isolate, the incubation periods recorded in individual tg338 and the number of ID50 inoculated to each mice (number of ID50 per 20 µL of the inoculated homogenate) (derived from Table 3) were plotted on a graph. On the basis of this data a four parameter logistic regression function was computed (Sigmaplot). This function was then used to estimate the infectious titre (number of Infectious Dose 50) contained in tissue samples on the basis of the incubation period observed in tg338 mice [29], [30], [31], [32], [33], [34].
CNS homogenate dilutions series from three different Atypical/Nor98 scrapie cases (case 1: cerebral cortex – case 8: cerebellum – case 9: cerebral cortex) were prepared by successive dilutions in negative brain homogenate.
The prepared dilutions were: 1/2, 1/5, 1/10, 1/20, 1/40, 1/80, 1/100, 1/200, 1/400, 1/800, 10−3, 10−4, 10−5, 10−6, 10−7, 10−8. The dilutions series were tested for PrPSc using TeSeE Sheep and Goat ELISA test and the WB as previously described in the text.
The neat sample and 10−5 to 10−7 dilutions from the same series were inoculated in groups of 6 tg338 in order to assess the infectious titre (see paragraph: Reference Central Nervous System samples endpoint titration).
Dilutions series of CNS homogenates were prepared from the Langlade scrapie (case 10: posterior brainstem) and PG127 (case 12: posterior brainstem) homogenates that were endpoint titrated in tg338 mice (Table 3). For these dilutions an aliquot of the 20% stock homogenate (stored at −80°C) was used as starting material. The dilution series (neat, 10−1, 10−2, 10−3, 10−4, 10−5, 10−6, 10−7) were then tested by WB (TeSeE WB kit – BioRad).
|
10.1371/journal.pcbi.1000910 | Gene Expression Variability within and between Human Populations and Implications toward Disease Susceptibility | Variations in gene expression level might lead to phenotypic diversity across individuals or populations. Although many human genes are found to have differential mRNA levels between populations, the extent of gene expression that could vary within and between populations largely remains elusive. To investigate the dynamic range of gene expression, we analyzed the expression variability of ∼18, 000 human genes across individuals within HapMap populations. Although ∼20% of human genes show differentiated mRNA levels between populations, our results show that expression variability of most human genes in one population is not significantly deviant from another population, except for a small fraction that do show substantially higher expression variability in a particular population. By associating expression variability with sequence polymorphism, intriguingly, we found SNPs in the untranslated regions (5′ and 3′UTRs) of these variable genes show consistently elevated population heterozygosity. We performed differential expression analysis on a genome-wide scale, and found substantially reduced expression variability for a large number of genes, prohibiting them from being differentially expressed between populations. Functional analysis revealed that genes with the greatest within-population expression variability are significantly enriched for chemokine signaling in HIV-1 infection, and for HIV-interacting proteins that control viral entry, replication, and propagation. This observation combined with the finding that known human HIV host factors show substantially elevated expression variability, collectively suggest that gene expression variability might explain differential HIV susceptibility across individuals.
| Many human genes have population-specific expression levels, which are linked to population-specific polymorphisms and copy-number variations. However, it is unclear whether human genes show similar dynamic range of expression between populations. In this work we analyzed HapMap gene expression compendium, and quantified the between-population and within-population expression variability for ∼18,000 human transcripts. We first concluded that the majority of the human genes have similar levels of within-population variability. However, a small fraction (∼4%) does show much higher expression variability in one population, and the deviation is consistently associated with increased SNP heterozygosity in their UTR regulatory regions. We further showed that genes with the greatest within-population expression variability are significantly enriched for chemokine signaling associated with HIV-1 infection. Combined with the finding that human HIV-1 host factors tend to have increased expression variability within populations, our analysis may explain, at least in part, different susceptibility to HIV infection within the human population. This work provides a fresh angle for analyzing gene expression variations in populations.
| In both prokaryotic and eukaryotic organisms, variations in gene expression exist widely within and between populations, which can be attributed to either genetic or non-genetic factors. Genetic factors are changes in DNA sequence that cause expression differences, such as single nucleotide polymorphisms (SNPs) and copy number variations (CNVs) on expression qualitative trait loci (eQTLs) [1], [2]. Non-genetic factors include epigenetic modifications [3], [4] and also innate expression stochasticity at the single-cell level [5], [6]. To date, extensive studies have investigated gene expression variation within and between natural populations of yeast [7], [8], fly [9]–[11], fish [12]–[14] and human [1], [2], [15]–[17]. These studies were mostly focused on identifying genes showing differential expression between populations or on localizing causal elements that affect expression changes among individuals (eQTL mapping). However, expression variation, as a manifested phenotype, in and of itself has complicated functional implications. It is established that the onset of many human diseases was associated with expression variation of some crucial genes [18], [19], and therefore gene expression variation is likely to be subject to selection. In this sense a systematic study on the expression variability within human populations is needed, which delineates the dynamic range of gene expression, i.e. to what degree a gene's expression could vary across individuals. This is of particular importance for several reasons. First, expression variability is conceptually distinct from differential expression (difference in mean expression level between populations); therefore studying expression variability might shed light on the evolution and differentiation of human gene expression. In analogy to sequence evolution, if a new advantageous expression level is rapidly fixed by natural selection in one population, a substantial reduction in expression variability might be expected. Second, expression variability is a natural estimate of dosage sensitivity of human genes. Due to natural selection, expression variability of dosage-sensitive genes is expected to be minimized; therefore investigation of expression variability might pave the way to future study of dosage sensitivity for human genes. Finally, recent genome-wide association studies have been based on the hypothesis of common disease-common variant (often abbreviated CD-CV), which carries the assumption that common variants might cause common aberrant expression of disease-associated genes, giving rise to pathological phenotypes. Given the widespread differential susceptibility to diseases within human populations, by circumventing the identification of causal sequence variants, a direct examination of expression variability of human genes and its implication towards disease susceptibility would highlight the importance of associating expression polymorphism to human disease.
In this paper, we sought to tackle the above questions by investigating the expression variability of human genes based on the previously published whole-genome expression profiling data [1], [2]. We found that, for most human genes, their within-population variability does not significantly differ between populations, with only a small group of genes exhibiting population-specific expression variability. Furthermore, this set of variable genes has SNPs in their untranslated regions (both 5′ UTRs and 3′UTRs) that show a pronounced elevated difference in population heterozygosity, which might explain, at least partially, their deviant expression variability between populations. We also found that a majority of human genes shows substantially reduced within-population variability, prohibiting the genes from differential expression between populations. Functional enrichment analysis revealed that genes with higher within-population variation are involved in a number of human diseases, particularly the early stage of HIV-1 entry into target cells, suggesting that expression variability is linked to variation in susceptibility to HIV infection among individuals.
The recently released whole-genome expression profiling data include 270 HapMap individuals spanning 4 ethnic populations [1], [2], including CHB (Chinese Han in Beijing), YRI (Yoruba people of Ibadan, Nigeria), CEU (U.S. residents with northern and western European ancestry) and JPT (Japanese from Tokyo). After preprocessing the expression data, we compiled expression profiles of 18, 081 human mRNA transcripts across all HapMap populations (CEU/YRI unrelated children, CEU/YRI unrelated parents, CHB and JPT, see Materials and Methods). After filtering out the Y-linked genes, we included both male and female samples since sex-biased expression is minimal (even for X-linked genes) in the lymphoblastoid cell line [20]. Although the subsequent analysis was based on CEU and YRI adult children (30 individuals in each population), all the conclusions hold for CEU/YRI parents, and also CHB and JPT, unless otherwise mentioned (see Figures S1, S2, S3, S4, S5).
We first sought to examine whether these genes have similar level of within-population variability in different populations. For each gene, we quantified the within-population expression variability by calculating its coefficient of variation η, which is the ratio of the standard deviation of its expression (across 30 individuals within a population) to the mean value [21]–[23]. Although other metrics can be used to quantify the expression variability, η is known to be one of the most robust and unbiased metrics [21]. Greater η implies higher expression variability for a particular gene across individuals within a population, while a significant reduction in η suggests that the gene might be dosage sensitive and thus under severe selection to minimize expression variability. The η values were calculated for each of the 18,081 mRNAs across individuals within the CEU and YRI populations separately (see Table S1 for genes with their calculated expression variability in each population). Between the CEU and YRI populations, most of the human genes exhibit a similar level of within-population variability, as η in CEU is well correlated with that in YRI (r = 0.88, P≈0; Figure 1). Pair-wise comparison of expression variability between all HapMap populations further confirmed this trend (r>0.85, P≈0). The same trend was recapitulated on another independent dataset of smaller sample size based on Affymetrix Human Focus Arrays [16], suggesting this observation was not resultant from a technical artifact. Therefore such a strong correlation of within-population expression variability between the two populations suggests either expression variability of most genes is subject to similar levels of constraints in both populations, or the cis- or trans- regulatory mechanisms of these genes have not diverged significantly.
Although the within-population variabilities of most human genes are tightly correlated between populations, a small number of genes do show noticeably different level of variability between CEU and YRI (Figure 1). To systematically identify those outliers with population-specific expression variability, we reciprocally regressed the values of η based on a linear model with random effects. Using residual analysis (see Materials and Methods) we were able to identify 919 and 898 genes as outliers for η's in YRI and CEU respectively as the explanatory variables. Among these outlier genes, 711 were found to be independent of the direction of the regression (either regressing ηCEU with ηYRI or regressing ηYRI with ηCEU, see Table S2 for a complete gene list). We noticed the presence of some annotated SNPs on the Illumina probes (affecting 4.5% of the 711 variable genes), so we removed the affected genes and only considered the remaining 679 outlier genes in our following analysis. We also noted that, among all the human genes, about 5% (916/18,081) had a probe overlapping with SNPs; this percentage is statistically indistinguishable from the percentage for the outlier genes (5% vs 4.5%, P-value = 0.50, Chi-square test). We thus eliminated the possibility that the observed expression variability was caused by the existence of SNPs in the microarray probes.
Could the observed asymmetric expression variability between populations be explained by their associated sequence variants? Supposing expression of a gene is only affected by a causative bi-allelic SNP, it is expected that the SNP with similar minor allele frequencies (MAFs) in both populations should have comparable expression variability of the associated gene. In other words, the observed increased expression variability of a particular gene is likely to be associated with some causative SNPs with divergent MAFs between two populations. Particularly under the assumption of Hardy-Weinberg Equilibrium for the diploid human populations, MAF of a SNP can be used to infer its expected heterozygosity θ (fraction of the heterozygous genotype) within a population [24]. Thus if a gene shows elevated expression variability in one population, the sequence variants affecting this gene are likely to have elevated expected heterozygosity within the population. Due to the difficulties in identifying trans-acting factors, we set out to examine this possibility for cis-SNPs surrounding the 679 genes showing population-specific expression variability.
We downloaded the promoter, 5′ UTR and 3′ UTR sequences for all human RefSeq genes (>20, 000) from UCSC Genome Browser [25], and mapped ∼3 million HapMap Phase II SNPs onto them (see Materials and Methods). We first examined the SNPs on 5′UTRs. We divided the 679 most variable genes into two groups: genes showing higher expression variability in CEU (383/679, termed CH group), and the remaining genes (296/679) showing higher expression variability in YRI (termed YH group). With the current SNP annotation, we were able to map SNPs onto the 5′ UTRs of 5, 690 human genes, including 130 CH genes and 94 YH genes. For each SNP on the CH genes, we calculated its difference in expected population heterozygosity between CEU and YRI (), and the same calculation was performed for all the SNPs on all the mapped 5, 690 human genes as background control (). As CH genes show elevated expression variability in CEU than in YRI, by comparing with genome background, we next tested if they are enriched for genes associated with higher population heterozygosity in CEU than in YRI (). As each gene often has multiple SNPs on its 5′UTRs, we first selected a cutoff, k, varying from 0 to 0.5 (the maximal ) with an increment of 0.04, and then compared the percentage of genes in each group (CH genes and background genes) bearing at least one SNP with greater than this cutoff. As seen in Figure 2(A), for all the cutoffs used, the CH genes consistently showed higher percentage than the genome background. To determine the statistical significance, we chose to use a stringent cutoff k = 0.04 (instead of using k = 0 to avoid numerical fluctuation), and found the percentage of genes in CH group bearing at least one SNP with >k is significantly higher than the genome background (P = 3.4×10−3, χ2 test). Similarly for YH genes, population heterozygosity was compared between YRI and CEU; thus and were calculated for each YH SNPs. With the same analysis, as shown in Figure 2(B), we reached the same conclusion that YH genes are significantly enriched for genes with elevated population heterozygosity in YRI (P = 0.05, χ2 test).
For 3′ UTR SNPs, we found the same enrichment for CH genes (P = 1.5×10−3, χ2 test), but not for the YH genes (P = 0.8, χ2 test). Moreover, neither CH nor YH genes show the trend on promoter SNPs (P>0.3, χ2 test). Taken together, the observed unequal expression variability between populations is likely to be explained, at least in part, by uneven MAF and population heterozygosity of the SNPs on UTR regions.
Among the 679 outlier genes that showed population-specific expression variability (see above), we were able to identify 184 genes that have differentiated expression levels between CEU and YRI (FDR≤0.01, 10,000 random permutations) after Benjamini and Hochberg FDR correction (see Materials and Methods), i.e. these genes on average have significantly higher expression levels in one population than in the other. For each of these 184 transcripts, we then plotted the distribution of within-population expression variabilities in CEU and YRI as a histogram in Figure 3, where the red diagonal line on the horizontal plane indicates equal expression variability in both CEU and YRI. Strikingly, we found among the total 184 transcripts, far more genes had higher expression variability in YRI (105 genes, 57%) than in CEU (79 genes or 43%). As we described in the above sections, among the total 679 outlier genes, 44% had higher expression variability in YRI, while among the 184 differentially expressed genes, a subset of the 679 outlier genes, the percentage substantially increased to 57%. With 10, 000 random permutation test, we confirmed such an enrichment of genes with higher expression variability in YRI is highly significant (P<10−5). Although the conclusion was drawn from 30 unrelated adult children from CEU and YRI, it also holds for the 60 unrelated parents in the two populations, suggesting our results are robust against sample size.
Among the majority of genes that have similar within-population expression variability in CEU and YRI (the non-outlier genes, see Materials and Methods), we also detected ∼20% (3, 429) that show differential expression levels between these populations with FDR = 0.01 (Benjamini and Hochberg FDR correction). Combined with the fact that only 184 among the 679 outlier genes (27%) show differential expression levels (see the above section), this clearly suggests the divergence of gene expression between populations is mostly manifested as a significant shift in expression levels without affecting within-population variability. We further quantified the degree of differential expression for each transcript between CEU and YRI through t-scores derived from a standard t-test (see Materials and Methods), which is the standardized distance of mean expression level between two populations. Higher absolute value of t-score is equivalent to a lower p-value, e.g. t = ±2 corresponding to p = 0.05 before Bonferroni correction, and t = ±5 corresponding to p = 0.05 after Bonferroni correction. As expression variability between CEU and YRI is almost perfectly correlated after removing the outliers in this study (r = 0.94), we only compared t-scores and expression variability for the transcripts in CEU (Figure 4, in which we used t = ±4 as a threshold to define differential expression levels between the populations, indicated by the two vertical lines, approximately corresponding to p = 2×10−4). As shown in Figure 4, a majority of genes has t-scores centered on 0 and has substantially reduced within-population expression variability compared with the genome background (the horizontal line). This observation indicates that a significant reeducation in expression variability within a population prohibits the genes from differential expression between populations. This group of genes is likely to be dosage-sensitive, which requires them to have similar expression levels between populations. It is also clear from Figure 4 that some genes have similar expression levels between two populations but also have very high expression variability (above the horizontal line); this implies these genes might be more dosage tolerant. We further noted a significant positive correlation between t-score and expression variability (r = 0.18, P<0.01) for genes shown in Figure 4, suggesting that genes with higher expression variability are more likely to develop more divergent expression levels between populations. Thus high expression variability is likely to confer higher expression evolvability. The conclusion stands when using another approach to identify the differentially expressed genes, which considers potential batch effects at the establishment of the cell lines [2].
Next we sought to determine whether genes with extreme within-population variability are specifically involved in any maladaptive processes. Since we are now studying the global trend of expression variability of human genes, we sought to exclude the genes that have population-specific expression variabilities. We excluded 1,106 of such genes from the total list of 18, 081 genes by either regressing ηCEU with ηYRI or regressing ηYRI with ηCEU (the union set, compared with the outliers as intersection set described above). In the end we retained a total of 16,975 mRNAs that showed similar variability in CEU and YRI. Since these transcripts have highly correlated within-population variability between these two populations, we focused the following analysis only on CEU population, unless otherwise mentioned.
The 16,975 mRNAs with homogeneous variability in the two populations were ranked according to their expression variability η from the lowest to the highest. By controlling the confidence level at 5%, we selected the top 2.5% and bottom 2.5% as the most and the least variable genes for further comparison respectively (424 out of 16, 975 genes for each group, see Table S3 for complete lists of genes). We performed an enrichment test by setting all 16, 975 transcripts in our study as background, then applied subsequent false discover rate (FDR) correction on each functional category using classifications in the DAVID biological database [26]. Functional enrichment analysis specifically included (1) Gene Ontology (GO) classifications (biological process, cellular component and molecular function at all levels), (2) KEGG pathways, (3) interaction with HIV-1 (human immunodeficiency virus 1) (from NCBI HIV-1, Human Interaction Database [27]), and (4) human disease annotations (from NIH Genetic Association Database [28] and OMIM).
As shown in Table 1, genes with the lowest expression variability are significantly enriched for fundamental biological processes such as translation and ribosome constituents (FDR = 0.02). The ribosomal genes are known to be dosage-sensitive [29]; this observation strongly suggests that expression variability within human populations indeed reflects intrinsic dosage-sensitivity of human genes. In sharp contrast with the least variable genes, genes with the greatest variability are enriched for behavior (FDR = 0.08), taxis (FDR = 0.02) and response to external stimulus (FDR≤0.05). While genes with the least expression variability are not associated with any human diseases reported from case-control studies deposited in GAD (NIH Genetic Association Database [28]), interestingly, genes with the highest expression variability are associated with seven human diseases (Table 1), mostly related to disease classes including ageing (FDR = 0.007) and neurological disorders (FDR = 0.036). Examination using disease associations documented in OMIM (Online Mendelian Inheritance in Man) did not find significant associations, however this might be due to the lower coverage of OMIM as compared to GAD, and the more stringent criteria used by OMIM in reporting disease associations. As GAD is primarily designed for collecting disease-associated genes bearing unevenly distributed biomarkers (e.g. SNPs), our observed disease association might be attributed to expression manifestation of these documented sequence polymorphisms.
In addition to being enriched for disease annotations listed above, genes with the highest expression variability also show significant enrichment for interaction with two HIV-1 proteins (see Materials and Methods). Notably, the highly variable genes are associated more frequently with the HIV-1 gene env (the precursor to HIV surface glycoprotein gp120; FDR = 0.018), and preferentially up-regulate the other HIV-1 gene, tat (FDR = 0.0024), whose protein product is of vital importance in regulating viral replication. Worthy of note, the HapMap samples used in this study were derived from lymphoblastoid B cells while the natural targets of HIV-1 are CD4+ T cells; however recent in vitro experiments have established that the lymphoblastoid cell line derived from B cells can well reflect the behavior of CD4+ T cells upon the infection of HIV-1 [30]. Therefore our observations suggest that the variation among individuals in their susceptibility to HIV viral entry or replication might be linked to the elevated expression variability of the host genes interacting with env and tat. Further lending support to this hypothesis, we found that variable genes are also enriched for chemokine receptors (FDR = 0.08). Since the HIV-1 virus fuses into target cells mainly through interactions between gp120 and chemokine receptors (e.g. CXCR4 and CCR3), this strongly supports that variability across populations is inherently linked to varied susceptibility to HIV-1.
The HIV-1 genome consists of 9 genes: env, gag, nef, pol, rev, tat, vif, vpr and vpu. To further explore the strong association between expression variability of host genes and HIV-1 pathogenesis, we next compared the expression variability of human host factors interacting with each of the 9 viral genes against human genome background (for CEU and YRI separately). The host-virus interactions were extracted from HIV-1, Human Protein Interaction Database [27]. We were able to identify 700, 194, 235, 211, 73, 853, 83, 215 and 30 human transcripts in our data set that have annotated interactions with the 9 HIV-1 genes respectively, and we examined the interactions in all categories (e.g. physical interaction, up-regulate or down-regulate, etc.). Strikingly, for 5 of the 9 HIV-1 genes (env, gag, nef, tat and vpr), the host factors exhibited significantly elevated expression variability in both populations (all p-values<0.05, Wilcoxon ranksum test; Figure 5a). For rev (regulator of virion) and vpu (viral protein U), only YRI population exhibited elevated expression variability (note that the relatively large error bars for vpu in both populations were due to small sample size as only 30 human genes were annotated to interact with vpu).
As the genome-wide expression profiling was performed in the lymphoblastoid cell line (an immune-related cell line that HIV virus can attack), combined with the observation that genes involved in immune system are enriched among the host factors interacting with the viral genes (P-value<0.05), it is tempting to trivially explain the above observation by the intrinsic variability of immunity genes [31]–[33]. To ascertain this possibility, we identified 361 human transcripts (∼16% of all the host factors in this study) that contain the keyword “immune” in their Gene Ontology annotations (Biological Processes, all hierarchies), and removed them from the host factors and repeated the above comparison. Again, we found host factors interacting with nef (negative regulatory factor), tat (trans-activator of transcription) and vpr (viral protein R) constantly show elevated expression variability in both CEU and YRI, which suggests that the elevated expression variability of the host genes cannot be fully explained by the enrichment of the immunity genes.
After ruling out the effect of immunity genes, we next applied two approaches to ascertain the possibility that the elevation of expression variability for HIV-interacting genes could be due to enrichment of highly variable GO functional categories. (i) Firstly, we pooled together the entire 1, 480 human genes that were annotated to interact with at least one HIV-1 genes, and removed 551 genes associated with the highly variable functions (based on GO terms derived from Table 1 and Supplemental Table S4, we removed all genes associated with these GO terms and their descendents in the GO hierarchy). For the remaining 929 HIV-interacting genes, again we observed their within-population expression variability is significantly higher than genome background in both CEU and YRI (showing ∼17% increase in comparison with expression variability of all human genes, P<10−11, Wilcoxon ranksum test). (ii) In the second approach, we generated “null” sets of genes, mirroring the GO functional categories of the 1,480 HIV-1 interacting genes and compared the variability of these null sets to the real gene set. Among the 1, 480 genes, we were able to consider 1, 284 genes, whose GO annotations (the most specific code) were also associated with at least one non-HIV interacting gene. We then chose a non-HIV-interacting gene with the same GO code and repeated this for every one of the 1,284 genes to make a null set. We repeated this procedure 1000 times by generating 1000 null gene sets, and asked among the 1000 simulations, how many times we observe the real data have significantly higher expression variability than the null set. Consistently, we found in all simulations, the real data always have average higher expression variability (on average 8% higher), and 991 out of the 1000 simulations are statistically significant. Thus we concluded that the observed elevation in expression variability of HIV-interacting genes is unlikely an artifact caused by the bias in the GO functional annotations.
Next we curated a list of human genes from the published literature that are known to induce differential susceptibility to HIV, and compared their expression variability with the genomic background. These genes included chemokine receptors (CCR2 [34]–[36], CXCR4 [37]), HIV-suppressive β-chemokines (CCL3 [38], CCL3L1 [39], CCL4 [40], CCL5 [41], [42], CXCL12 [43], [44]), a human endogenous HIV-1 replication inhibitor known to be involved in the mid stage of viral propagation (APOBEC3G [45]), and a newly identified inducible host factor implicated in the late stage of HIV-1 replication pathway (SOCS1) [46]. As shown in Figure 5(b), these key host factors have substantially elevated expression variability as compared to the genomic background. For example, CXCR4, one of the major chemokine receptors, has an almost 4.3-fold increase in expression variability, suggesting that it might have extremely low expression level in some individuals, leading to increased resistance to HIV entry (particularly for X4 strain, which utilizes CXCR4 for viral entry). Although we did not observe significantly elevated expression variability for CCR5 (η = 0.02, slightly higher than the genome background), we indeed found its ligand CCL3L1 had a 3-fold increase in expression variability. This is consistent with the previous observation that increased copy number of CCL3L1 in some individuals can effectively reduce the risk of HIV-1 infection [39]. Similarly, CXCL12 (SDF-1), the ligand of CXCR4, has a 4.4-fold increase in expression variability. These results collectively bolster the hypothesis that variation in genetic expression within a population may result in altered susceptibility to HIV-1 infection.
We further compared our results with a recent work by Loeuillet et al [30], in which the authors established a link between a SNP (rs2572886) to differential HIV susceptibility among European individuals by transduction of lymphoblastoid cells (the same cell line used in our study) with a HIV-1-based vector (HIV.GFP). The identified SNP is associated with 8 genes belonging to the LY6/uPAR family, and the authors prioritized 4 proteins (LYP6D, LYPD2, SLURP1 and GML) for over-expression study and 2 proteins (LY6D and LYPD2) for RNAi knockdown. However the authors did not observe HIV infectivity being significantly affected by these perturbations [30]. We re-examined expression variability among CEU individuals for these prioritized proteins, and found their expression variability is substantially below genome average (between 0.009–0.01, compared with the genome median of ∼0.0197). Among the remaining 4 tagged genes that were not examined in the original study, LY6E showed almost ∼1.8–2.5-fold increase in expression variability in comparison with that of background genes (expression variability of LY6E is 0.049 and 0.035 in CEU and YRI, respectively, in comparison with the background median of ∼0.0197). Therefore a re-examination of LY6E might be needed in future studies to elucidate the roles of this gene in affecting HIV susceptibility.
Although extensive efforts have been made to elucidate the effects of sequence variants on expression phenotypes, it is likely that not all expression variation can be fully explained by genetic factors [1], [47]. As gene expression is more pertinent to molecular functions, exploration of expression variability within and between human populations could provide additional insights into functional evolution of human genes. Unlike previous work that had focused on finding genes that are differentially expressed between populations [15], [16], [48], [49] or mapping eQTLs [17], [47], throughout this paper, we quantified expression variability for each human gene within individual human populations, and attempted to interpret the functional and evolutionary implications of such variations.
Our results revealed that the evolution of differential expression in human is largely manifested as a shift in mean expression level between populations without affecting their respective expression variability in each population. As within-population expression variability could be used to approximate dosage-sensitivity of a given gene, our observation also suggests that dosage-sensitivity of human genes is largely conserved between human populations. We also found that differentially expressed genes are more likely to have higher expression variability, which suggests variability might confer higher evolvability due to relaxed constraints.
For those genes that do have significantly different variability between distinct populations (referred as outliers), we also observed dissimilar minor allele frequencies (and thus population heterzygosity) between CEU and YRI in their UTRs, particularly on the 5′UTRs. It is possible that in addition to the cis-regulatory regions, other trans-acting and non-genetic factors might also take effect.
Our analysis revealed that genes with the highest expression variability within human populations are significantly associated with a number of human diseases, which may account for the differential susceptibility to diseases among human individuals. Although it is expected that sequence polymorphisms tend to be associated with elevated expression variability, other factors such as copy number variations (CNV) and epigenetics, could also cause variation in gene expression level. To this end, we compiled a list of ∼1, 800 RefSeq genes that reside in CNV regions identified from a recent fine-resolution mapping with pair-end sequencing [50]; however, we did not find the genes showing the highest expression variability are enriched for CNV genes. At the present time, it is difficult to separate the epigenetic effects from genetic effects based on available data, but it is important to note that epigenetic diversity across individuals and among populations can have profound impact in expression variability.
It has long been noted that susceptibility to HIV infection differs greatly among individuals, and individuals infected with HIV also have substantially varied rate of disease progression to full-blown AIDS. To explain such variation in viral resistance, several sequence variants of human genes have been identified, which is best exemplified by CCR5-Δ32 deletion [51], [52] and CCL3L1 copy number variants [39]. By circumventing the identification of the associated sequence variants, our analysis on gene expression posed an important question in understanding the differential HIV susceptibility, i.e. whether examining expression polymorphisms can directly assess such a difference. Our results corroborated such possibility, i.e. host factors interacting with several HIV genes, controlling viral entry, progression and replication cycles, show substantially elevated expression variability among individuals. Interestingly, although host factors involved in immune system are major targets in current HIV research, our results also demonstrated that non-immunity genes that interact with viral genes nef, tat and vpr also show significantly elevated expression variability. This observation might help expand the list of candidate genes that reduce HIV susceptibility. From an evolutionary perspective, our observation might also suggest that the virus can increase the chance of survival by preferentially targeting variable host factors.
The recently released whole-genome expression profiling of 270 HapMap individuals spanning 4 ethnic populations in the lymphoblastoid cell line [2], [20], includes CHB (Chinese Han in Beijing), YRI (Yoruba people of Ibadan, Nigeria), CEU (U.S. residents with northern and western European ancestry) and JPT (Japanese from Tokyo). Using an Illumina annotation table, we unambiguously mapped 18,127 utilized probes to human mRNA transcripts (only those with RefSeq NM_ identifiers). We then removed the 10% of genes with the lowest expression level (assuming they are not expressed in the lymphoblastoid cell line). The Illumina-annotated gene symbols were mapped onto officially approved HGNC (HUGO Gene Nomenclature Committee) symbols, allowing us to retain a total of 15, 554 unique HGNC genes. We filtered out Y-linked genes, and included both male and female samples in this study since sex-biased expression is minimal (even for X-linked genes) in the lymphoblastoid cell line [20]. We separated expression data of adult children from the unrelated parents because the trio family data might bring unnecessary dependency between data points because of parent-child inheritance in gene expression [2]. Finally we were able to retain 18,081 mRNA transcripts and 15,501 HGNC genes for each of the 30 individuals in both CEU and YRI populations. In addition, we were also concerned with the potential bias caused by the presence of SNPs on the designed microarray probes; however, after mapping the ∼3 million annotated HapMap SNPs onto the 18, 081 Illumina probes, we found the influence is minimal as ∼95% of the probes was not affected.
We used the same expression data as above to identify differentially expressed genes, but the data were median-normalized across composite population by pooling all populations together. This is of vital importance in differential expression analysis because in this way we could normalize the expression profiles of CEU and YRI using the same background scale. By excluding genes showing population-specific variability, we were able to consider 16, 878 transcripts in differential expression analysis.
We downloaded the annotated HIV-1, human protein interactions from NCBI (http://www.ncbi.nlm.nih.gov/RefSeq/HIVInteractions/) [27]. We considered human genes having “all” interactions with each of the nine HIV-1 genes, and mapped the Entrez ID to RefSeq mRNA IDs by using the DAVID ID conversion tool [26]. After overlapping with the transcripts in our study, we were able to consider 700, 194, 235, 211, 73, 853, 83, 215 and 30 transcripts interacting with HIV-1 genes env, gag, nef, pol, rev, tat, vif, vpr and vpu, respectively.
To identify genes with population-specific expression variability within CEU and YRI, we regressed expression fluctuation, η, in YRI and in CEU reciprocally and derived two lists of genes showing population-specific variation by using CEU and YRI as explanatory variables, respectively. About ∼70% of the genes on one list also appear on another list. The liner model was derived by minimizing the square errors between the observed η and the predicted values (). Taking YRI as an example, the residues, , were then normalized and Studentized. For each gene, by fitting a t-distribution, we calculated 95% confidence intervals (CIs) of its residue, and the outliers were defined as the genes away from the calculated 95% CIs of the fitted t-distribution.
We extracted promoter sequences (annotate by UCSC Genome Browser as upstream 1kb regions from transcription start site), 5′UTR, and 3′UTRs for both outlier genes and all annotated human genes in UCSC.
Our protocol is similar as described in [15], in which we performed 10, 000 permutation t-test followed by Benjamini and Hochberg FDR correction.
|
10.1371/journal.pbio.3000065 | Contest models highlight inherent inefficiencies of scientific funding competitions | Scientific research funding is allocated largely through a system of soliciting and ranking competitive grant proposals. In these competitions, the proposals themselves are not the deliverables that the funder seeks, but instead are used by the funder to screen for the most promising research ideas. Consequently, some of the funding program's impact on science is squandered because applying researchers must spend time writing proposals instead of doing science. To what extent does the community's aggregate investment in proposal preparation negate the scientific impact of the funding program? Are there alternative mechanisms for awarding funds that advance science more efficiently? We use the economic theory of contests to analyze how efficiently grant proposal competitions advance science, and compare them with recently proposed, partially randomized alternatives such as lotteries. We find that the effort researchers waste in writing proposals may be comparable to the total scientific value of the research that the funding supports, especially when only a few proposals can be funded. Moreover, when professional pressures motivate investigators to seek funding for reasons that extend beyond the value of the proposed science (e.g., promotion, prestige), the entire program can actually hamper scientific progress when the number of awards is small. We suggest that lost efficiency may be restored either by partial lotteries for funding or by funding researchers based on past scientific success instead of proposals for future work.
| The grant proposal system compels researchers to devote substantial time to writing proposals that could have instead been used to do science. Here, we use the economic theory of contests to show that as fewer grants are funded, the value of the science that researchers forgo while preparing proposals can approach or exceed the value of the science that the funding program supports. As a result, much of the scientific impact of the funding program is squandered. Unfortunately, increased waste and reduced efficiency is inevitable in a grant proposal competition when the number of awards is small. How can scarce funds be allocated efficiently, then? As one alternative, we show that a partial lottery that selects proposals for funding randomly from among those that pass a qualifying standard can restore lost efficiency by reducing investigators' incentives to invest heavily in preparing proposals. Lotteries could also improve efficiency by compelling administrators to de-emphasize grant success as a primary measure of professional achievement. If lotteries are politically untenable, another remedy would be to fund researchers based on their previous research successes, although in such a way that avoids establishing barriers to entry for junior scientists or scientists from historically underrepresented demographic groups.
| Over the past 50 years, research funding in the United States has failed to keep pace with growth in scientific activity. Funding rates in grant competitions have plummeted (S1 Fig, [1–4]) and researchers spend far more time writing grant proposals than they did in the past [5]. A large survey of top US universities found that, on average, faculty devote 8% of their total time—and 19% of their time available for research activities—towards preparing grant proposals [6]. Anecdotally, medical school faculty may spend fully half their time or more seeking grant funding [5, 7]. While the act of writing a proposal may have some intrinsic scientific value [8]—perhaps by helping an investigator sharpen ideas—much of the effort given to writing proposals is effort taken away from doing science [9]. With respect to scientific progress, this time is wasted [10].
Frustrated with the inefficiencies of the current funding system, some researchers have called for an overhaul of the prevailing funding model [9, 11–18]. In particular, Fang and Casadevall [16] recently suggested a partial lottery, in which proposals are rated as worthy of funding or not, and then a subset of the worthy proposals are randomly selected to receive funds. Arguments in favor of a partial lottery include reduced demographic and systemic bias, increased transparency, and a hedge against the impossibility of forecasting how scientific projects will unfold [16]. Indeed, at least three funding organizations—New Zealand's Health Research Council and their Science for Technological Innovation program, as well as the Volkswagen Foundation [19]—have recently begun using partial lotteries to fund riskier, more exploratory science.
Compared with a proposal competition, a lottery permits more proposals to qualify for funding and thus lowers the bar that applicants must clear. A lottery also offers a lower reward for success, as a successful proposal receives a chance at funding, not a guarantee of funding. Thus, we expect that investigators applying to a partial lottery will invest less time and fewer resources in writing a proposal. To a first approximation, then, a proposal competition funds high-value projects while wasting substantial researcher time on proposal preparation, whereas a partial lottery would fund lower-value projects on average but would reduce the time wasted writing proposals. It is not obvious which system will have the greater net benefit for scientific progress.
In this article, we study the merits and costs of traditional proposal competitions versus partial lotteries by situating both within the rich economic theory of contests. In this theory, competing participants make costly investments ("bids") in order to win one or more prizes [20, 21]. Participants differ in key attributes, such as ability and opportunity cost, that determine their optimal strategies. In an economics context, contests are often used by the organizer as a mechanism to elicit effort from the participants. For example, TopCoder and Kaggle are popular contest platforms for tech firms (the organizers) to solicit programming or data-analysis effort from freelance workers (the participants). However, because the participants' attributes influence their optimal strategies, the bids that participants submit reveal those attributes. Thus, screening of participants often arises as a side effect.
In funding competitions, the organizer is a funding body and the participants are competing investigators. Investigators pitch project ideas of varying scientific value by preparing costly proposals. However, unlike a traditional economic contest, the funding body's primary objective is to identify the most promising science, using proposals to screen for high-value ideas. The funding body has little interest in eliciting work during the competition itself, as the proposals are not the deliverables the funder seeks. All else equal, the funding agency would prefer to minimize the work that goes into preparing proposals, to leave as much time as possible for investigators to do science. In this case, how should the funder organize the contest to support promising science without squandering much of the program's benefit on time wasted writing proposals?
Below, we pursue this question by presenting and analyzing a contest model for scientific funding competitions. We first use the model to assess the efficiency of proposal competitions for promoting scientific progress and ask how that efficiency depends on how many proposals are funded. We next explore how efficiency is impacted when extrascientific incentives such as professional advancement motivate scientists to pursue funding, and compare the efficiency of proposal competitions with that of partial lotteries. Finally, we reflect on alternative ways to improve the efficiency of funding competitions without adding intentional randomness to the award process. All of our analyses focus on equilibrium behavior and thus pertain most directly to long-standing funding competitions, for which researchers can acquire experience that informs their future actions.
Our model draws upon a framework for contests developed by Moldovanu and Sela [20]. In our application, a large number of scientists (or research teams) compete for grants to be awarded by a funding body. The funder can fund a proportion p of the competing investigators. We call p the payline, although p could be smaller than the proportion of investigators who are funded if some investigators do not enter the competition.
Project ideas vary in their scientific value, which we write as v, where v≥0. In this case, scientific value combines the abilities of the investigator and the promise of the idea itself. Although we do not assign specific units to v, scientific value can be thought of as some measure of scientific progress, such as the expected number of publications or discoveries. We assume that the funder seeks to advance science by maximizing the scientific value of the projects that it funds, minus the value of the science that investigators forgo while writing proposals. However, the funder cannot observe the value of a project idea directly. Instead, the funder evaluates proposals for research projects, and awards grants to the top-ranked proposals. Assume that proposals can be prepared to different strengths, denoted x≥0, with a larger value of x corresponding to a stronger proposal. A scientist with a project idea of value v must decide how much effort to invest in writing a proposal, that is, to what strength x her proposal should be prepared. In our model, this decision is made by a cost–benefit optimization.
On the benefit side, if a proposal is funded, the investigator receives a reward equal to the scientific value of the project, or v. This reward is public, in the sense that it benefits both the investigator and the funder. Receiving a grant may also bestow an extrascientific reward on the recipient, such as prestige, promotion, or professional acclaim. Write this extrascientific reward as v0≥0. This extrascientific reward is private, as it benefits only the grant recipient and not the funder. Let η(x) be the equilibrium probability that a proposal of strength x is funded; η(x) will be a nondecreasing function of x. Thus, in expectation, an investigator with a project of value v who prepares a proposal of strength x receives a benefit of (v0+v)η(x).
Preparing a grant proposal also entails a disutility cost equal to the value of the science that the investigator could have produced with the time and resources invested in writing. Let c(v,x) give the disutility cost of preparing a proposal of strength x for a project of value v. Here, we study the case where c(v,x) is a separable function of v and x, so we set c(v,x) = g(v)h(x). Proposal competitions are effective screening devices because it is easier to write a strong proposal about a good idea than about a poor one. Therefore, g(v) is a decreasing function of v, i.e., g′(v)<0. For a given idea, it takes more work to write a stronger proposal, and thus h′(x)>0. Finally, we assume that preparing a zero-strength proposal is tantamount to opting out of the competition, which can be done at zero cost. Thus, h(0) = 0.
Preparing a proposal has some scientific value of its own through the sharpening of ideas that writing a proposal demands [8]. Let k∈[0,1) be the proportion of the disutility cost c(v,x) that an investigator recoups by honing her ideas. We call the recouped portion of the disutility cost the intrinsic scientific value of writing a proposal. The portion of the disutility cost that cannot be recouped is scientific waste.
All told, the total benefit to the investigator of preparing a proposal to strength x is (v0+v)η(x)+kc(v,x), and the total cost is c(v,x). The difference between the benefit and the cost is the investigator's payoff. The investigator's optimal proposal (or, in economic terms, her "bid") maximizes this payoff (Fig 1):
b(v)=argmaxx{(v0+v)η(x)−(1−k)c(v,x)}.
(1)
For simplicity, we assume that variation among projects is captured entirely in the distribution of v, which we write as F(v). We assume that v0 and k have common values shared by all investigators. In S1 Text we show that our results extend to cases where v0 or k vary among investigators, as long as they are perfectly correlated with v.
The challenge in finding the payoff-maximizing bid b(v) is that the equilibrium probability of funding, η(x), must be determined endogeneously, in a way that is consistent with both the payline p and the distribution of bids that investigators submit. In S1 Text, we follow Hoppe and colleagues [22] to show that, at equilibrium, the bid function is given by
b(v)=h−1[11−k∫0vv0+tg(t)ξ′(t)dt].
(2)
In Eq 2, ξ(v) = η(b(v)) is the equilibrium probability that an idea of value v is funded. The particular form of ξ(v) depends on how much randomness is introduced during the review process, which we discuss below.
By comparison, Moldovanu and Sela [20] considered a contest with a small number of competitors, in which the contest's judges observe x directly. In their setup, each contestant is uncertain about the strength of her competition (that is, her competitors' types, v), but she can be certain that the strongest bid will win the top prize. In our case, we assume that the applicant pool is large enough that the strength of the competition (i.e., the distribution of v among the applicants) is predictable. However, the funding agency does not observe x directly, but instead convenes a review panel to assess each proposal's strength. Variability among reviewers' opinions then introduces an element of chance into which proposals get funded.
We use the model to explore how efficiently the grant competition advances science. From the perspective of an individual investigator, the investigator's return on her investment (ROI) is the ratio of her payoff to the cost of her bid:
Investigator′sROI=(v0+v)η(b(v))−(1−k)c(v,b(v))c(v,b(v)).
(3)
An investigator will never choose to write a proposal that generates a negative payoff, because she can always obtain a payoff of 0 by opting out. (If the investigator opts out, Eq 3 evaluates to 0/0, in which case we define her ROI to be 0.) Thus, an investigator's equilibrium ROI must be ≥0.
To analyze the funding program's impact on scientific progress as a whole, we compare the total value of the science that the funding program supports with the total value of the science that has been squandered preparing proposals. Of course, both of these quantities will be confounded with the number of grants that are funded, so we standardize to a per-funded-proposal basis. In notation, the average scientific value per funded proposal is
1p∫vη(b(v))dF(v),
(4)
and the average scientific waste per funded proposal is
1p∫(1−k)c(v,b(v))dF(v).
(5)
We will refer to the difference between these two quantities as the scientific gain (or loss, should it be negative) per funded proposal, which is our measure of the funding program's scientific efficiency.
Note that while an investigator will never enter a grant competition against her own self interest, there is no guarantee that the scientific value per funded proposal will exceed the scientific waste. This is because the investigator's payoff includes private, extrascientific rewards obtained by winning a grant (v0), and (in our accounting, at least) these extrascientific rewards do not benefit the funding agency. If extrascientific motivations for winning grants are large enough, investigators may enter a grant competition even when doing so decreases their scientific productivity. If enough investigators are motivated accordingly, then the scientific progress sacrificed to writing proposals could exceed the scientific value of the funding program. In this case, the grant competition would operate at a loss to science, and the funding agency could do more for science by eschewing the proposal competition and spreading the money evenly among active researchers in the field, or by giving the money to researchers selected entirely at random.
We illustrate the model's behavior by choosing a few possible sets of parameter values. Our parameter choices are not directly informed by data. Thus, while the numerical examples illustrate the model's possible behavior, we highlight the results that are guaranteed to hold in general. Throughout, we use the following baseline set of parameters. We assume that the project values, v, have a triangular distribution ranging from vmin = 0.25 to vmax = 1 with a mode at vmin, such that low-value ideas are common and high-value ideas are rare (i.e., F(v) = 1−(16/9)(1−v)2). For the cost function, we choose c(v,x) = x2/v. We choose a convex dependence on x to suggest that the marginal cost of improving a proposal increases as the proposal becomes stronger. We assume that the intrinsic scientific value of writing a proposal allows investigators to recoup k = 1/3 of the disutility cost of proposal preparation. We first explore the case when investigators are motivated purely by the scientific value of their projects (v0 = 0) and then introduce extrascientific benefits (v0 = 0.25). In S1 Text and S3–S6 Figs, we provide parallel results with two alternative parameter sets.
The evaluation process by which review panels rank proposals introduces a layer of randomness to the awarding of grants [23–25]. To capture noisy assessment, we use a bivariate copula [26] to specify the joint distribution of a proposal's actual quantile, and its quantile as assessed by the funding agency's review panel. A bivariate copula is a probability distribution on the unit square that has uniformly distributed marginals, as all quantiles must. We use a Clayton copula [27], which allows for accurate assessment of weak proposals, but noisier assessment of strong proposals (S2 Fig). This choice is motivated by the pervasive notion that review panels can readily distinguish strong proposals from weak ones, but struggle to discriminate among strong proposals [16, 25, 28]. A Clayton copula has a single parameter (θ) that controls how tightly its two components are correlated. Rather arbitrarily, we use θ = 10 in the baseline parameter set. The Clayton copula has the important property that a proposal's probability of funding increases monotonically as its strength increases, regardless of the payline. Thus, we exclude the possibility that panels systematically favor weaker proposals. By using a copula, we implicitly assume that η(x) depends on x only through its rank. In S1 Text, we show how a copula leads to an equation for ξ′(v), which can then be plugged in to Eq 2.
Fig 2 shows numerical results for the baseline parameters at generous (p = 45%) and low (p = 15%) paylines. In this particular case, investigators' payoffs fall faster than costs as paylines drop, leading to a reduced ROI for everyone at the lower payline (Fig 2B). We will argue below that every investigator's ROI must inevitably fall when the payline becomes small (see S2 and S3 Figs for additional examples).
From the funding agency's perspective, with our baseline parameters, both the average scientific value and average waste per funded proposal increase as the payline falls, for paylines below 50% (Fig 3A). However, as the payline decreases, waste escalates more quickly than scientific value, reducing the scientific gain per funded project (Fig 3B). This same result also appears in our alternative parameter sets (S5 and S6 Figs). We will argue below that the decline in scientific efficiency at low paylines is an inevitable if unfortunate characteristic of proposal competitions.
Clearly, quantitative details of the model's predictions depend on the parameter inputs. To understand the robustness of these predictions, it helps to study the case in which panels discriminate perfectly among proposals. While perfect discrimination is obviously unrealistic in practice, it yields a powerful and general set of results that illuminate how the model behaves when discrimination is imperfect. Numerical results for perfect discrimination under the baseline parameter set appear in S7 and S8 Figs.
At equilibrium under perfect assessment, every project above a threshold value v* = F−1(1−p) will receive funding, and no project idea below this threshold will be funded. Investigators with projects of value v>v* all prepare proposals to the identical strength x* = h−1[(v0+v*)/((1−k)g(v*))] and are funded with certainty. Investigators with projects of value v<v* opt out (S7 Fig). All of the subsequent results follow (details appear in S1 Text). First, as paylines drop, all investigators realize either a diminishing or zero ROI, because investigators who remain in the competition must pay a higher cost for a reduced payoff. Second, the average scientific value per funded proposal must increase as paylines drop, because only the highest-value projects are funded under low paylines. Third, in the limiting case in which only one of many proposals can be funded (technically, the limit as p approaches 0 from above), the scientific value and scientific waste associated with the last funded project converge, and science is no better off than if no grant had been given at all (S8 Fig).
With perfect assessment, there is no general relationship between the scientific efficiency of a proposal competition and the payline that holds across the full range of paylines (but see Hoppe and colleagues [22] for a sharp result when the cost function is independent of v). Of course, we wouldn't expect scientific efficiency to decline monotonically with a falling payline, because there are likely to be some low-value projects that can be weeded out at low cost. However, our last result above guarantees that the scientific gain per funded proposal must eventually vanish as the payline declines to a single award.
Returning to the reality of imperfect discrimination, as long as review panels do not systematically favor weaker proposals, noisy assessment changes little about these qualitative results. That is, investigators' ROIs will drop as paylines fall, the average scientific value per funded proposal will increase as paylines decrease, and the scientific efficiency of the proposal competition must eventually decline as the payline approaches a single award. But efficiency need not drop to zero. Perhaps counterintuitively, imperfect discrimination is a saving grace at low paylines. Noisy assessment discourages top investigators from pouring excessive effort into grant writing as paylines fall, because the marginal benefit of writing an even better grant becomes small when review panels struggle to discriminate among top proposals. Indeed, noisy assessment, unlike perfect discrimination, allows a proposal competition to retain a positive impact on science, even with a single funded grant (compare Fig 3 and S8 Fig). This result hints at the salutary nature of randomness at low paylines, which we will see more vividly when we consider lotteries below.
Thus far, we have considered the case in which investigators are motivated only by the scientific value of the projects proposed (v0 = 0). Now, suppose that investigators are additionally motivated by the extrascientific benefits of receiving a grant, such as professional advancement or prestige (v0>0). Eq 2 shows that adding extrascientific motivation will increase the effort that investigators devote to preparing grant proposals. However, in our model, at least, this extra effort has no bearing on which grants are funded and thus does not affect the scientific value of the grants that are awarded. Increasing scientific costs without increasing scientific value will clearly be detrimental to the funding program's scientific efficiency. Extrascientific benefits to investigators can even cause the entire funding program to operate at a loss to science when paylines are low (Fig 3).
Our model can also be used to analyze the efficiency of a partial lottery for advancing science. Suppose that a fraction q≥p of proposals qualify for the lottery, and each qualifying proposal is equally likely to be chosen for funding. Call q the "lottery line." Now, the investigator's payoff is (p/q)(v0+v)ηl(x)−(1−k)c(v,x), where ηl(x) is the equilibrium probability that the proposal qualifies for the lottery. In S1 Text, we show that the investigator's bid is given by
b(v)=h−1[pq11−k∫0vv0+tg(t)ξl′(t)dt]
(6)
where ξl(v) = ηl(b(v)).
Our major result for lotteries is that measures of scientific efficiency—expressions 3, 4, and 5—depend on the lottery line q but are independent of the payline p (proofs appear in S1 Text). This result follows from the fact that, in a lottery, each investigator's benefit and cost are proportional to p. Thus, an investigator's ROI and the scientific efficiency of the funding program are determined by the lottery line but are not affected by the payline. To illustrate, Fig 4 compares an investigator's costs and benefits in a proposal competition with 45%, 30%, and 15% paylines versus a partial lottery with a q = 45% lottery line and the same three paylines. The key feature of Fig 4 is that the investigator's benefit curve in a partial lottery scales in such a way that her ROI is the same for any payline ≤q. Consequently, a partial lottery with a lottery line of q and any payline ≤q achieves the same scientific efficiency as a proposal competition with a payline of q.
Thus, our numerical results showing the investigator's ROI (Fig 2B) or the scientific efficiency (Fig 3) in a funding competition also show the efficiency of a lottery with the equivalent lottery line. That is, a lottery in which 45% of applicants qualify for the lottery has the same scientific efficiency as a proposal competition with a 45% payline, regardless of how many proposals in the lottery are randomly selected for funding. Thus, a lottery can restore the losses in efficiency that a proposal competition suffers as paylines become small.
In S1 Text, we also analyze a more general type of lottery in which proposals are placed into one of a small number of tiers, with proposals in more selective tiers awarded a greater chance of funding [13, 17, 29]. In a multitiered lottery, the efficiency is entirely determined by the number of tiers and the relative probabilities of funding in each, and is independent of the payline. Numerical results (S9 Fig) illustrate that the scientific value and waste of a multitiered lottery fall in between those of a proposal competition and a single-tiered lottery. Thus, a multitiered lottery offers an intermediate design that would partially reduce the waste associated with preparing proposals, while still allowing review panels to reward the best proposals with a higher probability of funding.
Our major result is that proposal competitions are inevitably and inescapably inefficient mechanisms for funding science when the number of awards is smaller than the number of meritorious proposals. The contest model presented here suggests that a partially randomized scheme for allocating funds—that is, a lottery—can restore the efficiency lost as paylines fall, albeit at the expense of reducing the average scientific value of the projects that are funded.
Why does a lottery disengage efficiency from the payline, while a proposal competition does not? For investigators, proposal competitions are, to a first approximation, all-or-nothing affairs, because an investigator only obtains a substantial payoff if her grant is funded. At high paylines (or, more precisely, when the number of awards matches the number of high-value projects), investigators with high-value projects can write proposals that win funding at modest cost to themselves. As the number of awards dwindles, however, competition stiffens. Depending on the details of the assessment process, an investigator with a high-value project must either work harder for the same chance of funding, or work just as hard for a smaller chance of funding. Either way, the return on her investment declines sharply. Thus, a contest is most efficient at the payline that weeds out low-value projects but does not attempt to discriminate among the high-value projects (e.g., S6 Fig). At lower paylines, however, the effort needed to signal which projects are most valuable begins to approach the value of those projects, making the funding program less worthwhile.
In a lottery, investigators do not compete for awards per se, but instead compete for admission to the lottery. The value to the investigator of being admitted to the lottery scales directly with the number of awards. It turns out that both the investigator's expected benefit and her costs of participation scale directly with the payline and thus the payline has no effect on efficiency. (In S1 Text, we follow Hoppe and colleagues [22] to show that this scaling can be explained by the economic principle of revenue equivalence.) If there are fewer awards than high-value projects, a lottery that weeds out the low-value projects but does not attempt to discriminate among high-value projects will facilitate scientific progress more efficiently than a contest.
Unfortunately, empirical comparisons between the efficiencies of funding competitions versus partial lotteries do not yet exist, to the best of our knowledge. However, two recent anecdotes support our prediction that the waste in proposal competitions is driven by the strategic dynamics of the contest itself. First, in 2012, the US National Science Foundation's Divisions of Environmental Biology and Integrative Organismal Systems switched from a twice-per-year, one-stage proposal competition to a once-per-year, two-stage competition, in part to reduce applicants' workload. However, the switch failed to reduce the applicants' aggregate workload meaningfully [30], and the two-stage mechanism was subsequently abandoned. Second, in 2014, the National Health and Medical Research Council of Australia streamlined the process of applying for their Project Grants, cutting the length of an application in half [31]. However, researchers spent more time, not less, preparing proposals after the process had been streamlined, both individually and in aggregate [31]. Both of these experiences are consistent with our prediction that, in a proposal competition, the effort that applicants expend is dictated by the value of funding to the applicants and the number of awards available, but does not depend on the particular format of the proposals.
A lottery is a radical alternative, and may be politically untenable [32]. If a lottery is not viable, an alternative approach to restoring efficiency is to design a contest in which the effort given to competing for awards has more direct scientific value. For example, a contest that rewards good science in its completed form—as opposed to rewarding well-crafted proposals that describe future science—motivates the actual practice of good science, and will be less wasteful at low paylines [9, 14]. Program officers could also be given the discretion to allocate some funds by proactively scouting for promising researchers or projects. Of course, a contest based on completed science or scouting has its own drawbacks, including rich-getting-richer feedback loops, a risk of new barriers to entry for investigators from historically underrepresented demographic groups, and the Goodhart's law phenomenon, whereby a metric that becomes a target ceases to be a good metric [33]. Nevertheless, it is tantalizing to envision a world in which the resources that universities currently devote to helping researchers write proposals are instead devoted to helping researchers do science.
This analysis also shows that extrascientific professional incentives to pursue grant funding can damage the scientific efficiency of a proposal competition. As many of these extrascientific incentives arise from administrators using grant success as a primary yardstick of professional achievement, perhaps one major benefit of adding explicit randomness to the funding mechanism would be to compel administrators to de-emphasize grant success in professional evaluations. Alternatively, to the degree that administrators value and reward grant success because of the associated overhead funds that flow to the university, funding agencies could reduce waste by distributing overhead separately from funding awards. Instead, perhaps overhead could be allocated based partially on the recent past productivity of investigators at qualifying institutions, among other possible criteria. Disengaging overhead from individual grants would encourage administrators to value grants for the science those grants enable (as opposed to the overhead they bring), while allocating overhead based on institutions’ aggregate scientific productivity would motivate universities to help their investigators produce good science.
Funding agencies often have pragmatic reasons to emphasize the meritocratic nature of their award processes. However, our model also suggests that downplaying elements of a funding competition's structure that introduce randomness to funding decisions can increase scientific waste. When applicants fail to recognize the degree to which the contest is already a lottery, they will overinvest effort in preparing proposals, to the detriment of science.
This model does not account for all of the costs or scientific benefits of a proposal competition, including the costs of administering the competition, the time lost to reviewing grant proposals, or the benefit of building scientific community through convening a review panel. Nonetheless, we suggest that the direct value of the science supported by funding awards and the disutility costs of preparing grant proposals are the predominant scientific benefits and costs of the usual proposal competition [13, 30], and provide a useful starting point for a more detailed accounting.
Our model also makes several simplifying assumptions, each of which may provide scope for interesting future work. First, researchers pay a time cost to prepare a proposal but receive money if the proposal is funded. In our model, we have converted both time and money into scientific productivity, in order to place both on a common footing. To be more explicit, though, scientific productivity requires both time and money (among other resources), and researchers may have vastly different needs for both. In S1 Text, we show that our model can be formally extended to encompass researchers' different needs for time and money if the marginal rate of technical substitution (that is, the rate at which time and money can be exchanged without altering scientific productivity) is exactly correlated with the project's scientific value. Our main results still hold in this case, as long as researchers with the best ideas do not value time so greatly that they write the weakest proposals. A more general exploration of researchers' heterogeneous needs for time and money—and of how researchers may adjust their portfolio of scientific activities when time or money is scarce—provide ample opportunity for future work.
Second, our model assumes that the distribution of the scientific value (v) across possible projects is exogeneous to the structure of the funding competition. This may not be the case if, for instance, a partial lottery encourages participation by investigators with unconventional views, reduces the psychological stigma of previous rejection [16], or discourages investigators either who have succeeded under the traditional proposal competition format or who perceive a lottery as riskier. In reality, such feedback loops may endogenize the distribution of v. Third, our model does not consider the savings that may accrue to investigators if they can submit a revised version of a rejected proposal to a different or subsequent competition. To a first approximation, submissions to multiple funders have the effect of increasing p, which can then be interpreted more generally as the proportion of ideas that get funded across all available funding programs. Iterations of revision and resubmission to the same funding program are likely to have more complex effects on efficiency and waste. Finally, our model is silent regarding whether many small or few large grants will promote scientific progress most efficiently, and is likewise silent about the factors that will influence this comparison.
To be sure, much more can be done to embellish this model. However, the qualitative results—that proposal competitions become increasingly inefficient as paylines drop and that professional pressure on investigators to pursue funding exacerbates these inefficiencies—are inherent to the structure of contests. Partial lotteries and contests that reward past success present radical alternatives for allocating funds and are sure to be controversial. Nevertheless, whatever their other merits and drawbacks, these alternatives could restore efficiency in distributing funds that has been lost as those funds have become increasingly scarce.
|
10.1371/journal.pbio.1001658 | Dectin-1 Is Essential for Reverse Transcytosis of Glycosylated SIgA-Antigen Complexes by Intestinal M Cells | Intestinal microfold (M) cells possess a high transcytosis capacity and are able to transport a broad range of materials including particulate antigens, soluble macromolecules, and pathogens from the intestinal lumen to inductive sites of the mucosal immune system. M cells are also the primary pathway for delivery of secretory IgA (SIgA) to the gut-associated lymphoid tissue. However, although the consequences of SIgA uptake by M cells are now well known and described, the mechanisms whereby SIgA is selectively bound and taken up remain poorly understood. Here we first demonstrate that both the Cα1 region and glycosylation, more particularly sialic acid residues, are involved in M cell–mediated reverse transcytosis. Second, we found that SIgA is taken up by M cells via the Dectin-1 receptor, with the possible involvement of Siglec-5 acting as a co-receptor. Third, we establish that transcytosed SIgA is taken up by mucosal CX3CR1+ dendritic cells (DCs) via the DC-SIGN receptor. Fourth, we show that mucosal and systemic antibody responses against the HIV p24-SIgA complexes administered orally is strictly dependent on the expression of Dectin-1. Having deciphered the mechanisms leading to specific targeting of SIgA-based Ag complexes paves the way to the use of such a vehicle for mucosal vaccination against various infectious diseases.
| Secretory IgA (SIgA) antibodies are secreted into the gut lumen and are considered to be a first line of defense in protecting the intestinal epithelium from gut pathogens. SIgA patrol the mucus and are usually known to help immune tolerance via entrapping dietary antigens and microorganisms and other mechanisms. SIgA, in complex with its antigens, can also be taken back up by the intestinal epithelium in a process known as reverse transcytosis. SIgA can thereby promote the uptake and delivery of antigens from the intestinal lumen to the Gut-Associated Lymphoid Tissues (GALT), influencing inflammatory responses. This reverse transcytosis of SIgA is mediated by specialized epithelial M cells. Because M cells possess the ability to take up antigens and are therefore important to the local immune system, they are a key target for the specific delivery of novel mucosal vaccines against various diseases. M cell receptors that take up the SIgA-antigen complexes, which serve as mucosal vaccine vehicles, represent an important aspect of this vaccine strategy. The identification of SIgA receptor(s) on the surface of M cells has, however, remained elusive for more than a decade. In this study, we now identify Dectin-1 and Siglec-5 as the key receptors for M cell–mediated reverse transcytosis of SIgA complexes. We further find that the glycosylation modification, and particularly sialylation, of SIgA is required for its uptake by M cells. We show that, when administered orally in complex with SIgA, the HIV p24 antigen is taken up in a strictly Dectin-1-dependent manner to stimulate a mucosal and systemic antibody response. These findings are considered important for understanding gut immunity.
| The mucosal immune system comprises the largest part of the entire immune system, and the mucosal surface represents the primary site of entry for pathogenic agents. SIgA has long been recognized as a first line of defense in protecting the intestinal epithelium from enteric pathogens and toxins. It is generally assumed that SIgA acts primarily through receptor blockade, steric hindrance, and/or immune exclusion. In recent years evidence has emerged indicating that SIgA promotes the uptake and delivery of Ags from the intestinal lumen to DC subsets located in gut-associated lymphoid tissues (GALTs), and influences inflammatory responses normally associated with the uptake of highly pathogenic bacteria and potentially allergenic antigens. This particular feature of SIgA, called reverse transcytosis, is mediated by epithelial M cells [1]. However, although the potentially useful properties of M cells on SIgA uptake are now well known, the receptor(s) whereby SIgA is taken up and transported by M cells remain(s) elusive.
SIgA reverse transcytosis was first invoked to account for the binding of rabbit SIgA to M cells in Peyer's patches (PPs) of suckling rabbits [2]. Colloidal gold particles coated with IgA were subsequently detected within M cell cytoplasmic vesicles and in the extracellular space of M cell pockets [3]. Endogenous SIgA was also shown to bind to human PP M cells in paraffin sections of human ileum [4]. In frozen sections, labeled SIgA could be visualized bound at the apical surface, in transit through intracellular vesicles, in the intraepithelial pocket, and on basolateral processes extending toward the basal lamina. In a mouse ligated ileal loop assay, mouse SIgA, human SIgA2, but not human SIgA1, bound to PP M cells [4]. Structural changes could explain the differences in reverse transcytosis between these subtypes. The IgA1 hinge features a 16 amino-acid insertion, lacking in IgA2, comprising a repeat of eight amino acids decorated with 3–5 O-linked oligosaccharides [5],[6]. Recombinant IgA1 with a deleted hinge region gained M cell binding function, which was interpreted as the M cell's binding site comprising both domains Cα1 and Cα2, juxtaposed in mouse IgA and human IgA2 [4]. Overall, IgA2 contains 4 N-glycosylation sites (Asn166, Asn263, Asn337, Asn459). In dimeric IgA, the Fc regions of the two monomers are linked end to end through disulfide bridges to the J chain [7]. IgA, with or without bound secretory component (SC), selectively adheres to the apical surfaces of mouse PP M cells [4].
To date, only a limited number of M cell receptors and their ligands have been identified, but most of these receptors are expressed in M cells and neighboring enterocytes as well. Some important pathogen recognition receptors, such as toll-like receptor-4, platelet-activating factor receptor, and α5β1 integrin have been identified on the surface of human and mouse M cells [8],[9]. The sialyl Lewis A (CA19.9) antigen lectin reacts with 80% of human M cells and, in contrast to the other ligands, binds only weakly to the enterocytes of the follicle-associated epithelium (FAE). Moreover, there is a wide variation in marker expression between M cells of different species and even between M cells at different portions of the intestine within the same species [10]. Indeed, M cells in murine, but not human, PP are preferentially bound with Ulex europaeus agglutinin–1 (UEA-1), a lectin specific to α-l-fucose residues [11]. A first mouse M cell–specific monoclonal antibody (mAb NKM 16-2-4) [12] displaying specificity for α(1,2)-fucose–containing carbohydrate moieties was produced. Glycoprotein 2 (GP2) was also shown to be specifically expressed on M cells of mouse and human PPs [13]–[15] and serves as an endocytic receptor for luminal antigens [16]. Another M cell marker, clusterin, is expressed in M cells and follicular DCs at inductive sites of human GALTs [14].
To date, the molecular partner(s) involved in SIgA reverse transcytosis has(have) not been identified in mice or in humans. In this work, we sought to map the structural feature(s) responsible for the selective interaction between murine SIgA and M cells. Since it is impossible to keep M cells in culture, one valuable approach consists in using cell culture models that mimic essential features of the FAE tissue. An in vitro model was used, based on the co-culture of polarized Caco-2 cells grown on inverted inserts and exposed to human Raji B lymphocytes [17],[18]. Following optimization in terms of functionality and reproducibility, we evaluated the transport of wild-type (wt) and mutant human IgA2 across newly differentiated M-like cells in comparison with other Ab isotypes. We found that glycosylation sites and in particular sialylation of the Cα1 region of IgA2 are required for M-like cell-mediated reverse transcytosis. We demonstrate for the first time that Dectin-1 expressed on the surface of M cells acts as a receptor involved in SIgA reverse transcytosis both in vitro and in vivo. Siglec-5 receptor seems also to participate in reverse transcytosis. Such a selective interaction has functional consequences in vivo, since targeting of HIV p24-SIgA complexes after oral delivery promotes the production of systemic and mucosal Ag-specific Abs in wt mice only, and not in Dectin-1 KO animals.
The model was adapted as described in the Methods section to optimize its reproducibility (Figure 1a). Prior to adding the lymphocytes, the tightness of the Caco-2 cell monolayer was checked by measuring transepithelial electrical resistance (TEER). The decrease in TEER observed after 5 d of co-culture is indicative of Caco-2 cell conversion into M cells (Figure 1b) [19] but not a result of the deterioration of tight junction organization, as reflected by preserved ZO-1 immunolabeling (Figure 1c), of either mono- or co-cultures.
M cells display a reduced brush border at their apical surface and an invaginated basolateral membrane, forming a pocket filled with immunoreactive cells [3]. Transmission electron microscopy shows that mono-cultures of Caco-2 cells exhibit a well-developed brush border with tightly packed microvilli, whereas in co-cultures with Raji cells, M-like cells characterized by the effacement of microvilli and enfolded lymphocytes are present (Figure 1d1). Moreover, the presence of desmosomes between M-like cells and the neighboring cells reveals their enterocytic origin (Figure 1d1, inset). Using scanning electron microscopy analysis, we observed in mono-cultures that all Caco-2 cells possessed a regular brush border and well-developed tight junctions, whereas in co-cultures, approximately 20–30% of Caco-2 cells expressed short and irregular microvilli (Figure 1d2) [20]. Immunolabeling of M-like cells with CA19.9 and enterocytes with UEA-1 (Ulex europaeus isoagglutinin I) [21] indicated a similar percentage of conversion, assuming a surface equivalence for M and Caco-2 cells (Figure 1e1). This was further verified by co-localization of human IgA2 with M cells labeled with CA19.9 mAb (Figure 1e2), in agreement with Mantis et al. [4].
To verify functional Caco-2 cell conversion into M cells, the transport of yellow/green-conjugated, 0.2 µm nanoparticles (NPs) across mono- and co-cultures was examined. NPs have previously been used to study transcytosis in various M-like cell models in vitro and in vivo [22]. The number of transported NPs recovered in the basal medium was 5.5-fold higher in the co-cultures, compared to Caco-2 cell mono-cultures (p<0.001) (Figure 1f). The sum of these data confirmed that the in vitro model of human FAE allowed efficient Caco-2 cells to M-like cell conversion to occur (20–30%), and importantly, with a high level of reproducibility.
Wt and truncated/mutated Ab constructs depicted in Figure 2 were cloned in the pGTRIO expression vector, stably transfected in CHO cells, produced in the culture supernatant, and purified by affinity chromatography as described in the Methods section. SDS-PAGE performed under reducing and nonreducing conditions confirmed the expected molecular weight for the light and heavy chains of the various constructs produced, and indicated assembly despite reduced formation of disulfide bridges between heavy and light chains (Figure 2), a feature commonly encountered while expressing IgA Abs in CHO cells [23].
One feature of M cells is their ability to transport a broad range of materials including Abs from the lumen to the underlying follicles. Specific retro-transport of Abs was compared between mono- and co-cultures using a luciferase (Luc)-IgA fusion protein. The Luc tag did not affect the Ab functionality (unpublished data) and allowed for sensitive quantification. As shown in Figure 3a, a significant exclusive transport of the IgA2 monomer (m-IgA2) across the cell monolayer harboring M-like cells was observed (p = 0.03). No significant transport of m-IgA1, IgG, or IgE was detected. Specificity of IgA reverse transcytosis was further confirmed in vivo by using a ligated murine intestinal loop. IgA positive cells were 30 times more abundant than IgG positive cells in PPs (Figure 3b). Dimerization, by incorporation of the J chain, or association with human SC did not modify IgA2 uptake by M-like cells (Figure 3a).
Next, mapping of regions and domains involved in IgA2 reverse transcytosis was performed with recombinant IgA2 lacking various portions of the heavy chain C-terminus (Figure 3a). IgA2 monomer depleted of Cα2, Cα3, and the tailpiece (m-IgA2 dCα2/3) crossed M-like cells as well as m-IgA2 wt, whereas IgA2 monomer depleted of the tailpiece (m-IgA2 dPB) and IgA2 monomer depleted of both the Cα3 and tailpiece (m-IgA2 dCα3) were not transported. The hinge region did not influence uptake as m-IgA2 dCα2/3 and m-IgA2 Cα1 (IgA2 with only the Cα1 constant region) gave similar results. Strikingly, M-like cell-mediated transport of IgA2 with only the Cα1 constant region was equivalent to wt m-IgA2. These results demonstrate that in the in vitro model, the Cα1 region of IgA2 is sufficient to allow reverse transcytosis through M-like cells.
Subclasses of human IgA are also different with respect to the number of N-glycosylation sites. In order to determine whether N-glycans present on IgA2 could influence their uptake and transport by M-like cells, transcytosis of a battery of constructs with engineered glycosylation sites was compared. As shown in Figure 4a, the efficiency of reverse transcytosis was highly dependent on the number of glycosylation sites. Indeed, there was a significant decrease in the transport of m-IgA2 G2, G1, G0, and m-IgA2 Cα1 G0 compared with m-IgA2. These findings were confirmed by enzymatic digestion of m-IgA2 by PNGase, an amidase that cleaves between the innermost GlcNAc and asparagine residues of high mannose, hybrid, and complex oligosaccharides from N-linked glycoproteins.
Sialic acid (Sia) can occur in different glycosidic linkages, most typically at the exposed, nonreduced ends of oligosaccharide chains attached to a wide variety of proteins like IgA [24]. To assess the function of Sia in IgA binding to M-like cells, IgA was exposed to neuraminidase, which has the capacity to selectively cleave the glycosidic linkages of neuraminic acids. An important and significant decrease in transport of IgA2 lacking Sia, resembling that measured for m-IgA2 G0 or m-IgA2+PNGase, was observed. Identical results were obtained with another recombinant IgA2 Ab molecule specific for CD20, with SIgA purified from colostrum and with plasma IgA treated by neuraminidase or PNGase (Figure 4b). The absence of remaining carbohydrates or Sia on the different IgA was verified by Western blot using labeling with lectins (Figure 4c), while the integrity of the IgA2 polypeptide following enzymatic treatment was verified by analysis on SDS-PAA gels (unpublished data). These results demonstrate the essential role of IgA glycosylation sites, and in particular, Sia in the reverse transcytosis of IgA2 by M-like cells.
As the above results provide solid evidence of the contribution of glycosylation to reverse transcytosis, we postulated that the IgA2 receptor of M-like cells is a glucan receptor. Blocking experiments were performed using a series of β-glucans, mono-, and disaccharides. A statistically significant decrease in IgA2 transport was observed in the presence of β-glucans including curdlan, laminarin, and zymosan (Figure 5a). No inhibition was observed with other members of the family or with mono- or disaccharides.
To further explore the possible involvement of glycans in IgA2 binding to M-like cells, Abs directed against the most common sugar receptors were used in blocking experiments. The use of an anti-Dectin-1 mAb targeting this β-glucan receptor led to an almost complete inhibition of IgA2 reverse transcytosis (Figure 5b). In contrast, blocking of the mannose receptor with an anti-CD206 mAb or of the lipopolysaccharide receptor with anti-TLR4 and anti-CD14 mAbs did not influence IgA2 transport. Consistently, the presence of Dectin-1 was observed on M-like cells present in co-culture conditions only (Figure 5c–e). Other receptors that have been described as being involved in IgA transport were evaluated. Transferrin receptor expressed by enterocytes (CD71) [25], which binds IgA1 Abs, did not block IgA2 passage, thus confirming the exclusive transport of IgA2 by M-like cells. Similarly, targeting of the human myeloid IgA Fc receptor (CD89) [26] with a specific mAb did not block the transport of IgA2 (Figure 5b).
To confirm the inability of desialylated IgA2 to target M-like cells in vitro, blocking experiments were also carried out using mAbs directed against various Siglecs, a family of receptors that specifically recognize Sia [27]. The unique involvement of Siglec-5 in IgA2 reverse transcytosis was demonstrated (Figure 5b), in contrast to all the other members of the family. The surface of M-like cells was Siglec-5+ in co-culture conditions only, with no labeling observed in mono-cultures (Figure 5c–e). Moreover, N-Acetylneuraminic acid severely affected SIgA2 reverse transcytosis (Figure 5a). Mabs to either Dectin-1 or Siglec-5 strongly inhibited transport of IgA2 in vitro, reaching up to 90% when added together (Figure 5b). Binding of monomeric IgA2 and SIgA to Dectin-1 and Siglec-5 was verified by ELISA using recombinant Dectin-1 and Siglec-5 as coating molecules (Figure 6a). In support of previous data, deglycosylated IgA2 and IgA1 were unable to recognize Dectin-1 and Siglec-5. Specificity of IgA2 recognition was further confirmed by immunofluorescence (Figure 6b) and flow cytometry (Figure 6c) using HEK cell transfectants expressing both Dectin-1 and Siglec-5. Importantly, co-localization between partners of the triad was observed in both types of analyses. Taken together, these results highlight the prominent role of Dectin-1 and Siglec-5 as receptors that mediate intestinal IgA2 reverse transcytosis.
To verify the validity of data obtained using the in vitro model of human FAE, SIgA transport was also analyzed in vivo in a mouse ligated intestinal loop containing a PP [28]. As shown in Figure 7a1 and 7a2, mouse SIgA-Cy3 was present on the surface of, and inside, UEA-1+ or GP2+ M cells, thus confirming the in vitro binding data in the in vivo context. Co-localization of mouse SIgA on Dectin-1+ cells in the FAE confirmed the role of Dectin-1 in SIgA binding in vivo as well (Figure 7b). In support of these data, in a Dectin-1 KO mouse model, we observed no co-localization between SIgA-Cy3 and UEA-1+ M cells and no reverse transcytosis of SIgA-Cy3 in PPs (Figure 7c). The interaction between Dectin-1 or Siglec-5 with IgA2 was similarly observed in human PPs. Immunolabeling with green-labeled IgA2 and red-labeled Dectin-1 or Siglec-5 of patient biopsies displayed specific co-localization between the Ab and Dectin-1 (Figure 7d) or Siglec-5 (Figure 7e). No specific immunofluorescence of secondary IgG Abs was obtained on human M cells (Figure 7f).
These results prompted us to compare the outcome of oral immunization in wt C57BL/6 mice and Dectin-1 KO mice using SIgA as an intestinal delivery system targeting M cells. As Dectin-1 is also expressed by DCs or macrophages, one can argue that such cells intercalating within the FAE may “pollute” the Dectin-1 signal on M cells in vivo. To solve this issue, Dectin-1 KO mice reconstituted with wt bone marrow cells (chimeric-KO:wt) and wt mice reconstituted with Dectin-1 KO bone marrow cells (chimeric-wt:KO) were immunized. Confirmation of the correct reconstitution in the chimeric mice was obtained by flow cytometry (Figure 8a/b) and immunofluorescence (Figure 8c) analysis on peripheral blood leukocytes. Positive control of immunization was obtained by subcutaneous administration of nanoparticulated vaccine polylactic acid (PLA)-p24, which induced strong immune response in mice (Figure 8d/e) [29].
HIVp24 was chosen as a vaccine candidate antigen for its relatively low molecular weight, thus reducing the risk of disturbing the overall structure of SIgA after covalent coupling. Administration of p24-SIgA in an intestinal ligated loop resulted in the presence of the complex in the SED region of PPs (unpublished data). Moreover, p24-SIgA complexes administered orally co-localized with Dectin-1+ cells in the FAE region (Figure 9a). Oral immunizations with p24-SIgA were performed in wt, Dectin-1 KO, chimeric-wt:KO, and chimeric-KO:wt mice as described in the Methods section. As intestinal immunization is well known to induce both mucosal and systemic responses [30], serum and feces samples were collected 1 wk after the last immunization. p24-specific IgG and IgA titers were measured following immunization of wt and chimeric-wt:KO mice only (Figure 9b and 9c). Moreover, the levels of p24-specific IgG and IgA responses in these mice were 25-fold higher than those obtained after oral immunization with the p24 polypeptide only. No antigen-specific response was measured in Dectin-1 KO and chimeric-KO:wt mice, thus confirming the essential role of Dectin-1 in SIgA reverse transcytosis.
Taken together, these results indicate that reverse transcytosis of the p24-SIgA complex is strictly Dectin-1-dependent and results in the potentiated passage of the hooked Ag, which is subsequently processed to trigger the onset of mucosal and systemic Ab responses.
In order to examine SIgA2 transport from the intestinal lumen to DCs located in the SED region of PPs, we took advantage of the recent demonstration that SIgA is recognized by DCs via the DC-SIGN receptor [31]. SIgA uptake by DCs was analyzed in vivo in a PP-containing ligated intestinal loop from wt mice with DC-SIGN-specific immunostaining and from CX3CR1-GFP transgenic mice. Figure 10a and 10b show specific localizations of p24-SIgA on DC-SIGN+ and SIgA-Cy3 on CX3CR1-GFP+ DCs present in the SED region. However, the strictly equivalent of DC-SIGN has not been described in mice, yet several homologues have been documented [32]. We assume that DC-SIGN-positive staining results from cross-reactivity with one of these murine homologues.
We next assess the relevance of these findings in the human in vitro system. HeLa transfectants stably expressing DC-SIGN added to the compartment bathing the basolateral pole of Caco-2 cells were used as surrogates of DCs populating the SED region of PPs [33]. Control of DC-SIGN expression was demonstrated by the inhibition of gp120 binding on HeLa-DC-SIGN+ cells by specific blocking mAbs (unpublished data). The binding of IgA2 that had previously crossed the monolayer containing M-like cells was observed by immunostaining of HeLa-DC-SIGN (Figure 10c), but not with wt HeLa cells used as a negative control. In another control, CHO expressing Langerin placed in the basolateral compartment did not bind transcytosed IgA2 (unpublished data). The transport of integral IgA2 through M cells, and also their preserved capacity to interact with DC-SIGN+ DC–expressing cells is another indication of the steps involved to ultimately lead to immune responses as detected above. These findings have also been confirmed by flow cytometry using human monocyte-derived DCs known to express DC-SIGN (Figure 10d).
The sum of these data shed light on the biochemical partners involved in reverse transcytosis of SIgA by PPs. SIgA is first taken up by M cells via the Dectin-1 receptor and/or Siglec-5, and is subsequently targeted to mucosal CX3CR1+ DCs bearing the DC-SIGN receptor. In the context of immune complexes, this process explains the functional production of mucosal and systemic Ab responses to the associated antigen.
M cells possess a high transcytotic capacity, allowing a wide range of materials to be transported including particulate Ags, soluble macromolecules, and pathogens. They are delivered from the intestinal lumen to inductive sites of the mucosal immune system. M cells are also the primary route through which SIgA are delivered to the GALT. Corthésy et al. have previously shown that after selective interaction with M cells, SIgA are targeted to DCs located in the SED region of PP, resulting in limited mucosal and systemic immune responses against a non-self-associated protein Ag [34]. Selective adherence to the apical surface of M cells is a prerequisite for efficient transepithelial transport, but the identity of receptors involved in SIgA endocytosis has remained elusive. In the current study, we investigated the transport of human SIgA2 across a model mimicking human FAE. At the level of IgA2, we provide evidence that both the Cα1 domain and associated glycosylation, more particularly Sia residues, are involved in M-like cell-mediated reverse transcytosis, while at the receptor level, both Dectin-1 and Siglec-5 have been identified as essential partner in the process. Finally, we validate our in vitro results upon analysis of murine and human tissues, ultimately demonstrating that Dectin-1/Siglec-5-mediated uptake of SIgA-based complexes results in productive mucosal and systemic antigen-specific Ab responses.
Initially, we studied reverse transcytosis of IgA2 across human M-like cells using a cell culture model that reproduces features of the FAE tissue. We confirmed that human IgA2, with or without J chain and/or bound SC, but not IgA1, IgG, or IgE, selectively bound to the apical surface of in vitro differentiated human M-like cells. Using a battery of deletion mutants, we demonstrated that domains Cα2 and Cα3 of IgA2 are dispensable to keep reverse transcytosis through M-like cells highly active. Low or absent transport of m-IgA2 dCα3 and m-IgA2 dPB comprising the Cα1 region suggests that subtle structural changes may affect optimal folding of these two particular recombinant proteins. Our in vitro results obtained with human cells do not totally correlate with the in vivo results of Mantis et al., who showed that both domains Cα1 and Cα2 were required for IgA binding to mouse PP M cells [4]. Differences in the expression systems for IgA constructs (deletion versus domain swab) and the glycosylation pattern may explain this discrepancy. It is conceivable that a critical density of glycans must be present to ensure uptake, as was recently described for Dectin-1 efficiently binding β-glucan polymers [35]. In conclusion, our study unequivocally demonstrates that IgA transport requires the presence of the properly glycosylated Cα1 domain within the Ab structure. The model opens the path toward in vitro assays of transport across reconstituted FAE, examination of the mechanisms of uptake, and investigation into vaccine or intestinal microbe delivery.
Sia residues on pathogens interact with Siglecs, which are expressed in the hemopoietic, immune, and nervous systems. Glycosylation patterns on pathogens are frequently used for adherence to, and passage across, the mucosal epithelium and in particular M cells in the FAE [36]–[38]. Similarly, it is conceivable that abundant carbohydrates located on the surface of SIgA may intervene in the process of selective recognition of M cells. The sum of our data confirms this working hypothesis, and demonstrates the prominent influence of glycosylation on the uptake of IgA2 by M cells. Additional experiments dealing with deletion of particular glycosylation sites and enzymatic desialylation allowed us to confirm the role of Sia residues in reverse transcytosis (Figure 4). Several members of the β-glucan superfamily were also identified as competitors of IgA2 transcytosis (Figure 5). This adds to the multiple functions of carbohydrates in SIgA including, for example, neutralization of bacterial toxins [39] and interaction with commensal bacteria [40].
Having unraveled the structural features responsible for the selective transport of SIgA in the reconstituted FAE model, we sought to identity the receptor(s) by which SIgA is taken up and transported by M cells. The use of blocking Abs against known IgA receptors including CD89 and CD71 did not prevent SIgA2 reverse transcytosis. These data, combined with the sufficient role of the Cα1 region of IgA2 in M-cell-mediated reverse transcytosis, led to the conclusion that no other known IgA receptor (pIgR, Fcα/μ receptor, and the asialoglycoprotein receptor) was involved in the process. Given the established involvement of Sia and β-glucan moieties, we speculated that the IgA2 receptor of M cells is a glucan- and/or Sia-receptor. Our work provides evidence of the presence of Dectin-1 on M-like cells, together with its involvement in reverse transcytosis of SIgA2. Dectin-1 is a type II transmembrane protein of the C-type lectin family, expressed by myeloid phagocytes (macrophages, DCs and neutrophils), which recognizes β-glucans in fungal cell walls and transduces signals triggering phagocytosis and the production of reactive oxygen species [41],[42]. In contrast, as recognition of soluble ligands by Dectin-1 does not lead to inappropriate activation signaling [35], its presence on M cells is consistent with simple SIgA capture and internalization.
Co-operation between Fc galactosylation and Dectin-1–inducing anti-inflammatory activities suggests that Dectin-1 is capable of working in combination with other partners in the cell plasma membrane. In view of the involvement of Sia in IgA2 reverse transcytosis via M cells, we investigated whether a Siglec receptor could serve this function. The majority of Siglecs, including CD33-related Siglecs like Siglec-5, appears to be naturally masked owing to cis-interactions with adjacent Sia. Unmasking of Siglecs can also occur in some cases by cellular activation or by exposure to sialidases. The unmasked Siglec would then be capable of de novo interactions with surrounding ligands in the environment. This could result in increased interactions with exogenous materials including glycosylated SIgA. Such a scenario of Siglec serving as a co-receptor has been reported in the case of HIV-1 entry mediated by CD4 in macrophages.
Preparation of murine duodenal ligated loops validated the results generated in the in vitro model of human FAE. This method has proven valuable in documenting the interaction of mouse IgA with PP M cells [4]. Tissue immunolabeling both confirmed the transport of SIgA2 by UEA-1+ and GP2+ M cells, and that of murine SIgA by Dectin-1 (Figure 7a and b). However, the absence of cross-reactivity of the anti-human CD170 mAb prevented us from confirming the role of Siglec-5 in the reverse transcytosis of SIgA in mice. Consistent with the in vitro data gathered in the model based on human cells, human biopsy analyses resulted in specific co-localization between IgA2-GFP and Dectin-1 (Figure 7d) or Siglec-5 (Figure 7e).
Finally, oral immunization of wt, Dectin-1 KO, or chimeric mice with p24-SIgA complexes unambiguously demonstrated that reverse transcytosis of SIgA is strictly dependent on Dectin-1 expressed on M cells. The further confirmation of the essential role of Dectin-1 in the in vivo context provides an explanation to the uptake of antigen-bearing SIgA by M cells, a feature resulting in systemic and mucosal immune responses [43]. The lack of a murine functional ortholog of human Siglec-5 prevented us from confirming the associated role of Siglec-5 in SIgA reverse transcytosis in vivo [44]–[46].
In vivo, the uptake of murine SIgA by murine CX3CR1+ DCs present in the GALT could also be documented (Figure 10b). In the SED region, CX3CR1+ DCs play a central role in antigen sampling [47]. In contrast to CD103+ DCs, CX3CR1+ cells represent a nonmigratory gut-resident population, which displays poor T-cell stimulatory capacity [48],[49]. In contrast to CD103+ DCs that serve classical DC functions and initiate adaptive immune responses in local lymph nodes, CX3CR1+ populations might modulate immune responses directly in the mucosa and serve as a first line barrier against invading enteropathogens. This supports the low activation properties of SIgA targeting antigen to DCs in the SED region [50],[51]. A recent study has shown that small intestine goblet cells function as passages delivering the low molecular weight soluble dextran (10 kDa) to CD103+ DCs [52], which promote IgA production, imprint gut homing on lymphocytes, and induce the development of regulatory T cells. As HIVp24 is administered in the form of a complex with SIgA (400 kDa), we believe that this pathway need additional characterization before it can be considered as operative for large molecules.
Transcytosis across M cells is known to enable the selective transport of particulate antigens in the absence of any assessable damage [53],[54]. This holds true for soluble SIgA, as the transcytosed Ab released by M cells in the human in vitro and murine in vivo models was still able to specifically target cells expressing DC-SIGN in the basolateral environment (Figure 10a/c/d). In mucosal tissues such as the rectum, uterus, and cervix, DC-SIGN is abundantly expressed by DCs present in the lamina propria and PPs, further substantiating the importance of the localization of DC-SIGN+ DCs as a first line of defense against viruses and pathogens. Delivery in the form of SIgA-based immune complexes may thus combine the onset of limited immune responses, which translates into the absence of spurious inflammatory reactions. Moreover, this receptor, by binding to ICAM-3, favors the generation of antigen-specific suppressive CD4+ T cells, which produce IL-10 [55], a cytokine that intervenes in both intestinal homeostasis and the production of local IgA.
This work defines Dectin-1 expressed on the surface of M cells as a receptor involved in SIgA reverse transcytosis both in vitro and in vivo. Besides bringing new information on the mechanism involved in SIgA retro-transport, deciphering the identity of such receptors may lead to the further development of mucosal vaccines targeting M cells. In future work, it will be critical to test the expression of Dectin-1 on other mucous membranes such as nasal/bronchial, endocervical, or buccal mucosa order to evaluate the broad applicability of this finding to active and passive immunization. As a perspective to future works, one can argue that intestinal villous M cells serving as an antigen gateway for the sampling of gut bacteria and inducing Ag-specific immune responses in a PP-independent manner [56] may contribute to SIgA reverse transcytosis as well.
Pullulan from Aureobasidium pullulans, mannan from Saccharomyces cerevisiae, α-Lactose, L-fructose, glycogen from bovine liver, sucrose, curdlan from Alcaligenes faecalis, laminarin from Laminaria digitata, and zymosan from Saccharomyces cerevisiae were all purchased from Sigma-Aldrich.
Anti-human Dectin-1/CLEC 7A polyclonal Ab (pAb) (goat IgG), anti-human CD14 mAb (mouse IgG1), anti-human TLR4 pAb (goat IgG), anti-human CD170 mAb (Siglec-5) (mouse IgG1), and anti-human CD329 mAb (Siglec-9) (mouse IgG2a) were all purchased from R&D Systems. Anti-human CD206 mAb (mouse IgG1) (mannose receptor) was purchased from Ozyme. Anti-human CD22 mAb (Siglec-2) (mouse IgG1), anti-human CD33 mAb (Siglec-3) (mouse IgG1), anti-human CDw328 mAb (Siglec-7) (mouse IgG1), and anti-human CD169 mAb (Siglec-1) (mouse IgG1) were purchased from AbD Serotec. Anti-human CD71 mAb (mouse IgG1) was purchased from Cliniscience. Anti-human CD89 mAb (mouse IgG1) was purchased from Abcam. All Abs were blocking and used according to the procedure provided by the manufacturer.
Yellow-green carboxylated or aminated latex particles (FluoSpheres) with a mean diameter of 0.2 µm were purchased from Molecular Probes.
Both the human intestinal cell line Caco-2 cell (clone 1) (obtained from Dr. Maria Rescigno, University of Milan-Bicocca, Milan, Italy) [57] and CHO cells were cultured in Dulbecco's modified Eagle's medium (DMEM) (PAA) supplemented with 10% (v/v) fetal bovine serum (FBS, Thermo-Fisher), 1% (v/v) nonessential amino-acids (PAA), and 1% (v/v) penicillin-streptomycin (PAA). The human Burkitt's lymphoma cell line Raji B (American Type Culture Collection) was cultured in RPMI 1640 supplemented with 10% (v/v) FBS, 1% (v/v) nonessential amino-acids, 1% (v/v) L-glutamine, and 1% (v/v) penicillin-streptomycin.
The inverted FAE model (Figure 1a) has been previously reported [20]. Several major changes were made and are listed below. Inverted Transwell polycarbonate inserts (12 wells, pore diameter of 3.0 µm, Corning) were coated with Matrigel, a basement membrane matrix (BD Biosciences) prepared in pure DMEM to a final protein concentration of 100 µg/ml for 1 h at room temperature. The coating solution was removed and inverted inserts washed with 300 µl of DMEM. Caco-2 cells (3×105), resuspended in 300 µl of supplemented DMEM, were seeded on the lower insert side and cultured overnight. The inserts were then inverted and placed in a 12-well culture dish and kept for 9 d. Raji B cells (5×105), resuspended in supplemented DMEM, were then added to the basolateral compartment of the Caco-2 cells, and co-cultures were maintained for 5 d. Mono-cultures of Caco-2 cells, cultivated as above but without the Raji B cells, were used as controls. Finally, the inserts were inverted in six-well plates, and a piece of silicon tubing (14×20 mm, Labomoderne) was placed on the basolateral side of each insert. Cell monolayer integrity, both in mono- and co-cultures, was controlled by measurement of TEER using an Endohm tissue resistance chamber (Endohm-12, World Precision Instruments) connected to a Millicell-ERS Ohmmeter (Millipore). The resistance of medium alone (9 Ω×cm2) was considered as background resistance and subtracted from each TEER value. Barrier function of the tight junctions was also analyzed by zonula occludens-1 (ZO-1) immunolabeling (see next section).
Cells morphologically similar to M cells were discriminated from Caco-2 cells using transmission electron microscopy (TEM) and scanning electron microscopy (SEM). TEM and SEM were used to evaluate morphological cell changes after co-culture with Raji cells. Mono- and co-cultures were washed twice in HBSS and fixed in 4% (v/v) formaldehyde. Ultra-thin sections of cell-covered filters were prepared for TEM analysis by standard methods, as previously described [58]. Observations were made using a Hitachi H-800 and a Digital camera Hamamatsu AMT XR40. Samples processed for SEM analyses were dehydrated, dried at critical point, and gold coated. Pictures of cell monolayers were obtained with a Thermo Noran Quest 2 L Hitachi S 3000N. Since no human-specific M cell markers have yet been identified, the microvilli-free morphology of M-like cells was used to identify and quantify them by SEM. Mono-cultures were used as controls.
Characterization and quantification of M-like cells in co-cultures was further verified by immunolabeling. Inserts were washed in HBSS to eliminate residual medium, incubated in 4% paraformaldehyde for 30 min, permeabilized with 0.1% Triton X-100 (Sigma-Aldrich), and blocked with PBS containing 5% FBS for 15 min at room temperature. Immunolabeling was performed using a combination of GFP-IgA2, anti-human ZO-1 mAb (Invitrogen), and mouse anti-human CA19.9 (Dako) [10]. Each reagent was diluted to 1/100, and incubated for 2 h at room temperature. 1/200 dilutions of secondary antibodies labeled with a fluorochrome were incubated for 1 h at room temperature. After two washes, inserts were air-dried, mounted with Fluoprep (BioMerieux), and observed by Immunofluorescence microscopy (Eclipse Ti, Nikon).
Nanoparticle (NP) (yellow-green fluorescent, 0.2 µm carboxylate-modified FluoSpheres beads) transport by polarized Caco-2 cells was evaluated in HBSS medium. NP concentration was adjusted to 4.5×109 NPs/ml and vortexed for 1 min to dissociate possible aggregates. NP suspension was added to the apical side of cell monolayers (400 µl) and the inserts were incubated at 37°C for 90 min. Basolateral solutions were then sampled and the number of transported particles was measured by flow cytometry (Facs Calibur, Becton Dickinson). The measurements were based on both fluorescence and particle size.
Light and heavy chain encoding genes from a human TNF-alpha–specific IgA Ab were cloned in a single vector (pGTRIO) designed for efficient Ab expression in HEK293 and CHO cell lines. pGTRIO is a derivative of pVITRO2 (Cayla-InvivoGen, Toulouse, France), a multigenic plasmid that contains two distinct transcription units. In pGTRIO, the antibiotic resistance gene is under the control of the EF1 alpha/HTLV promoter combined with the CMV enhancer that together constitutes a third transcription unit with the EF1 polyadenylation signal. The kappa constant region was cloned downstream of the FerL promoter together with the CMV enhancer, and the heavy chain constant regions were cloned downstream of the FerH promoter together with the human aldolase A enhancer. Unique restriction sites were introduced upstream of each constant region in order to allow the cloning of the variable region as SgrAI-BsiWI and AgeI-NheI fragments for VL and VH, respectively. All variable heavy chain regions were fused at the C-terminal end of secreted luciferase. CHO cells were transfected with pGTRIO constructs using the LyoVec system (Cayla-InvivoGen) in accordance with the manufacturer's instructions. Stable transfectants were selected in antibiotic-containing medium and screened for the production of Abs with a neutralizing activity on the HEK-Blue TNF-alpha/IL1-beta reporter cells (Cayla-InvivoGen) stimulated with TNF-alpha. IgA preparations were purified using Kappa affinity chromatography, IgG preparations were purified using protein G affinity chromatography, and IgE preparations were purified using protein L affinity chromatography. Ig-Luc constructs specific for TNF-alpha maintained their ability to block the cytokine, indicating proper assembly and folding. The following Abs were obtained by this method (Figure 2): human m-IgA2 (monomer); human GFP-IgA2 (monomer); human d-IgA2 (+ J chain - dimer); murine m-IgA (monomer); human IgE; human IgG1; human m-IgA1 (monomer); human m-IgA2 G0 (no glycosylation – monomer); human m-IgA2 G1 (1 glycosylation (Asn263)- monomer); human m-IgA2 G2 (2 glycosylations (Asn263 and Asn469) – monomer); human m-IgA2 dPB (without basal part – monomer); human m-IgA2 dCα3 (without basal part and Cα3 – monomer); human m-IgA2 dCα2/3 (without basal part, Cα3 and Cα2 – monomer); human m-IgA2 Cα1 G0 (only Cα1 without glycosylation); and human m-IgA2 Cα1 (only Cα1 with glycosylation). M-IgA2 was desialylated and deglycosylated with neuraminidase and PNGase, respectively (Enzymatic CarboRelease Kit, QA-Bio). Purity and assembly of the Abs were controlled by SDS-PAGE (Figure 2b/c). IgA deglycosylation was detected under normal conditions by using standard Western blot protocol with a combination of UEA-1-HRP and WGA-HRP lectins (Sigma-Aldrich).
For visualization of the mouse SIgA retrotranscytosis, a polymeric IgA Ab from the hybridoma clone IgAC5 specific to S. flexneri serotype 5a LPS [59] was obtained as previously described [60]. Purified free human SC was produced in Chinese hamster ovary cells [61]. SIgA molecules were obtained by combining in PBS pIgA molecules with a 2-fold excess of human SC for 2 h at room temperature according to the conditions described in the study by Rindisbacher et al. [62]. Cy3-SIgA molecules were obtained by conjugation with indocarbocyanine (Cy3) using the FluoroLink mAb Cy3 labeling kit (Amersham Biosciences) according to the procedure provided by the manufacturer.
Transport experiments were performed in HBSS at 37°C for 90 min with 10 µg of Ab conjugated with luciferase (Luc). Basolateral solutions were then recovered and the number of retro-transcytosed Ab-Luc measured by luminometry (Tristar LB941, Berthold Technologies) using the Gaussia Luc Assay Kit (Biolux) according to the procedure provided by the manufacturer. Ab-Luc transport was expressed as a mean value ± S.E.M. For inhibition experiments, cell monolayers were first preincubated apically with 5 mg of inhibitor in HBSS for 90 min at 37°C, and washed with HBSS, before adding the Ab-Luc suspension. All transport experiments were carried out in triplicate and were standardized with m-IgA2 (ratio RLU/µg of Ab).
Maxisorp 96-well plates were either coated with 50 µl of recombinant human Dectin-1/CLEC7A (5 µg/ml) (R&D Systems), 50 µl of recombinant human Siglec-5 (5 µg/ml) (R&D Systems), or 50 µl of an equal mixture of both Dectin-1 and Siglec-5 proteins and incubated O/N at 4°C. The wells were then washed three times with PBS and saturated with 200 µl of blocking solution (PBS+3% BSA) at room temperature for 1 h. The blocking solution was then discarded and 100 µl of m-IgA2, colostrum IgA, m-IgA2+PNGase, or m-IgA1 were added at a concentration of 5 µg/ml. After 1 h of incubation at room temperature, wells were washed three times with PBS, and bound IgA was detected using biotinylated goat anti-human IgA (Southern Biotech) followed by streptavidin-HRP (Amersham). Results are expressed as the means of OD ± SEM.
Six-week-old C57BL/6 mice were purchased from Charles River Laboratories (Lyon, France). CX3CR1-GFP transgenic mice were obtained from Maryline Cossin (Joseph Fourier University, France). Dectin-1 knockout mice [63], chimeric-KO:wt mice, chimeric-wt:KO mice, C57BL/6 mice, and CX3CR1-GFP transgenic mice were hosted at the University Hospital Unit for animal testing (Saint-Etienne, France). For ileal loop preparation, mice were starved overnight, anesthetized by intra-peritoneal injection of a mix of ketamine and xylazine (100 and 10 mg/kg animal weight, respectively), and kept warm at 37°C throughout the surgical procedure. We administered 100 µl of a 1 mg/ml solution of SIgA-Cy3 or p24-SIgA diluted in PBS into a 1.5-cm ileal loop containing a PP. Upon completion of the experiment, the mice were sacrificed by cervical dislocation and the piece of intestine was removed, extensively washed with PBS, fixed for 2 h in 3% paraformaldehyde, and included in optimal cutting tissue (OCT) embedding solution. We captured 7-µm sections (Leica cryostat model CM1950, Leica Microsystems) on Ultra+ superfrost microscope slides (VWR International) and stained for M cells. Slides were washed in PBS to eliminate residual OCT embedding solution, and blocked with PBS containing 5% FBS for 30 min at room temperature. Abs diluted to 20 µg/ml were incubated for 2 h at room temperature. The slides were then washed in PBS, air-dried, and mounted with Fluoprep (Biomérieux). Slides were observed by immunofluorescence microscopy (Eclipse, Nikon). Immunolabeling was performed using a combination of UEA-1-FITC (Sigma-Aldrich), anti-human Dectin-1/CLEC 7A pAb, anti-human CD170 mAb (R&D Systems), anti-human GP2 mAb (MBL), and p24-specific Ab directly labeled with PE (Santa Cruz Biotechnology). The protocol followed the guidance of the regional Ethics Committee for Animal Testing (CREEA) (Permit Number No. 69387487).
Informed and consenting patients who had undergone upper duodenal endoscopy for routine diagnostic purposes (e.g., dyspepsia and chronic diarrhea) with normal intestinal mucosa provided four to six biopsy samples from the distal duodenum. Biopsies were fixed for 2 h in 3% paraformaldehyde and included in OCT embedding solution, before being cryosectioned using a Leica cryostat model CM1950. We captured 7 µm sections on Ultra+ Superfrost microscope slides, and they were stained for M cells as described for mouse intestine and observed by immunofluorescence microscopy. Immunolabeling was performed using a combination of GFP-IgA2, anti-human Dectin-1/CLEC 7A pAb, and anti-human CD170 mAb (R&D Systems).
We housed 5–9-wk-old Dectin-1 KO and wt C57BL/6 males in individually ventilated cages at least 7 d prior to being irradiated. During this time and throughout the remainder of the experiment, animals were also maintained on sterile food and acidified in sterile water (containing 0.004% HCl). Animals received two doses of full body irradiation at 5 Gy (2×500 rads). Each dose was separated by a 3 h interval to limit gastrointestinal problems. Irradiated mice were returned to individually ventilated cages for 24 h. Bone marrow was isolated from the femurs and tibia of donor Dectin-1 KO and wt C57BL/6 males under sterile conditions in the absence of red blood cell lysis. Nucleated cells were counted on a haemocytometer. Irradiated Dectin-1 KO mice each received 2×106 total nucleated bone marrow cells from wt C57BL/6 mice intravenously via the lateral tail vein. The phenotype of these mice, named chimeric-KO:wt, is thus wt at the systemic level and Dectin-1 KO at the mucosal level. Irradiated wt C57BL/6 mice were similarly injected with the same number of Dectin-1 KO donor cells. The phenotype of these mice, referred to as chimeric-wt:KO, is Dectin-1 KO at the systemic level and wt at the local level. Animals were maintained in individually ventilated cages as described above for a further 6 wk. Five weeks after bone marrow injections, 50 µl of tail vein blood was taken from each animal to characterize the cell phenotype, and red blood cells lysed for 2 min at room temperature in 1× Pharmlyse buffer. Cells were then washed twice in phosphate-buffered saline and counted. Cells were incubated for 15 min in FACS block (HBSS+2 mM NaN3, 0.5% BSA, and 5% heat inactivated rabbit serum) containing 6 µg/ml Fc-receptor blocking mAb (clone 24G2), prior to addition of 10 µg/ml biotinylated anti-Dectin-1 mAb (clone 2A11) or the rat biotinylated isotype control IgG2b for 30 min on ice. After three washes in FACS wash (HBSS complemented with 2 mM NaN3 and 0.5% BSA), cells were incubated in FACS block containing 1/200 APC-conjugated streptavidin (Invitrogen) for 20 min on ice. Cells washed three times were analyzed on a FACSCalibur (Becton-Dickinson) and data analyzed using FlowJo software.
Statistical analyses were performed using the InStat version 2.01 from the GraphPad Software, and the unpaired two-tail Mann–Whitney U test was applied. Significance limit was set at p≤0.05.
|
10.1371/journal.pgen.1007633 | Glycine promotes longevity in Caenorhabditis elegans in a methionine cycle-dependent fashion | The deregulation of metabolism is a hallmark of aging. As such, changes in the expression of metabolic genes and the profiles of amino acid levels are features associated with aging animals. We previously reported that the levels of most amino acids decline with age in Caenorhabditis elegans (C. elegans). Glycine, in contrast, substantially accumulates in aging C. elegans. In this study we show that this is coupled to a decrease in gene expression of enzymes important for glycine catabolism. We further show that supplementation of glycine significantly prolongs C. elegans lifespan, and early adulthood is important for its salutary effects. Moreover, supplementation of glycine ameliorates specific transcriptional changes that are associated with aging. Glycine feeds into the methionine cycle. We find that mutations in components of this cycle, methionine synthase (metr-1) and S-adenosylmethionine synthetase (sams-1), completely abrogate glycine-induced lifespan extension. Strikingly, the beneficial effects of glycine supplementation are conserved when we supplement with serine, which also feeds into the methionine cycle. RNA-sequencing reveals a similar transcriptional landscape in serine- and glycine-supplemented worms both demarked by widespread gene repression. Taken together, these data uncover a novel role of glycine in the deceleration of aging through its function in the methionine cycle.
| There is a growing number of studies showing that amino acids function as signal metabolites that influence aging and health. Although contemporary -OMICs studies have uncovered various associations between metabolite levels and aging, in many cases the directionality of the relationships is unclear. In a recent metabolomics study, we found that glycine accumulates in aged C. elegans while other amino acids decrease. The present study shows that glycine supplementation increases lifespan and drives a genome-wide inhibition effect on C. elegans gene expression. Glycine as a one-carbon donor fuels the methyl pool of one-carbon metabolism composed of the folate and methionine cycles. We find that the glycine-mediated longevity effect is fully dependent on the methionine cycle, and that all of our observations are conserved with supplementation of the other one-carbon amino acid, serine. These results provide a novel role for glycine as a promoter of longevity and bring new insight into the role of one-carbon amino acids in the regulation of aging that may ultimately be beneficial for humans.
| Aging is characterized by a progressive deterioration of the functional capacity of tissues and organs. Pioneering studies in the nematode C. elegans have identified longevity-associated genes and provided us with great insights into the plasticity of aging [1–3]. In the last few decades, genetic and nutritional interventions have been employed in multiple organisms including Saccharomyces cerevisiae, C. elegans, Drosophila melanogaster, rodents, and more recently fish [4–6]. These models have set the stage for characterizing the genetic basis of physiological aging and for developing efficient strategies to control the rate of aging.
To date, metabolic pathways including the mTOR, insulin/IGF-1, and AMP-activated protein kinase (AMPK) signaling pathways have emerged as playing a critical role in aging [reviewed in [7]]. Several studies demonstrate that the levels of specific amino acids effectively influence lifespan by affecting these pathways. For example, the branched-chain amino acids valine, leucine, and isoleucine when administered to C. elegans can function as signaling metabolites that mediate a mTOR-dependent neuronal-endocrine signal that in turn promotes a longer lifespan [8]. Moreover, inhibition of threonine and tryptophan degradation also contributes to lifespan extension by enhancing protein homeostasis in C. elegans [9,10]. Additionally, restriction of methionine extends the lifespan of flies in a mTOR-dependent manner [11]. However, supplementation of other amino acids such as methionine, serine, glycine, histidine, arginine, and lysine have been shown to promote lifespan in C. elegans by mechanisms that are to date not known [12].
In addition to the mTOR signaling pathway, alterations in one-carbon metabolism involving the folate and methionine cycles couple amino acid metabolism to the regulation of human health and disease [13]. Glycine, as one of the input amino acids that feeds into one-carbon metabolism, provides a single carbon unit to the folate cycle to yield a variety of one-carbon bound tetrahydrofolates (THFs) [14]. These function as coenzymes in methylation reactions including the production of methionine through methionine synthase (METR-1 in C. elegans) as well as the universal methyl donor, S-adenosylmethionine (SAMe) through S-adenosyl methionine synthetase (SAMS-1 in C. elegans) [14]. These output metabolites of one-carbon metabolism support a range of biological functions [14]. In C. elegans, mutations in the metabolic gene sams-1 and the levels of SAMe and S-adenosylhomocysteine (SAH) have been implicated in the regulation of aging [15,16]. Although the underlying mechanism of how SAMe/SAH status influences aging needs further investigation, studies in vivo have provided evidences that the level of SAMe couples with the trimethylation status of lysine 4 on histone H3 (H3K4me3) and affects gene regulation [17]. Another study in mouse pluripotent stem cells demonstrates that threonine catabolism contributes one carbon to SAMe synthesis and histone methylation through glycine cleavage pathway [18]. Of particular note, several histone methyl-transferases and de-methyltransferases in C. elegans have been identified as longevity regulators [19–21]. Taken together, these studies all suggest that altering one-carbon metabolism is a mechanism that controls the aging process.
We recently showed that glycine accumulates with age in a large scale metabolomics study profiling levels of fatty acids, amino acids, and phospholipids across the lifespan of C. elegans, [22]. Another study in human fibroblasts suggested that epigenetic suppression of two nuclear-coded genes, glycine C-acetyltransferase (GCAT) and serine hydroxymethyltransferase 2 (SHMT2) which are both involved in glycine synthesis in mitochondria, was partly responsible for aging-associated mitochondrial respiration defects [23]. This study went on to report that glycine treatment rejuvenated the respiration capacity of fibroblasts derived from elderly individuals [23]. However, to date the role of glycine has not been systematically defined in animal models of longevity. In this study, we build upon our previous observations that suggest glycine accumulation in aging animals may play a unique and as-of-yet unexplored role in the regulation of eukaryote lifespan.
We previously measured amino acid levels throughout the life of C. elegans, including four larval phases (L1-L4) and ten days of adulthood from young worms to aged ones (days 1–10) [22]. We reported that the concentrations of most amino acids peaked at the later larval stage or early adult phase and then began declining at different adult stages, reaching low levels by the latest stages of the animals’ life [22]. One stark exception was glycine, which continued to accumulate in aged worms [22].
To determine if the accumulation of glycine with age is due to increased synthesis or reduced degradation, we measured the expression levels of genes directly involved in glycine metabolism in worms collected at different ages (Fig 1A). Specifically, we observed that the expression levels of most genes in glycine degradation and consumption pathways including glycine decarboxylase (gldc-1), glycine cleavage system H protein (gcsh-1), phosphoribosylamine-glycine ligase (F38B6.4), and D-amino acid oxidase (daao-1) were dramatically lower at day 9 of adulthood (D9) (Fig 1B). In contrast, the expression levels of most genes involved in glycine synthesis including threonine aldolase (R102.4), serine hydroxymethyltransferase (mel-32), alanine-glyoxylate aminotransferase 2 (T09B4.8), and alanine-glyoxylate amino transferase (agxt-1) remained unchanged in aged worms (Fig 1C). These data suggest that the accumulation of glycine observed in aged worms is predominantly due to a reduction in the expression of genes required for its degradation.
To gain a better understanding of the significance of glycine accumulation in aged animals, we next asked whether this phenomenon is prevalent among long-lived worms such as daf-2(e1370) and eat-2(ad465), the C. elegans models of impaired insulin signalling [2] and dietary restriction [24], respectively. To characterize the changes in the levels of glycine during aging in the daf-2(e1370) and eat-2(ad465) mutant lines, we measured glycine levels in long-lived worms at young and old stages, specifically L3 (larval stage 3) and day 10 (D10) of adulthood. Interestingly, the levels of glycine in daf-2(e1370) and eat-2(ad465) at D10 were significantly increased relative to both their levels at L3 (S1 Fig). These results suggest that there might be a generic regulatory mechanism mediating the level of glycine during aging in both wild type and long-lived C. elegans.
To confirm the metabolic branch points of glycine metabolism in the control of glycine levels in C. elegans, we subjected worms to RNAi against the genes in glycine metabolism (Fig 1A) at the time of hatching, and then measured the level of glycine in D1 worms. Worms treated with RNAi against the genes in glycine synthesis pathways showed no effect on the levels of glycine (Fig 1D). This is perhaps because glycine from bacteria may compensate for the reduction of glycine synthesis in worms. Interestingly, a notable exception to this pattern was found for the knockdown of mel-32, encoding a worm homologue of mammalian SHMT1 and SHMT2, which acts to interconvert serine and glycine in one-carbon pathway (Fig 1A) [25]. mel-32 RNAi led to a strong increase in the level of glycine (Fig 1D) and a concomitant slight decrease in the level of serine (S2 Fig). Thus, these results indicate that MEL-32 is prone to synthesize serine from glycine in C. elegans (Fig 1D). Likewise, RNAi of T25B9.1 also led to a subtle increase of glycine in worms, suggesting that T25B9.1 tends to degrade glycine in C. elegans (Fig 1D). RNAi knockdown of genes in glycine catabolic pathways including daao-1, gss-1, gcsh-1, glycine cleavage system T-protein (gcst-1), and dihydrolipoamide dehydrogenase (dld-1), all significantly increased endogenous glycine compared to control (Fig 1E). Surprisingly, RNAi of phosphoribosylformylglycinamidine synthase (pfas-1), which is thought to block glycine being used in purine synthesis, was found to lower the level of glycine (Fig 1E). This implies that there are complex metabolic consequences from the perturbation of de novo purine synthesis. Collectively, these data suggest that the majority of glycine in C. elegans is influenced by two metabolic branches, namely one-carbon metabolism via glycine cleavage complex and serine synthesis via MEL-32.
We next verified the effects of glycine on lifespan by administering various concentrations of glycine to worms. To avoid influences of glycine on bacterial metabolism and vice versa, we killed E. coli OP50 with a combination of ultraviolet (UV)-irradiation and antibiotic (carbenicillin) supplementation.
In line with a previously reported observation [26], worms being fed UV- and carbenicillin-killed E. coli OP50 (referred to hereafter as “killed bacteria”) live significantly longer compared to those being fed live E. coli OP50 (S3 Fig). Therefore, to confirm if glycine still accumulates in aged worms after switching to a killed bacteria diet, we again quantified amino acids levels in worms at different stages of C. elegans lifespan including L3 (larval stage 3), day 1 (D1), day 3 (D3), day 6 (D6) and day 9 (D9) of adulthood. We found that most amino acids remained unchanged from L3 to D9, including valine, tryptophan, lysine, isoleucine, glutamine, asparagine, aspartate, arginine, serine, and proline (S4A Fig). Some amino acids change with age such as leucine, methionine, ornithine, and glutamate (S4A Fig). The levels of these either peak at the L3 stage or at D3, then decrease with age (S4A Fig). Interestingly, the level of tyrosine peaked at both L3 and D9, and the level of alanine peaked at D3, then maintained stable from D3 to D9 (S4A Fig). Although the levels of some of these amino acids in worms fed killed bacteria were more stable with age compared to those in worms fed live OP50 [22], we consistently found that the levels of leucine, methionine, ornithine, and glutamate decreased and the level of glycine increased in aged worms (Fig 2A and S4A Fig). The results suggest that the changes of these amino acids during aging are robust phenotypes that are independent of the worms being fed live or killed bacteria.
On killed bacteria we tested how a range of glycine concentrations from 5 μM to 10 mM affects the lifespan of C. elegans. We observed a significant increase in median lifespan at concentrations of 5 μM, 50 μM, and 500 μM of dietary glycine as compared to untreated controls, with a 7.7% (p < 0.0001), 19.2% (p < 0.0001), and 19.2% (p < 0.0001) extension observed, respectively (Fig 2B). Higher concentrations, however, including 5 mM and 10 mM, failed to extend worm lifespan suggesting a dose response where only low concentrations of glycine between 5–500 μM are beneficial to lifespan (Fig 2B, S4B Fig). Our findings are in agreement with previous studies suggesting a dose-effect of amino acids on worm lifespan [12].
Next, we measured amino acid levels in C. elegans at D1 of the adult stage to investigate how glycine supplementations at different doses affect glycine levels in vivo. We did not detect obvious changes of glycine abundance itself or of the other amino acids at any concentration of supplementations such as serine and threonine, the substrates for glycine synthesis (S5A–S5E Fig). These results suggest that in worms, glycine immediately fuels metabolic pathways, likely through glycine cleavage complex and MEL-32 (Fig 1D and 1E). Moreover, the results suggest that the levels of amino acids are tightly controlled in the early adult stage when glycine metabolic genes are actively expressed (Fig 1B and 1C).
To test whether a specific timing is required for the beneficial effect of glycine on C. elegans lifespan, we administered glycine to worms at different times in the animal’s life. These times included (a) the development phase from the time of hatching to L4, (b) the early adult period from D1 to D3, (c) the adulthood starting from D1 until death, and (d) during the entire life (Fig 2C). Intriguingly, the longevity-promotion function of glycine was observed to act exclusively during adulthood, especially the early adult phase, as supplementation from D1 to D3 was sufficient to prolong lifespan by 19.0% (p < 0.0001) (Fig 2D). In contrast, glycine supplementation during development did not exert a beneficial effect on lifespan (Fig 2D). These data suggest that the first three days of adulthood are important for glycine to confer its longevity effects in C. elegans.
RNAi knockdown of mel-32 or the components in glycine cleavage complex (e.g. gcst-1) resulted in pronounced increases of the endogenous levels of glycine in C. elegans (Fig 1D and 1E). As such, we wanted to determine whether such increase by blocking two distinct metabolic branches also impacts lifespan in C. elegans. We hence assayed the lifespan of animals fed RNAi targeting mel-32 and gcst-1 from the time of hatching. Intriguingly, we found that while RNAi of gcst-1 increased the endogenous level of glycine 6.9-fold (Fig 1E), it failed to increase lifespan in C. elegans (Fig 2E). In contrast, mel-32 RNAi not only increased the endogenous level of glycine robustly (6.7-fold) (Fig 1D), but also improved the survival of wild-type N2 C. elegans significantly (Fig 2F). Thus, these results suggest that elevated endogenous or exogenous glycine levels drive lifespan extension, and that an intact glycine cleavage complex is likely required for this extension.
As glycine is an important one-carbon donor and an essential substrate for de novo purine synthesis, we next sought to understand how one-carbon and purine metabolism are influenced by aging. We therefore quantified the expression levels of genes involved in one-carbon metabolism (Fig 3A) at five developmental distinct time points (L3, D1, D3, D6, and D9) in the worm lifespan. We found that with the exception of tyms-1 and dhfr-1 (thymidylate synthetase and dihydrofolate reductase) which were upregulated in aged worms, the expression of genes participating in transferring the one-carbon moiety of glycine to form SAMe including gcst-1 (glycine cleavage system T-protein), dld-1 (dihydrolipoamide dehydrogenase), gcsh-1 (glycine cleavage system H-protein), mthf-1 (methylene tetrahydrofolate reductase), metr-1 (methionine synthase), and sams-1 (S-adenosyl methionine synthetase), dropped dramatically in aged animals (Fig 3B) [27]. Furthermore, the de novo purine synthesis genes atic-1 (5-aminoimidazole-4-carboxamide ribonucleotide formyltransferase/IMP cyclohydrolase) (Fig 3B) and F38B6.4 (Fig 1C) were markedly reduced in aged worms compared to their expression levels in worms at D1. Together these data suggest a downregulation of both one-carbon and purine metabolism during aging in C. elegans.
To resolve in greater detail how glycine supplementation counteracts the age-related changes in one-carbon and purine metabolic genes, we performed next generation RNA-sequencing on D1 worms which were fed control diet (UV-killed E. coli OP50) and 500 μM glycine-supplemented diet from the time of hatching, respectively. In contrast to the changes occurring with age, glycine induced a marked increase in the expression of genes in purine metabolic pathways (Fig 3C). Concomitantly, several genes in one-carbon metabolism (indicated in red in S6 Fig) were differentially expressed, including upregulation of atic-1, F38B6.4, dhfr-1, tyms-1, mel-32, and gcsh-1, and downregulation of sams-1, mthf-1, and gldc-1 (S6 Fig). These results suggest that the expression of purine metabolism genes is subjected to an overall upregulation by exogenous supplementation of dietary glycine, while genes in one-carbon metabolism are under complex regulations upon glycine supplementation.
Having determined one of the effects of glycine on worm lifespan and its ability to partly counteract age-related declines in gene expressions in one-carbon and purine metabolic pathways, we next aimed to investigate whether similar gene regulatory events are also present in long-lived mutant worms. To test this, we turned to microarray datasets from three long-lived worm models, daf-2(e1370) and eat-2(ad465) [28] or mrps-5 RNAi worms (reported here). We specifically looked at glycine-associated metabolic pathways in the Kyoto Encyclopedia of Genes and Genomes (KEGG) database including ‘glycine serine and threonine metabolism’, ‘one carbon pool by folate’, ‘cysteine and methionine metabolism’, and ‘purine metabolism’. Strikingly, although distinct longevity pathways are known to be active in these long-lived worms [2,29,30], all these longevity worm models consistently showed a transcriptional activation of glycine metabolism, folate-dependent one-carbon metabolism, and methionine metabolism (S7A–S7C Fig). This suggests that transcriptional activation of glycine and one-carbon metabolism is a prominent signature of longevity shared by these long-lived worms. In contrast, overall purine metabolism including de novo synthesis, the salvage pathways, and purine degradation was only mildly deactivated across the three long-lived strains (S7A–S7C Fig).
To identify prominent transcriptional features in the three long-lived strains, we next examined the expression profile of individual genes belonging to “glycine, serine and threonine metabolism”, as well as folate-mediated one-carbon metabolism and purine metabolism. We found that genes involved in glycine anabolism including T09B4.8, agxt-1, C15B12.1, R102.4, and T25B9.1 were upregulated in daf-2(e1370), eat-2(ad465) [28] and mrps-5 RNAi worms (S7D–S7F Fig). Moreover, a concomitant rise in the expression levels of genes in glycine catabolism and de novo purine synthesis occurred, including gldc-1, gcst-1, gcsh-1, F38B6.4, and atic-1 (S7D–S7I Fig), suggesting a stimulation of metabolic activity of glycine-associated processes in these long-lived worms. To query the expression of genes in the production of SAMe in the long-lived worm models, we specifically checked the expression of five homologues of SAMe synthetases in C. elegans from the microarray data including sams-1, sams-2, sams-3, sams-4, and sams-5 (S8 Fig). Interestingly, the expression of sams-1, which encodes the enzyme accounting for the majority of overall SAMe production in C. elegans [31], was significantly upregulated in all these long-lived worms (S8 Fig). Moreover, sams-5 was increased in daf-2(e1370) and eat-2(ad465) [28], while sams-2, sams-3, and sams-4 were suppressed in mrps-5 RNAi and daf-2(e1370) (S8 Fig). Collectively, the data point to elevated methylation activities in these long-lived worms.
In contrast to an overall upregulation of purine metabolism genes in response to dietary glycine treatment, all three longevity worm models specifically induced the expression of genes in de novo purine synthesis compared to control, including F38B6.4, F10F2.2 (pfas-1, phosphoribosylformylglycinamidine synthase), B0286.3 (pacs-1, phosphoribosylaminoimidazole succinocarboxamide synthetase), and atic-1 (S9A–S9C Fig).
Taken together, our data reveal that transcriptional activations of glycine and glycine-associated pathways, including one-carbon and de novo purine synthesis, are present in three distinct longevity models.
To understand the mechanism of glycine-mediated lifespan extension on a more global scale, we returned to our next-generation RNA-sequencing dataset and performed unsupervised Principle Component Analysis (PCA) on the individual libraries. We found a clear separation between glycine-treated versus untreated samples (Fig 4A), corresponding to a large difference in gene expression (Fig 4B). Interestingly, more genes were transcriptionally repressed in response to glycine treatment in which 2629 genes were differentially down-regulated, and 983 genes were up-regulated compared to control (Fig 4B). This suggests an inhibition propensity of glycine on gene expression.
To probe the processes changed upon glycine supplementation, we performed gene ontology (GO) term enrichment analysis on the differentially expressed genes using the Database for Annotation, Visualization and Integrated Discovery (DAVID) bioinformatics resource [32]. A larger number of GO terms were found enriched in the downregulated gene list, among which were GO terms for body morphogenesis, growth regulation, post-embryonic development, molting cycle, cuticle development, multicellular organism growth, and positive regulation of growth rate (Fig 4C). These enrichments are all related to growth control, suggesting a decelerating effect of glycine upon development and growth which is a phenomenon known to be associated to longevity [33]. Additionally, unlike some mutations that confer longevity to the soma at the cost of a reduction in fecundity [34], supplementation of glycine mildly enhanced the expression of genes in reproduction-related biological processes such as the GO terms of vitellogenesis, meiosis cell cycle, and gamete generation (Fig 4D). While we did not observe clear differences in the number or the size of embryos, the progenies from glycine- and serine-supplemented worms seem to be healthy and normal. Overall, with the observations of the lifespan extending effects of glycine, these data suggest that the beneficial role of glycine slows down pathways that are traditionally ameliorated in healthy aging models.
MEL-32 and glycine cleavage complex are major gatekeepers responsible for the flux of glycine into metabolic pathways (Fig 1A). Particularly, although suppression of mel-32 or the components in glycine cleavage complex in C. elegans dramatically elevated the level of endogenous glycine (Fig 1D and 1E), only mel-32 RNAi showed a potent lifespan extension effect on worms likely by favoring the flux of glycine into one-carbon metabolism (Fig 2E and 2F). Additionally, by fueling the one-carbon metabolic network through glycine cleavage complex, glycine contributes to the synthesis of SAMe (Fig 3A), the availability of which has been implicated in the regulation of histone methylation patterns and subsequently gene expressions [35]. Collectively, this led us to hypothesize that the methionine cycle may be required for the lifespan-extending effect of glycine. To test if the methionine cycle is necessary for the longevity effect of glycine, we performed lifespan analyses with the methionine cycle-deficient mutants metr-1(ok521) and sams-1(ok3033). In these mutants, 500 μM glycine failed to promote lifespan (Fig 5A–5C), demonstrating that the effects of glycine on worm longevity depend on the methionine cycle.
Serine is another important one-carbon donor and the major precursor for glycine synthesis in vivo [14]. Thus, serine and glycine are closely related to each other in one-carbon metabolism. Given their similarities, we next investigated whether serine can also exert beneficial effects on worm lifespan. Serine supplementation at a concentration from 1 mM to 10 mM has been shown previously to extend worm lifespan [12]. We therefore administered 5 mM serine to worms and measured lifespan. Similar to glycine, we confirmed that serine prolonged the lifespan of worms (+20.8% in median lifespan) (Fig 6A).
Given that the adulthood is important for glycine-mediated lifespan extension, we investigated whether serine increases lifespan in the same fashion as glycine. Similarly, we treated worms with serine at different times in worm’s lifespan including (a) developmental stage, (b) the beginning of adulthood from D1 to D3, (c) adulthood from D1 until death, and (d) during the whole lifetime. Similar to the effects of glycine on lifespan, serine treatment from D1 to D3 is sufficient to recapitulate the beneficial effects on lifespan as did the treatment throughout the entire life or during adulthood only (Fig 6B). In contrast, supplementation during the developmental stage failed to increase lifespan (Fig 6B). This further implies that serine acts on the same downstream longevity signalling pathways to influence aging as does glycine.
We further investigated whether the anti-aging effects of serine also rely on the methionine cycle. In agreement with the results observed with glycine supplementation, the lifespan extending effect of serine was also ablated by mutations of metr-1 and sams-1 (Fig 6C and 6D). These results further suggest that serine and glycine prolong C. elegans lifespan via a similar mechanism.
To determine the common regulators in both glycine and serine-mediated longevity, we performed RNA-sequencing on serine-supplemented worms. As expected, PCA analysis showed a clear separation between worms treated with either of the amino acids when compared to non-treated worms, and a strong similarity between glycine- and serine-treated worm groups (Fig 6E). Statistical analysis found one significantly differentially expressed gene, F38B6.4 (S10A Fig), an enzyme that consumes glycine for purine synthesis. In addition, visualizing the data as a volcano plot showed again a greater number of genes repressed by serine treatment (2865 downregulated genes vs 973 upregulated genes), in line with the same gene expression suppression pattern of glycine-supplemented worms (S10B Fig, Fig 4B). Likewise, we found a strong overlap between glycine- and serine-treated worms when looking at the up- and downregulated genes, as shown in the Venn diagrams where 82.8% (2335) of the downregulated and 72.3% (804) of the upregulated genes are shared (Fig 6E–6G). Taken together, we suggest a model whereby both glycine and serine supplementation stimulate longevity in a methionine cycle-dependent fashion and through common signaling pathways. Moreover, this seems to be dependent on the expression of sams-1 and metr-1. This model is illustrated in Fig 6H.
Our work sheds light onto the means by which the amino acid glycine can increase C. elegans lifespan when supplemented to the diet. Using a metabolomics approach, we found that glycine steadily and significantly accumulates in aging C. elegans [22]. Furthermore, we demonstrated that this accumulation is mainly coupled to a decrease in the expression levels of genes in glycine cleavage pathway which control the majority of glycine breakdown in C. elegans. We found that mel-32 RNAi causes a marked rise in the endogenous level of glycine which in turn extends lifespan. Moreover, supplementing dietary glycine extends lifespan at concentrations between 5–500 μM, while mutations in methionine synthase [metr-1(ok521)] and S-adenosyl methionine synthetase [sams-1(ok3033)], two enzymes involved in methionine cycle, can fully abrogate this lifespan extension. Furthermore, we found that serine, another amino acid that feeds into one-carbon metabolism, shows similar transcriptional changes, metr-1 and sams-1 dependency, and lifespan extension upon dietary supplementation as does glycine. These results confirm an important role for the methionine cycle in the longevity effects of glycine.
Our work reveals a timing requirement of glycine supplementation in the promotion of longevity in C. elegans. Specifically, the first three days of adulthood (from D1 to D3) are crucial for glycine to confer the benefits on longevity. Given that the DAF-2 pathway also acts exclusively during adulthood and throughout the reproductive period to affect lifespan in C. elegans [36], further investigation is warranted to see if this classical longevity pathway is fully or only partially required for the beneficial effects of glycine. Furthermore, the first three days of adulthood coincide with the reproductive period of worms, raising an interesting question for future studies about the crosstalk between the reproductive system and glycine-activated longevity pathways.
Our work identified a counterintuitive biological phenomenon, whereby glycine accumulation was observed during the aging process in worms while supplementation of glycine was nonetheless able to prolong worm lifespan. However, it is not uncommon for changes that occur with age to also benefit lifespan when artificially induced. For example, suppression of IGF1 signaling may extend lifespan in many model organisms [2,37,38], while IGF1 levels themselves have been observed to decline with age [39,40]. Moreover, methionine restriction is beneficial to lifespan in a variety of model organisms [11,41], while methionine abundance in vivo has been observed to decline with age [observed in this study (S4A Fig) and [22]]. Similar to these phenomena, glycine supplementation may activate protective cellular pathways that promote longevity when exogenously applied, while a natural glycine accumulation with age may reflect the organism’s need to upregulate these same cytoprotective pathways to deal with the damage and detrimental changes occurring during aging.
Studies in rodents have suggested glycine supplementation to have pro-longevity effects [42,43], anti-inflammatory effects [44], to be cytoprotective [45], and to ameliorate metabolic disorders [46]. In humans, glycine supplementation in patients with metabolic disorders has a protective effect against oxidative stress and inflammation [47–49]. In line with these observations in mammalian systems, our data demonstrated that glycine promotes longevity in C. elegans. Furthermore, we show this benefit to occur in a metr-1 and sams-1 dependent manner, implicating the methionine cycle in longevity regulation. Finally, we show glycine supplementation induces widespread suppression of genes including many that are hallmarks of the aging process. Taken together, our findings suggest that dietary glycine is an effective strategy to increase lifespan and warrant further investigation for life- and healthspan studies in humans.
C. elegans strains N2 Bristol, RB2204 sams-1(ok3033)X, RB755 metr-1(ok521)II, eat-2(ad465), and daf-2(e1370) were obtained from the Caenorhabditis Genetics Center (CGC, University of Minnesota). Nematodes were grown and maintained on Nematode growth media (NGM) agar plates at 20°C as previously described [50].
E. coli OP50 and E. coli HT115 (DE3) with the empty vector L4440 was obtained from the CGC. Bacterial feeding RNAi experiments were carried out as described [51]. RNAi E. coli feeding clones used were mrps-5 (E02A10.1), T09B4.8, agxt-1 (T14D7.1), C15B12.1, R102.4, T25B9.1, gss-1 (M176.2), gcsh-1 (D1025.2), gcst-1 (F25B4.1), pfas-1 (F10F2.2), and dld-1 (LLC1.3). Clones of agxt-1 (T14D7.1), C15B12.1, R102.4, T25B9.1, gss-1 (M176.2), gcsh-1 (D1025.2), gcst-1 (F25B4.1), and pfas-1 (F10F2.2) were derived from the Ahringer RNAi library [52]; Clones of mrps-5 (E02A10.1), T09B4.8, and dld-1 (LLC1.3) were derived from the Vidal RNAi library [53]; Worms were fed RNAi bacteria from the time of hatching unless otherwise indicated.
Glycine and serine were purchased from Merck Millipore (no. 8.1603.0250) and Sigma (no. S4500), respectively. A stock of concentration of 1 M glycine and serine was made by dissolving glycine and serine in water. The pH of glycine and serine stock solution was adjusted to 6.0–6.5 with sodium hydroxide and then sterilized with 0.45 μm Millipore filter. The concentrations of 5 μM, 50 μM, 500 μM, 5 mM and 10 mM glycine, and 5 mM serine were used in the present study.
Overnight cultures of E. coli OP50 were seeded on standard NGM plates containing carbenicillin (25 μg ml-1) to prevent the bacterial growth. After drying overnight at room temperature, the bacterial lawn was irradiated with 254 nm UV light using a Stratalinker UV Crosslinker model 1800 (Stratagene, USA) at 999900 μJ/cm2 for 5 min. A sample of UV-exposed E. coli OP50 was collected and cultured in LB medium overnight at 37°C to confirm the bacteria were completely killed. Plates seeded with UV-killed bacteria were stored in 4°C and used within 1 week after seeding.
Lifespan experiments of amino acids supplementation were performed at 20°C without fluorouracil as described with the exception that UV-killed E. coli OP50 was used [54]. Briefly, for treatment throughout lifespan, worms were cultured on glycine- or serine- supplemented plates from the time of hatching until death; For the treatment during larval development only, worms were cultured on glycine- or serine- supplemented plates from egg until L4, and then transferred onto control plates until death; For the treatment from D1 to D3, worms were cultured on glycine- or serine- supplemented plates from D1 (one day after L4) to D3, and transferred on control plates. For amino acid treatment during adulthood only, worms were cultured on glycine or serine supplemented plates from D1 until death. During the reproductive period (≈ day 1–8), worms were transferred to fresh plates every other day to separate them from their progeny.
For RNAi lifespan experiments, worms were cultured on NGM plates containing 2mM IPTG and seeded with HT115 (DE3) bacteria transformed with either pL4440 empty vector or the indicated RNAi construct from the time of hatching. Worms were transferred to fresh plates containing 10 μM fluorouracil at L4 larval stage to prevent egg laying. 100–150 worms per condition were used for every lifespan. Survival was scored every other day throughout the lifespan and a worm was considered as dead when they did not respond to three taps. Worms that were missing, displaying internal egg hatching, losing vulva integrity, and burrowing into NGM agar were censored. Statistical analyses of lifespan were calculated by Log-rank (Mantel-Cox) tests on the Kaplan-Meier curves in GraphPad Prism.
Amino acids were extracted and analyzed as described before [22] and each experiment was performed in three biological replicates. About amino acids profiles change with age in wild type N2 related to Fig 2A and S4A Fig, worms were cultured on UV-killed E. coli OP50 and collected at the desired stage (L3, D1, D3, D6, and D9) for amino acids extraction; About amino acids profiles change with age in wild type N2, daf-2(e1370), and eat-2(ad465) related to S1 Fig, worms were cultured on alive E. coli OP50 and collected at L3 and D10 for amino acids extraction; About glycine supplementation experiments related to S5A–S5E Fig, worms were fed UV-killed E. coli OP50 and supplemented with glycine at the desired concentration (5 μM, 50 μM, 500 μM, 5 mM, and 10 mM) or with water control from the time of hatching. Then, D1 worms were harvested for amino acids extraction; About RNAi experiments related to Fig 1D and 1E, and S2 Fig, worms were cultured on NGM plates containing 2mM IPTG and seeded with HT115 (DE3) bacteria transformed with either pL4440 empty vector or the indicated RNAi construct from the time of hatching. Then, D1 worms were harvested for amino acids extraction.
Around 1500 synchronized worms at the desired stage were collected, freeze-dried and stored at room temperature until use. Worm lysates were obtained by homogenization and subsequent tip sonication. Amino acids were extracted from worm lysate containing 50 μg protein and measure by UPLC-MS/MS analysis.
In Fig 1B and 1C, worms were cultured on live E. coli OP50 and harvested at D1, D4, and D9 for mRNA extraction; In Fig 3B, worms were cultured on UV-killed E. coli OP50 and collected at L3, D1, D3, D6, and D9 for mRNA extraction. Approximately 500 worms were collected in three biological replicates per condition at the desired stage. Total RNA was isolated according to the manufacturer’s protocol. Briefly, samples were homogenized in TRIzol (Invitrogen) with a 5 mm steel metal bead and shaken using a TissueLyser II (Qiagen) for 5 min at a frequency of 30 times/sec. RNA was quantified with a NanoDrop 2000 spectrophotometer (Thermo Scientific) and stored at -80°C until use. Genomic DNA was eliminated, and cDNA was synthesized using the QuantiTect Reverse Transcription kit (QIAGEN). The qPCR reaction was carried out in 8 μL with a primer concentration of 1 μM and SYBR Green Master mix (Roche) in a Roche LightCycler 480 system. In all analyses, the geometric mean of two reference genes, eif-3.C and F35G12.2, was used for normalization and the oligonucleotides used for PCR are listed in S2 Table.
daf-2(e1370) and eat-2(ad465) worms were fed HT115 E. coli expressing empty vector from the time of hatching. Wild type N2 worms were fed HT115 E. coli expressing dsRNA against mrps-5 from the larval stage 4 of parental worms, and this exposure was continued in the first filial population (F1). mrps-5 RNAi-treated F1 worms were used for total RNA extraction. Microarray experiment was performed as described [28]. Approximately 500 young adult worms were collected in four replicates per condition and total RNA was extracted as described above. RNA quality and quantity were assessed after DNase clean-up using a 2100 Bioanalyzer (Agilent Technologies). RNA was amplified and labeled using a Low Input QuickAmp Labeling Kit (Agilent Technologies) and hybridized using the Agilent Gene Expression Hybridization Kit (Agilent Technologies). An ArrayXS-068300 with WormBase WS241 genome build (OakLabs) was used and fluorescence signals were detected by the SureScan microarray Scanner (Agilent Technologies). Data of all samples were quantile normalized using the ranked median quantiles as described previously [55].
Worms were cultured on UV-killed E. coli OP50 upon 500 μM glycine treatment or 5 mM serine treatment from the time of hatching. Approximately 500 worms at D1 were collected in quadruplicates per condition for total RNA extraction as described above. Genomic DNA residues were eliminated with RNase-Free DNase (Qiagen), followed with the cleaning up with the RNeasy MinElute Cleanup Kit (Qiagen). Samples were sent to GenomeScan B.V. (Leiden, The Netherlands) for RNA library preparation and sequencing at a 20 million read-depth (see methods below).
Samples were processed for Illumina using the NEBNext Ultra Directional RNA Library Prep Kit (NEB #E7420) according to manufacturer’s description. Briefly, rRNA was depleted using the rRNA depletion kit (NEB# E6310). A cDNA synthesis was performed in order to ligate with the sequencing adapters. Quality and yield after sample preparation was measured with the Fragment Analyzer. Size of the resulting products was consistent with the expected size distribution (a broad peak between 300–500 bp). Clustering and DNA sequencing using the Illumina cBot and HiSeq 4000 was performed according to manufacturer's protocol with a concentration of 3.0 nM of DNA. HiSeq control software HCS v3.4.0, image analysis, base calling, and quality check was performed with the Illumina data analysis pipeline RTA v2.7.7 asnd Bcl2fastq v2.17.
Reads were subjected to quality control FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc), trimmed using Trimmomatic v0.32 [56] and aligned to the C. elegans genome obtained from Ensembl, wbcel235.v91 using HISAT2 v2.0.4 [57]. Counts were obtained using HTSeq (v0.6.1, default parameters) [58] using the corresponding GTF taking into account the directions of the reads. Statistical analyses were performed using the edgeR [59] and limma/voom [60] R packages. All genes with no counts in any of the samples were removed whilst genes with more than 2 reads in at least 4 of the samples were kept. Count data were transformed to log2-counts per million (logCPM), normalized by applying the trimmed mean of M-values method (Robinson et al., 2010) and precision weighted using voom [61]. Differential expression was assessed using an empirical Bayes moderated t-test within limma’s linear model framework including the precision weights estimated by voom [61]. Resulting p-values were corrected for multiple testing using the Benjamini-Hochberg false discovery rate. Genes were re-annotated using biomaRt using the Ensembl genome databases (v91). RNA-seq samples were compared using principal component analysis (PCA) and Partial least squares discriminant analysis (PLS-DA) using mixomics [62]. Heatmaps, venn diagrams, and volcano plots were generated using ggplot2 [63] in combination with ggrepel (https://CRAN.R-project.org/package=ggrepel), and venneuler (https://CRAN.R-project.org). Data processing and visualization was performed using R v3.4.3 and Bioconductor v3.6.
Gene ontology (GO) analyses were conducted using DAVID bioinformatics resource [32]. Genes found to be significantly up- or downregulated with an adjusted p-value < 0.05 and | log2 (fold change) | > 0.5 were subjected to functional annotation clustering. To retrieve significantly enriched GO terms, enrichment threshold (EASE score) was set as 0.05 for all analyses and the category of each annotation cluster generated by David was curated manually.
Heat maps of gene expression profile in S7–S9 Figs were plotted using “R2”, Genomics analysis and Visualization platform (http://r2.amc.nl). Gene set map analyses were performed on “R2” with defined gene sets from the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways (http://www.genome.jp/kegg/) [64].
Data were analyzed by two-tailed unpaired Student’s t-test or by one-way ANOVA with Tukey’s post hoc test for multiple comparisons, except for survival curves, which were calculated using the log-rank (Mantel-Cox) method. For all experiments, data are shown as mean ± SD and a p-value < 0.05 was considered significant.
|
10.1371/journal.pntd.0007311 | Molecular characterization of Brucella species from Zimbabwe | Brucella abortus and B. melitensis have been reported in several studies in animals in Zimbabwe but the extent of the disease remains poorly known. Thus, characterizing the circulating strains is a critical first step in understanding brucellosis in the country. In this study we used an array of molecular assays including AMOS-PCR, Bruce-ladder, multiple locus variable number tandem repeats analysis (MLVA) and single nucleotide polymorphisms from whole genome sequencing (WGS-SNP) to characterize Brucella isolates to the species, biovar, and individual strain level. Sixteen Brucella strains isolated in Zimbabwe at the Central Veterinary laboratory from various hosts were characterized using all or some of these assays. The strains were identified as B. ovis, B. abortus, B. canis and B. suis, with B. canis being the first report of this species in Zimbabwe. Zimbabwean strains identified as B. suis and B. abortus were further characterized with whole genome sequencing and were closely related to reference strains 1330 and 86/8/59, respectively. We demonstrate the range of different tests that can be performed from simple assays that can be run in laboratories lacking sophisticated instrumentation to whole genome analyses that currently require substantial expertise and infrastructure often not available in the developing world.
| Brucellosis is endemic in Zimbabwe. This article describes the use of various assays such as AMOS, Bruce-ladder, MLVA, and whole genome sequencing to characterize Brucella species isolated from different animals in Zimbabwe. Choice of which assays to use in the laboratory is generally done considering reproducibility, robustness, expertise and affordability in a given setting. As evidenced in this study, most laboratories in Africa lack resources especially finances, equipments and expertise to perform necessary tests for diagnosis and identification of specific pathogens. The study shows that the differentiation of species can be correctly concluded from the analysis with AMOS, Bruce-ladder and MLVA16 assays. Furthermore, MLVA16 can be used as an epidemiological tool and traceback of outbreaks. These PCR assays can therefore add to the control and eradication of brucellosis, since the Brucella species (B. ovis, B. abortus, B. suis and B. canis) existing in Zimbabwe could be identified and characterized.
| Brucellosis is a worldwide infectious disease affecting a wide range of domestic and wildlife animals and humans [1]. Brucellosis is caused by species in the genus Brucella, which consists of six classic species, Brucella abortus, B. melitensis, B. suis, B. ovis, B. canis, and B. neotomae [2]. Recently, the genus has expanded to include B. ceti and B. pinnipedialis from marine animals [3], B. microti from the common voles (Microtus arvalis) [4] and red foxes [5], B. inopinata isolated from a human breast implant [6], B. papionis from baboons (Papio spp.), B. vulpis from red foxes (Vulpes vulpes) and novel Brucella spp. in amphibians and fish [7, 8, 9, 10].
In African countries, brucellosis is reported to be a serious threat; although, under-reported due to a limited number of studies conducted and the lack of epidemiological evidence [11]. B. abortus and B. melitensis have been reported frequently when livestock have been tested. However, there is limited information on the prevalence of brucellosis in small ruminants as compared to cattle [12]. In Zimbabwe, only B. abortus and B. melitensis have been reported to cause brucellosis in animals [13], with B. abortus biovar (bv.) 1 and to a lesser extent B. abortus bv. 2 reported to be the most prominent cause of bovine brucellosis [13]. However, reports might be biased because investigations/ testing targeted mostly bovine rather than other species. The authors used biotyping and AMOS-PCR to identify Brucella isolates from commercial and communal cattle farms in Zimbabwe and also reported a single B. melitensis bv. 1 isolate from a goat. Brucellosis has been demonstrated by serology to be present in Zimbabwean wildlife including African buffalo, eland, zebra, giraffe and impala [14] as well as in domestic dogs [15]. B. abortus bv. 1 was isolated from waterbuck (Kobus ellipsiprymnus) and eland (Taurotragus oryx) [16]. This complicates the control of the disease since the animals in areas bordering the National parks interact with wildlife and it is almost impossible to vaccinate wildlife. Bovine brucellosis is endemic in the country in most regions with high sero-prevalence of up to 53% reported in commercial herds as compared to 16% from small-scale farmers in Zimbabwe [17, 18].
General classification of Brucella species and biovars is still based on phenotypic characteristics, with minimal standards previously defined [19] well before the development of modern genomics and the discovery of new Brucella species. Biotyping is time-consuming and often difficult to interpret due to limited standardization of the typing reagents [20]. Moreover, the efficacy of biotyping is moderate and since it includes the manipulation of the live agent, it poses a biosafety and public health risks of laboratory infections to the personnel involved [20]. Initial assays based on DNA analysis by PCR amplification were genus-specific and not sufficient to assist brucellosis control programs [21, 22] in the endemic regions of Zimbabwe. Most programs for brucellosis control employ genus-specific serology tests which are confirmed by species-specific culturing since the associated regulatory methods are species dependent [23]. AMOS-PCR is a multiplex PCR assay that differentiates B. abortus bv. 1, 2 and 4, B. melitensis, B. ovis, B. suis bv. 1, B. abortus vaccine strains S19 and RB51 based on the genetic element IS711 [24, 25]. AMOS and Bruce-ladder multiplex PCRs use species- and strain-specific genetic differences to distinguish among Brucella species [26, 27]. The initial Bruce-ladder assay identifies almost all Brucella species including the vaccine strains B. abortus S19, RB51 and B. melitensis Rev1 but will occasionally incorrectly identify some B. canis strains as B. suis [27]. The original Bruce-ladder assay had limited utility for distinguishing the more recently described species such as B. ceti, B. pinnipedialis, B. microti, and B. inopinata; but was later updated [28; 29, 30]. Multi-locus variable number tandem repeats (VNTR) assays (MLVA) is a genetic approach with high discriminatory power in the Brucella genus, clearly identifying species and providing fine-scale resolution among isolates [31, 32]. The most commonly used MLVA scheme consists of 16 VNTR markers, including eight moderately variable minisatellites (panel 1) and eight highly polymorphic microsatellites (panel 2A and 2B) [20, 33] that has the capacity to distinguish Brucella species and their biovars. Accurate discrimination between species and biovars achieved with the high resolution MLVA is necessary to determine the source, origin and geographical spread of infection [34]. Finally, characterisation of the genome of Brucella species with whole genome sequencing (WGS) provides the ultimate genetic resolution and can enable the determination of other features such as virulence factors [35]. The availability of whole genome sequences covering B. melitensis [36], B. suis [37], and B. abortus [38] has contributed to our understanding of the pathogenicity and diagnosis of brucellosis [35, 36]. WGS combined with single nucleotide polymorphism (SNP) analysis provides greater resolution and fine-scale differentiation of Brucella species [39, 40, 41] that cannot be obtained with multiplex PCR assays or MLVA.
Brucella abortus bv.1 is the most frequently isolated species in the cattle industry in Zimbabwe, with B. abortus bv.2 occasionally detected [13]. However, the control program is compulsory for commercial farming but is only optional for communal cattle production systems so may be missing most cases of brucellosis [14]. Various strains have been isolated from samples collected between 1990 and 2009 from various host animals throughout the country at the Central Veterinary Laboratory (CVL) in Zimbabwe, with only some of the isolates previously identified to the species level using biotyping. The aim of this study was to characterize these Brucella strains using AMOS-PCR, Bruce-ladder and MLVA, to evaluate genotyping approaches and develop a toolkit to support a nation-wide eradication program at a sustainable cost. Finally, based on the data obtained with the abovementioned techniques, three isolates were further characterized with WGS [42].
All experimental protocols were approved by the Animal Experiments and Ethics Committee of the University of Pretoria (V096-15 AEC Approval) and the Section 20 approval obtained from DAFF (SDAH-Epidem 15012613530_ Section 20) for the use of animals and animal products.
Sixteen Brucella strains (Table 1) were isolated at CVL from samples of domestic animals collected and isolated isolated between 1990 and 2009 in Zimbabwe and used to evaluate the feasibility and the need of large scale surveillance in the country. At the time of the study (2011–2013) there was no surveillance going on, the study isolates were obtained from farms/clients samples submitted to CVL for routine screening. They were characterized as Brucella by bacteriological methods (urease, catalase, oxidase, H2S, indole and sensitivity to dyes (thionin and basic fuchsin)) as indicated by previously [43]. Due to financial constraints, it was not possible for the laboratory to buy PCR reagents at that time; thus, only 7 of the 16 cultures were further classified to species level with the available reagents (S1 Table) according to standard bacteriological methods (excluding the phage lysis test) [43].
DNAs from 17 reference strains obtained from National and OIE/FAO Animal Brucellosis Reference Laboratory in France were included as controls for PCR assays (Table 1). Genotyping information of Brucella strains from previous studies [20, 31, 32, 44] that were used in MLVA in this study can be accessed from MLVA database [45].
DNA was extracted from each strain grown on Brucella selective media and blood agar using Qiagen DNA mini kit (Qiagen) at CVL in Zimbabwe and quantified with BioTek Take3 Micro-Volume Plate used in BioTek Microplate reader using the Gen5 pre-programmed quantification protocol at the University of Pretoria, South Africa. The study controls were amplified with Genomiphi DNA Amplification Kit (GE Healthcare Life Sciences AEC-Amersham) to increase their quantity.
AMOS-PCR was done as described previously [25, 26]. The PCR mixture contained 1X MyTaq mix (Bioline), a combination of five primer sets specific for B. abortus, B. melitensis, B. ovis, B. suis (0.2 μM) and IS711 (1 μM), respectively, and 10 ng DNA per 25 μl reaction. The PCR conditions consisted of an initial denaturation at 95°C for three minutes followed by 35 cycles of 95°C for one minute, 55.5°C for two minutes and 72°C for two minutes.
Bruce-ladder PCR was also done as described previously [27]. PCR reactions (25 μl) composed of 1X MyTaq mix (Bioline), 0.4μM of each primer of the eight primer pairs and 10ng template DNA. PCR conditions consisted of initial denaturation at 95°C for three minutes, followed by 25 cycles at 95°C for 30 sec, 64°C for 45 sec and 72°C for three minutes and a final extension of 72°C for five minutes on an ABI 2720 Thermal Cycler (Applied Biosystems).
To confirm the identity of strains identified as B. suis and B. canis with Bruce-ladder, the previously described Suis-ladder multiplex PCR assay [46] was used.
PCR products were separated by gel electrophoresis on a 1.5% agarose gel subsequently stained with ethidium bromide and photographed under UV light.
MLVA16 was performed as previously described [20, 31]. The 16 locus set was divided in three groups namely panel 1 (bruce06, bruce08, bruce11, bruce12, bruce42, bruce43, bruce45 and bruce55), panel 2A (bruce18, bruce19, bruce21) and panel 2B (bruce04, bruce07, bruce09, bruce16 and bruce30). PCR was performed in 15 μl reactions containing 3–15 ng of DNA template, 1X PCR buffer (Promega), 200 μM of each deoxynucleotide triphosphate, 0.5 μM of each flanking primer [20,31] and 1U GoTaq Hotstart polymerase (Promega). The PCR conditions included an initial denaturation step of 96°C for five minutes, followed by 30 cycles of 96°C for 30 seconds, 60°C for 30 seconds, extension at 72°C for one minute, followed by a final extension step of 72°C for 5 minutes. The PCR reaction products (5 μl) were separated on agarose gels in 1x TAE buffer using electrophoresis until the bromophenol blue has run for 20 cm on the agarose gel. The 16M B. melitensis reference strain was included as a control since each VNTR locus size is known. Brucella reference strains that have already been characterized using the MLVA16 markers panel 1, panel 2A and panel 2B were included to ensure accurate evaluation of field strain genotypes. For Panel 1 VNTRs, 2% agarose gel was used with GeneRuler 100 bp plus DNA ladder (Thermo Scientific). For panel 2 VNTRs, 3% standard agarose gel and low molecular weight DNA ladder 766–25 bp (New England Biolabs) were used. The ethidium bromide stained gels were visualized by UV light. Genotype was scored by visual analysis of the gel images or BioNumerics software version 6.6 (Applied-Maths).
Band size estimates were converted to repeat units following the published allele numbering system version 3.6 [45] (S1 Table). MLVA data were analysed as a character data set within BioNumerics software (version 6.6) (Applied Maths). Clustering analysis was performed using the categorical coefficient and UPGMA (unweighted pair group method using arithmetic averages). A different weight was given to the markers depending on their panel: Panel 1 markers were assigned an individual weight of 2 (total weight for panel 1: 16), panel 2A markers a weight of 1 (total weight for panel 2A: 3), and markers of panel 2B a weight of 0.2 (total weight for panel 2B: 1) [20]. The MLVA16 results were compared with MLVA16 published data of Brucella reference and other strains [20, 31, 44] (S1 Table). Minimum spanning tree (MST) analysis was performed using MLVA8 (panel 1) in BioNumerics as well.
Zimbabwean B. suis strains ZW043 (GenBank accession CP009094.1 and CP009095.1) and ZW046 (GenBank accession CP009096.1 and CP009097.1) and B. abortus strain ZW053 (GenBank CP009098.1 and CP009099.1) [47] were selected for WGS since these strains were isolated from cattle in different regions of Zimbabwe and represented different MLVA genotype subclades. In addition, 23 B. abortus and 17 B. suis complete genomes were retrieved from GenBank and used for comparison and phylogenetic analyses (S2 Table). Sequenced reads from B. abortus and B. suis strains were aligned to B. abortus str. 9–941 (Accession no: NC_006932.1, NC_006933.1) and B. suis 1330 (Accession no: NC_017251, NC_017250) respectively, using Burrows-Wheeler Aligner (BWA) [48]. SAMtools [49] was used to sort and index the aligned reads of Brucella genomes. Sequence reads of the complete and draft Brucella genomes were simulated using SAMtools [49]. Picard-tools (http://picard.sourceforge.net/) were used to mark duplicate reads and to build binary index of the samples. Repeated regions of the Brucella sequenced reads were excluded from this analysis. For variant detection, Unified Genotyper method in GATK [50] was used to call for SNPs. Variant filtration and selection of SNPs was achieved using GATK. SNPs positioning sets were deducted from the aligned genomes using molecular evolutionary genetics analysis (MEGA) tool version 6 [51]. Only SNP positions that could be called in all genome sequences were used (core genome analysis) for phylogenetic analysis. A phylogenetic tree was constructed using (MEGA) tool version 6 [51] from the coreSNPs of the Brucella genomes. The trees were generated using maximum likelihood method with 500 bootstrap replicates.
All Brucella spp. strains from Zimbabwe were non-motile, gram-negative coccobacilli, positive for modified Ziehl-Neelsen stain, negative for indole production, and oxidase and catalase production positive. Only a few of the strains (S1 Table) were further characterized using growth characteristics and biochemical profiles (phage lysis was not determined).
All strains except ZW100 and ZW377 were successfully genotyped using AMOS-PCR (Fig 1A). ZW002 and ZW005 were identified as B. ovis, ZW011, ZW040, ZW043, ZW045-048, ZW201 as B. suis and ZW053, ZW248, ZW283 and ZW323 as B. abortus (Fig 1A and Table 2).
Bruce-ladder gave identical results (Fig 1B and Table 2) as AMOS-PCR and in addition could identify strains ZW100 and ZW377 as B. canis. Using the complementary Suis-ladder multiplex PCR [46], both strains were confirmed to be B. canis and strains ZW011, ZW40, ZW043, ZW045-048, ZW201 were confirmed as B. suis bv. 1 (S1 Fig).
Due to lack of sufficient DNA, strains ZW002 and ZW005 (both B. ovis), and ZW248 and ZW283 (both B. abortus) could not be genotyped using MLVA16. MLVA data derived from the seven reference strains were as expected from previously published data, with the exception of reference strains 16M, 63/290 and RM 6/66. The 16M strain used in the present study differs from the B. melitensis reference 16M strain at locus Bruce07, which is not unexpected due to high variability at this locus [25]. The reference strain RM 6/66 we used differed from B. canis reference RM 6/66 strain at Bruce07, Bruce09 and Bruce16 loci whereas 63/290 differs at loci Bruce09 and Bruce16. The difference between control strains used in this study and the reference strains may be due to amplification of the reference DNA using Genomiphi (GE Healthcare Life Sciences AEC-Amersham) due to low quantities of DNA from these strains available in our study.
The Zimbabwean strains consisted of eight MLVA16 genotypes and clustered into three groups when analyzed together with MLVA data from [45]. All eight B. suis bv. 1 strains (ZW011, 040, 043, 045, 046, 047, 048 and 201) belong to MLVA8 genotype 6 like the vast majority of B. suis bv. 1 strains in the MLVA bank and are most closely related to B. suis bv. 1 reference strain 1330 in the B. suis bv. 1, 3, 4 / B. canis MLVA cluster (Fig 2). ZW100 and ZW377 (both isolated from dogs in Harare) formed a sub-cluster with B. canis REF RM 6/66 in the B. suis bv. 1, 3, 4 / B. canis cluster (Fig 2). B. abortus strain ZW323 (MLVA8 genotype 28) was identical at all 16 VNTR loci to B. abortus bv. 1 strain (LNIV-416Ba1-07) from Portugal [44] while B. abortus bv. 1 ZW053 strain also belonged to MLVA8 genotype 28. The clustering obtained with the Minimum Spanning Tree (MST) analysis is similar to the UPGMA clustering (S2 Fig).
WGS-SNP phylogenetic analysis of 19 B. suis and 24 B. abortus genomes was defined by 7104 and 4549 core SNPs respectively. Phylogenetic analysis of the Brucella genomes showed that B. abortus ZW053 clustered in B. abortus bv. 1 and 2 clade alongside B. abortus bv.2 86/8/59, while B. suis ZW043 and ZW046 strains are grouped within the B. suis bv. 1 clade (Fig 3). Comparative SNP analysis between ZW053 and B. abortus bv. 2 str. 86/8/59 resulted in 35 SNPs as compared to 90 SNPs obtained when comparing the strain with B. abortus bv.1 str 9–941.
Fast and accurate diagnosis of brucellosis is important for control programs [23] and since the choice of the assay to use depends on the affordability and availability of expertise in a given country, it is always a trade-off between the two requirements. Eradication and control program based on compulsory calf vaccination with B. abortus strain S19 was introduced in Zimbabwe in the early 1980s, but only to commercial farms and was voluntary to communal ones [17, 18]. However, infections caused by B. abortus and B. melitensis have been reported from both the communal and the commercial areas of Zimbabwe [13, 14, 15]. PCR-based assays can be used as a supplement or even a replacement to biotyping for the identification of Brucella species and/or biovars [23, 25], as genotyping is often essential for accurate epidemiological inference. Biotyping is time consuming, labour intensive and requires good expertise specific for this pathogen. In addition, it involves handling of live cultures that poses risks of laboratory exposure and infection [52]. The purpose of the study was to explore the practical suitability of PCR assays (MLVA, AMOS-PCR and Bruce-ladder) for laboratories that do not have biotyping capabilities as was the case with CVL, Zimbabwe at the time of the identification of these Brucella strains isolated from cattle, pigs, dogs and sheep. B. abortus and B. melitensis are the most prominent species in Africa and were previously reported in Zimbabwe from livestock and wildlife [13, 14]. The occurrence of these species in wildlife complicates the control of bovine brucellosis since it is almost impossible to vaccinate wildlife. Furthermore, interaction between wild life and animals in areas bordering the National parks could result in possible transmission of the disease.
In the present report, eight strains including five isolated from bovine and two from pigs were identified as B. suis bv. 1. The strains were identified as B. suis using AMOS-PCR [24, 25] and Bruce-ladder [26]. Suis-ladder [43] and MLVA identified these isolates to be B. suis bv. 1. Four strains isolated from cattle were identified as B. abortus bv.1 with AMOS-PCR and Bruce-ladder in this study. The identification could be confirmed by MLVA in two cases only due to limited DNA availability. In a previous study [13], B. abortus bv. 1 was shown to be the main cause of bovine brucellosis in Zimbabwe; however, in this study B. suis bv. 1 was most frequently isolated strain even from cattle. The isolation of B. suis bv.1 from both pigs and cattle might be the result of either mixed farming or the interaction of animal species in the grazing areas and drinking points. MLVA, Bruce-ladder and Suis-ladder assays identified two strains ZW100 and ZW377 as B. canis. This is the first report of B. canis in Zimbabwe. Due to low quantity of DNA, two B. ovis strains (ZW002 and ZW005) were only identified with AMOS and Bruce-ladder PCR but not with MLVA. Brucella ovis has been indicated by OIE reports as present in Zimbabwe [19].
Two B. suis bv. 1 strains isolated from cattle were selected for draft whole genome sequencing since B. suis had not been reported from pigs in Zimbabwe in literature but was detected in samples from both cattle and pigs in this study. WGS indicated that the two strains are separated from B. suis bv. 1 reference strain 1330 [37] by only five SNPs. A third strain, identified as B. abortus bv.1 was shown by WGS-SNP analysis to be closest to a strain independently recovered from Zimbabwe.
A previous study [24] compared the AMOS, Bruce-ladder and MLVA8 assays for typing of Brucella species and found only Bruce-ladder correctly identified all tested Brucella strains as MLVA8 does not resolve the very closely related B. canis and B. suis bv. 4. Both MLVA11 and MLVA16 resolve the two species however and also allows comparison to a worldwide Brucella MLVA dataset [53]. As shown in previous studies [39, 40, 41], WGS-SNP analysis provides better resolution than MLVA16, and much stronger phylogenetic support although there are still fewer strains from more limited geographic areas available for comparisons as compared to the MLVA database. Importantly the number of public whole genome sequences, particularly sequence reads archives, is rapidly growing with already more than 1000 datasets available.
The status of B. suis as a single species has been questioned in light of a broader host specificity [54]. Isolation of B. suis bv. 1 from bovines in Zimbabwe was first reported in 2014 [47]. The present study further emphasizes the occurrence of B. suis bv. 1 in cattle and pigs. There are several reports of isolation of B. suis bv. 1 from cattle [55, 56] in which the infection appears to be noncontagious with limited induced pathology and no induction of abortions [19, 54]. The presence of B. suis bv. 1 in pigs and bovines in Zimbabwe could be due to the predominance of smallholdings with mixed populations of livestock [57]. Therefore, the use of multiplex PCR assays that will distinguish the four species (B. ovis, B. abortus, B. suis and B. canis) present in Zimbabwe as confirmatory test will strengthen the control programs since most serology assays are based on smooth lipopolysaccharides (LPS) which cannot detect B. ovis and B. canis as they are rough strains.
WGS analysis showed that the ZW053 strain from a bovine in Zimbabwe [47] has large insertions and deletions as described in other B. abortus genomes [38, 58]. In spite of the variations observed in the genome sequences (S2 Table), whole-genome sequencing of the three strains and their comparison to reference genomes indicate that the isolates were B. suis (ZW043 and ZW046) and B. abortus (ZW053) respectively, thus corresponding with the data obtained with the Bruce-ladder, AMOS, Suis-ladder and MLVA PCR assays. Isolates from sub-Saharan countries and those from Europe have been shown to respectively cluster together, although heterogeneity within these species especially B. abortus do exist [12, 59]. This was also the case with ZW053 as it grouped with a Portuguese strain, and we hypothesize that this might be the result of socio-economic, migration or colonization links among Zimbabwe, Mozambique and Portugal or more generally European countries. Clustering of B. abortus bv. 2 strain 86/8/59 within biovar 1 and 2 clade and alongside ZW053 (Fig 3) and other B. abortus in WGS-SNP analysis was also shown in a previous study [41] that indicated that it might either be due to the paraphyletic nature of the biovar 1, 2 and 4 clade and the biovar classification not consistently reflecting genetic relationship in this species and/or that the biochemical biotyping to biovar level is unreliable. The present results further indicate the usefulness of MLVA and WGS-SNP in support of disease control. However, to perform the abovementioned assays requires a purified DNA template which may prove difficult to obtain due to the difficulty of culturing Brucella. Furthermore, brucellosis is endemic in sub-Saharan countries including Zimbabwe thus, the use of affordable high-throughput assays is necessary. More importantly, tests that can detect all the species that exit in a specific country should be considered.
Since most laboratories in Africa lack resources and expertise to do biotyping of Brucella to the species level, PCR assays like Bruce-ladder, AMOS and MLVA can contribute to the identification and can furthermore be used as an epidemiological tool and traceback of outbreaks. However, the choice of assays should be made considering reproducibility, robustness, expertise and affordability in a given setting and in most cases this choice will be a compromise. Brucellosis control programs in most countries are based on serological tests which includes Rose Bengal test (RBT), milk ring test, (MRT), complement fixation test (CFT), enzyme-linked immunosorbent assay (ELISA), the fluorescence polarisation assay (FPA) etc. [60]. These tests have varying sensitivity and specificity and they are prone to cross-reactions with other bacteria that have the smooth lipopolysaccharide used as the antigen in these assays [60]. Therefore, to complement these limitations, molecular assays can be used since most of them are robust, less expensive and can differentiate between Brucella spp. at genus, species and biovar levels [25, 26, 27, 28, 33, 44]. The development of standardised, safe and efficient DNA extraction procedures sufficient to produce a few micrograms of DNA of a good quality allowing long term conservation will be essential for this purpose.
Bruce-ladder and AMOS assays are species-specific simple and robust multiplex PCRs. Even though the initial AMOS PCR assay was more limiting as it has the capability of detecting only B. abortus bv 1, 2 and 4, B. melitensis bv. 1 and B. suis but not B. canis; it was subsequently enhanced and currently can detect B. abortus biovars 5, 6 and 9 and the new subgroup 3b of biovar 3 as well [61]. Furthermore, its subsequent use alongside Bruce-ladder is also an advantage. Moreover, a previous study [13] in which AMOS PCR assay was used, also indicated the presence of brucellosis in Zimbabwe with infections mainly caused by B. abortus bv. 1 (84.6%) and B. abortus bv. 2 (15.4%). The MLVA16 assay provides a clustering of strains that is in accordance with all currently recognized Brucella species and biovars [11, 32, 43].
Considering affordability and reproducibility; Bruce-ladder can be used as it allows identification of all known Brucella species including the vaccine strains simultaneously in one run. This study has confirmed that species differentiation can be correctly deduced from both MLVA16 and Bruce-ladder analysis. These PCR assays can therefore add to the control and eradication of brucellosis, since B. ovis, B. abortus, B. suis and B. canis could be identified. The latter two species are reported for the first time in Zimbabwe. Additionally, more strains, whole genome sequences, and epidemiological data from Zimbabwe are needed to accurately draw conclusions on the clustering and circulation of strains.
|
10.1371/journal.pcbi.1000011 | Entropy Measures Quantify Global Splicing Disorders in Cancer | Most mammalian genes are able to express several splice variants in a phenomenon known as alternative splicing. Serious alterations of alternative splicing occur in cancer tissues, leading to expression of multiple aberrant splice forms. Most studies of alternative splicing defects have focused on the identification of cancer-specific splice variants as potential therapeutic targets. Here, we examine instead the bulk of non-specific transcript isoforms and analyze their level of disorder using a measure of uncertainty called Shannon's entropy. We compare isoform expression entropy in normal and cancer tissues from the same anatomical site for different classes of transcript variations: alternative splicing, polyadenylation, and transcription initiation. Whereas alternative initiation and polyadenylation show no significant gain or loss of entropy between normal and cancer tissues, alternative splicing shows highly significant entropy gains for 13 of the 27 cancers studied. This entropy gain is characterized by a flattening in the expression profile of normal isoforms and is correlated to the level of estimated cellular proliferation in the cancer tissue. Interestingly, the genes that present the highest entropy gain are enriched in splicing factors. We provide here the first quantitative estimate of splicing disruption in cancer. The expression of normal splice variants is widely and significantly disrupted in at least half of the cancers studied. We postulate that such splicing disorders may develop in part from splicing alteration in key splice factors, which in turn significantly impact multiple target genes.
| RNA splicing is the process by which gene products are pieced together to form a mature messenger RNA (mRNA). In normal cells, RNA splicing is a tightly controlled process that leads to production of a well-defined set of mRNAs. Cancer cells, however, often produce aberrant, mis-spliced mRNAs. Such disorders have not been quantified to date. To this end, we use a well-known measure of disorder called Shannon's entropy. We show that overall splicing disorders are highly significant in many cancers, and that the extent of disorder may be correlated to the level of cell proliferation in each tumor. Surprisingly, genes that control the splicing mechanism are unusually frequent among genes affected by splicing disorders. This suggests that cancer cells may withstand harmful chain reactions in which splicing defects in key regulatory genes would in turn cause extensive splicing damage. As mis-spliced mRNAs are widely studied for cancer diagnosis, awareness of these global disorders is important to distinguish reliable cancer markers from background noise.
| The majority of mammalian genes produce alternative transcripts as part of their normal expression program [1]–[4]. Alternative transcripts include splicing, polyadenylation and transcription initiation variants which can be expressed differentially in different tissues [4]–[7] providing the fine tuning of gene expression required for cell differentiation and tissue-specific functions. Disruptions in the balance of alternative transcripts, especially at the splicing level, are known to affect angiogenesis [8], cell differentiation [9] and invasion [10]. A large body of evidence has established connections between alternative splicing defects and cancer, so that the identification of transcript isoforms is now considered an important avenue in cancer diagnosis and therapy [11],[12].
The disruption of splicing isoform expression in cancer may result from very different underlying genetic events. On one hand, mutations in cis-regulatory sequences lead to the abnormal expression of specific isoforms, as observed for example in the BRCA1 gene in breast and ovarian cancer [13]. Another class of event includes alterations of the mRNA processing machinery or its signalling pathway. These may affect the splicing of specific genes such as CD44 [14]–[16], but may also cause wider perturbations of isoform expression as the processing of multiple genes can be simultaneously affected [17]–[20]. Evidence for wider changes in alternative transcription linked with cancer are present for instance in EST databases, where a large fraction of splice variant are actually tumor-specific [21]. However, while most studies of splicing and cancer attempt to isolate “signature” splice variants with significant over-expression in disease cells, no published work to date has focused on the bulk of splicing disruption that potentially arises when the splicing machinery is impaired.
The aim of the present study is to evaluate the extent and modalities of non-specific alternative transcript disruptions in cancer. Instead of seeking “interesting” signature isoforms, we analyzed the distribution of all isoforms from a single gene in a given tissue. We postulated that, in a tissue where the splicing machinery is impaired, the distribution of isoforms may be more disordered than in a control tissue. To measure the level of disorder in cDNA and cDNA tag libraries, we borrowed the notion of entropy from information theory. We applied this measure to all three types of alternative transcription, comparing isoform distributions in pairs of disease and normal tissues. Our results show that neither alternative polyadenylation nor alternative transcription initiation are associated with a disordered isoform expression. However, in half of the cancers studied, alternative splicing showed a highly significant entropy gain relative to the corresponding normal tissues. We analyze this entropy gain and discuss its possible causes.
Given a random variable X with probabilities P(xi) for discrete set of events x1,….,k, Shannon's entropy, also known as Information Entropy, is defined by:The entropy, and thus the disorder, is maximal when the probability of all the events P(xi) are equal and thus the outcome is most uncertain. Here, Shannon's entropy is applied to the expression profiles of different transcript isoforms for a given context. In the Figure 1 example, Gene1 has 4 alternative splice forms (SP1…SP4) and we are interested in their expression in normal cerebellum and cerebellum tumor tissues. For each splice form, we count the number of transcripts observed in different tissue types (for instance ESTs/cDNAs matching splice form SP1 are observed 4 times in cerebellum tumor libraries and once in normal tissue libraries). For this gene, isoform entropy across the four splice forms is higher in tumor than in normal cerebellum tissues, reflecting a more uniform tissue distribution of isoforms in the tumor libraries.
We hypothesised that impairment of the transcriptional or post-transcriptional control machinery in cancer or other diseases should result in the loss of a tissue-specific expression pattern of certain transcript isoforms. This loss can be measured by a gain of entropy in the expression pattern of isoforms of a given gene. By averaging entropy gains or losses on a sufficient number of genes expressed in a disease/normal tissue pair, we should observe a significant entropy bias if isoform expression is altered in this disease.
We obtained transcript isoform collections from the FANTOM3 database [1] for initiation variants and the ATD database [22] for polyadenylation and splicing variants. We then related isoforms to cDNA or cDNA tag counts and mapped each cDNA or tag to its tissue/disease information using the EvoC ontology [23] for ESTs/cDNAs or direct parsing of CAGE/SAGE databases as explained in Materials and Methods. A gene was considered in the entropy calculation only if it had at least two alternative isoforms supported by at least 10 different transcripts from three separate libraries, thus a total of at least 20 transcripts mapped to each gene considered. In order to measure isoform entropy changes in a disease/normal tissue pair, we required that at least 50 genes and 100 isoforms were found expressed in both the normal and disease tissues. By considering only isoforms that were observed in both states, we excluded from our analysis spurious isoforms that are prevalent in many cancer EST libraries [24].
We define the entropy ratio of a gene as the ratio of the entropy of this gene in the disease to the entropy of the same gene in the normal tissue. The entropy ratio of a disease/normal tissue pair is the average of the entropy ratios of all genes available in this tissue pair. Figure 2 presents entropy ratios for different diseases with respect to alternative initiation (A), polyadenylation (B) and splicing (C). An entropy ratio of one means that isoform entropy does not vary between disease and normal tissue (thick line in Figure 2). To estimate significance boundaries, random assays were performed by dividing the average entropy of 1000 randomly picked genes from any disease/tissue state by that of another randomly picked set of 1000 genes from any other disease/tissue state and repeating this process 10,000 times. This process was performed independently on the three isoform datasets. Values for the highest and lowest percentile are represented by red and green vertical lines, respectively.
Entropy ratios for alternative initiation and polyadenylation did not ever exceed the significance boundaries (Figure 2A and 2B) in the 6+8 cancer/normal tissue pair studied. This suggests that expression of alternative polyadenylation and initiation isoforms does not present large scale alterations in cancer. Alternative splicing however was quite different with 24 of the 27 cancer tissues studied showing a higher level of entropy than their normal counterpart (Figure 2C and Table S1). This entropy gain was highly significant in 13 cases, suggesting that the expression of splicing isoforms is strongly disrupted in certain cancers. In none of the 27 cases studied did the normal tissues show significantly higher entropy than disease tissues, and none of the three non-cancer diseases (arthritis, ascites and schizophrenia) presented a significant entropy change between normal and disease tissues.
The observed entropy bias is not imputable to sampling differences in normal and cancer libraries. The number of ESTs/cDNAs used to calculate entropy did not differ significantly between normal or disease tissues (Table S1), mainly due to the fact that we considered only isoforms that are expressed both in disease and normal tissues. Furthermore, Pearson's correlation tests (Table S1) showed no relationship between the entropy ratio and differences in the numbers of ESTs/cDNAs between normal and disease tissues (P = 0.28) or between the entropy ratio and the total size of libraries (P = 0.12). The observed gain in entropy can therefore not be attributed to a size effect of cancer EST libraries.
In the ten most disrupted cancer tissues, splicing entropy gains were caused by 16 to 258 significantly disrupted genes, or 30%–68% of the gene set available for entropy calculation in these tissues. This suggests that splicing perturbation is caused by factors that regulate multiple genes at the same time. Sets of splice-disrupted genes from different tissues show little overlap therefore we cannot isolate a list of genes displaying a generally higher rate of splicing disruption. However, a clear functional trend appears when high entropy gain tissues are pooled together. In the ten cancer tissues that displayed the highest gain in splicing entropy (from stomach/carcinoma to brain/astrocytoma, Figure 2), we analyzed all genes showing a splicing entropy gain (414 genes) for functional enrichment. Interestingly, the most over-represented terms among splice-disrupted genes either contain “RNA splicing” or are higher level terms that incorporate RNA splicing (Table 1). The “RNA splicing” class mostly comprises splice factors. This suggests that splicing alterations in a few key splice factors could be involved in the more extensive splicing disruption observed in the high entropy-gain tissues. This enrichment is observable only after cancer tissues are pooled, which means the number of disrupted splice factors in a single disease is low. A total of 13 splice factors show a significant increase in splicing entropy in the cancer tissues studied (Table S2). Most are constitutive splice factors, only three (TRA2B, U2AF1, SF3A2) being involved in alternative splicing regulation.
Splice factors are subject to alternative splicing at higher rates than average genes: 72% of the 58 annotated splice factors in Gene Ontology [22] have at least one alternative splice form in the ATD database [25], with an average of 5.4 isoform per gene, compared to 62% alternative splicing and 3.4 isoform per gene in the total ATD gene set. To test whether this bias could explain the over-representation of splice factors among disrupted genes in the high entropy gain cancers, we performed the same GO-term analysis among splice-disrupted genes in the ten disease categories displaying the lowest entropy gain. We could not observe any functional bias in this gene set (not shown). Therefore, splicing deregulation of splice factors is a hallmark of tissues where overall splicing is deregulated. This again designates misplicing of splice factors as a possible cause of wider splicing disruption in these tissues.
Although tumors are diverse and heterogeneous, they all share the key ability to proliferate at a higher level than normal tissue and this despite the very tight control that the organism usually exerts on cell proliferation. To test potential links between disordered isoform expression and higher levels of proliferation, we classified the cancer types that deregulate the splicing mechanism (Figure 2C) in function of their proliferative potential. To evaluate proliferation, we extracted the 188 genes from the “cell cycle” module of Stuart et al. [26], a cluster of coexpressed genes shown to be enriched in elements that are overexpressed in highly proliferative cells and whose high expression is a marker of entry into the cell cycle [27]. We manually verified each of these 188 genes (Table S3) and confirmed that 92 were shown to be specifically over-expressed during one of the replicative phases of the cell cycle and another 17 bore significant proof of being over-expressed in proliferating cells. We thus used a high expression of these markers as a surrogate for a high level of proliferation. In order to obtain a “proliferation index” of cancer samples, we computed the median expression level of the 188 markers in each of 3787 published Affymetrix microarray experiments performed on cancer samples [28]. Samples were then binned into five categories from low to high proliferation, as shown in Figure 3. To relate proliferation levels to splicing entropy results, we considered only microarray samples that contained the exact same keywords as disease tissues in Figure 2C. Results are shown in Figure 4. Cell proliferation, as measured from the expression of cell cycle genes, is significantly correlated to splicing entropy gains.
This observation led us to question the possible correlation between splicing entropy and cellular proliferation in a non-pathological context. We compared the splice isoform entropy of foetal and adult tissues in the same manner we compared disease and normal tissues (Figure 5). While foetal tissues are expected to present higher levels of proliferation than their adult counterparts, we could not observe any significant entropy gain in foetal tissues. This suggests the higher isoform entropy observed in highly proliferating cancers is only indirectly related to proliferation (proliferation indices of foetal tissues could not be obtained due to insufficient foetal microarray data).
While previous studies of cancer-related splicing alterations have focused mainly on the discovery of “aberrant” splice variants, we looked instead at changes in the balance of variants expressed in both healthy and cancer tissues. This new perspective enabled us to characterize another kind of splicing disorder in which splice variant expression profiles are significantly flattened in tumors. While isoforms from the same gene are usually differentially expressed in a given tissue, with clear minor and major forms, these expression differences are reduced in cancer and this leads to a raise of isoform entropy. Although controlled over/under-expression events may in principle produce a flattened profile, we find unlikely that the generalized entropy gain observed in cancer could result from a combination of multiple controlled changes in isoform expression. The entropy gain is more likely a sign of a general loss of regulation involving widespread, non-specific perturbations of alternative splicing. We did not observe such cancer-related disorders in alternative transcription initiation and alternative polyadenylation, the two other processes associated with expression of disease-specific isoforms.
Previous efforts to identify cancer-specific splice forms, either through EST analysis or experimental means, have mostly ignored non-specific, large-scale disruptions. An exception is the study by Xu and Lee [29] which sought splice forms with statistically significant expression changes between normal and tumor EST libraries. In that sense, these authors were looking for events that would cause an entropy reduction, not an entropy gain. However, they also discussed the impact of unspecific disruptions and analyzed expression patterns that may lead to cancer-specific isoforms (Figure 6). The most frequent patterns leading to cancer-specific events were the loss of a normal isoform S, and the switch in expression between normal (S) and cancer-specific (S') isoforms. A general entropy gain would go against the occurrence of such events, which makes these patterns even more interesting on a background of entropy gain. Contrarily, the “gain of S'” category is directly correlated to a rise of entropy (i.e. the “tumor” situation has higher entropy). Therefore, in a context of general entropy gain, events of the “gain of S'” category, even when statistically significant, could merely reflect the wider splicing disruption and should be considered with caution. Xu and Lee rightly noted that this category, which produces only a small fraction of cancer-specific splice forms, may be related to a loss of splicing specificity in tumors.
There is now ample evidence that changes in splice factor expression, due for instance to kinase activation [14], disrupt splicing patterns in tumors [16], [18]–[20],[30],[31]. Figure 7, box A presents the most common of these effects, where an up-regulated splice factor causes expression of a rare or aberrant splice form. Splice factors previously analyzed for such dysfunctions include SF2/ASF, U2AF-65, SFRS2, SFRS3, SRm160, hnRNP A1/A2, and TRA2-β, all acting both in alternative and constitutive splicing. Although these factors may potentially target many genes, studies have focused on specific targets such as CD44 and have not examined more widespread splice defects. The splicing disruptions that we observed apparently affect a larger number of transcripts and are characterized by a loss of splice form regulation. Although this phenomenon might occur as a byproduct of the above mechanism, its association with the mis-splicing of splice factors, prevalently of the constitutive type, leads us to postulate a second process (Figure 7, box B) in which mis-splicing of general splice factors would cascade into a wider splicing disruption and entropy gains. Among the 13 splice factors that displayed splicing disruptions in our study, two were already known to regulate their own splicing: SFRS3 and TRA2-β [15],[28]. In each case, overexpression of the splice factor activated the inclusion of stop codon-containing exons [15],[28] producing transcripts subject to nonsense-mediated decay [32],[33]. Both genes have additional isoforms that are not NMD-prone (Figure S1) and may contribute to the mis-splicing of other genes.
A possible link between the two pathways in Figure 7 naturally comes to mind when considering that a change in splice factor expression in pathway “A” could alter the splice variant balance of other splice factors in pathway “B”. This transition may occur preferentially in highly proliferating tumors, where we observed the strongest splicing disruption. Splicing perturbation is knowingly correlated to proliferation [31] however no causal relationship between these events has been identified yet. Perhaps the splicing mechanism has trouble in trying to keep up with the accelerated pace of cell proliferation or a general disorder in splicing is causing failure in the regulation of cell cycle. Independently of any mechanistic hypothesis, splicing entropy measures show that widespread splicing disruption may be prevalent in most cancer tissues. In such a context of high splicing entropy, therapeutic avenues involving the reprogrammation of mis-spliced isoforms [34] would have a limited interest. As already recognized in different studies [35],[36] splice factors or their regulatory machinery may turn out as better therapeutic targets.
Transcripts and expression data for each type of transcriptional variation (initiation, splicing, polyadenylation) were obtained from the following sources.
Alternative initiation isoforms were obtained from the CAGE Basic/Analysis databases at http://fantom31p.gsc.riken.jp/cage_analysis/hg17/. This database classifies 3,106,472 CAGE tags into 450,228 transcription clusters (TC) further grouped into 32,351 transcription units (TU). TCs and TUs are two operationally defined units proposed in FANTOM3 [1] used to characterize promoters and genes respectively. We considered only those TCs that bore proof from at least 3 different CAGE libraries and 10 transcripts. These TCs were downloaded from the RIKEN website as well as the mappings of CAGE transcripts to these TCs in a given tissue type. This allowed us to create a relational database in which each TC could be queried to display its mapped CAGEs in each tissue type and the TU to which it belongs. For each normal/disease tissue pair we could therefore query a list of TCs common to both tissue types, link these TCs to their specific TUs and obtain the number of CAGEs mapped to a each of these TCs from the normal tissue library and from the disease tissue library.
Alternative polyadenylation isoforms were downloaded from the EBI ATD database, Human Release 1 (31 May 2005) [25] at http://www.ebi.ac.uk/atd/humrel1.html. Here, we only considered poly(A) sites located in the 3′-most exon of the gene because poly(A) sites located in upstream exons can belong to different splice forms. Since alternative splicing and polyadenylation can interfere [37], such events cannot be safely attributed to either phenomena. Again, each alternative polyadenylation event had to be supported by three different cDNA libraries and 10 transcripts, giving a total of 206,138 transcripts mapped to 13,367 poly(A) sites for 4400 genes. These 13,367 poly(A) sites were downloaded from the ATD website as well as the mapping of ESTs, cDNAs and SAGES to these isoforms. cDNA and EST transcripts were then linked to the eVOC 2.6 ontology through their Genbank accession identifiers and SAGE transcripts were manually parsed for simple tissue descriptors that were identical to eVOC 2.6 ontology terms (39 descriptors from the Gene Expression Omnibus [27]). This allowed us to create a relational database in which each poly(A) isoform could be queried to display its mapped transcripts in each tissue type and the Ensembl gene ID to which it belonged. For each normal/disease tissue pair we could therefore query a list of poly(A) isoforms common to both tissue types, link these isoforms to their specific Ensembl gene identifier and obtain the number of transcripts mapped to a each of these isoforms from the normal tissue library and from the disease tissue library.
Alternate splice isoforms were also downloaded from the EBI ATD database, Human Release 1. Again, 3 separate libraries and 10 transcripts were required to establish a splice form. Transcripts that mapped to multiple isoforms were excluded from the study bringing the total number of transcripts/isoforms/genes in the database from 808845 / 52742 / 14791 to 444799 / 47308 / 12281. These 47,308 alternative splice sites were downloaded from the ATD website as well as the mapping of ESTs and cDNAs to these isoforms. cDNA and EST transcripts were then linked to the eVOC 2.6 ontology through their Genbank accession identifiers. This allowed us to create a relational database in which each alternative splicing isoform could be queried to display its mapped transcripts in each tissue type and the Ensembl gene ID to which it belonged. For each normal/disease tissue pair we could therefore query a list of splicing isoforms common to both tissue types, link these isoforms to their specific Ensembl gene identifier and obtain the number of transcripts mapped to a each of these isoforms from the normal tissue library and from the disease tissue library.
Cell-cycle specific genes were extracted from the conserved co-expression network defined by Stuart et al. [26] and available for download at http://cmgm.stanford.edu/kimlab/multispecies. A matrix of gene-gene Euclidean distances was computed and used for hierarchical clustering using R software. The tree obtained was then split into several groups by specifying a cutoff height of 10. All genes in the “cell cycle” cluster were extracted and their respective Locuslink ID used for annotation.
Microarray expression data was obtained from the Gene Expression Omnibus [28] selecting Affymetrix GPL96 platform (8340 different samples). We parsed microarray sample descriptions for the presence of any EvoC ontology keyword inherited from the top level term ≪neoplasia≫ and then manually checked to see if the description genuinely corresponded to a cancer-related experiment. From a set of 8340 microarray samples studied, 3787 samples corresponded to cancer-related microarray experiments. Proliferation categories were then attributed to each sample based on the median ranking (MR) of the expression level of the 188 genes from the cell cycle node, as follows: High proliferation : MR in the top 20% of the genes on array.; Medium-high proliferation : MR between top 20% and top 40% of genes on array; Medium proliferation : MR between the top 40% and top 60% of the genes on array; Medium-low proliferation: MR between bottom 20% and bottom 40% of genes on array; Low proliferation: MR in the bottom 20% of genes on array. |
10.1371/journal.ppat.1002876 | A Unique Bivalent Binding and Inhibition Mechanism by the Yatapoxvirus Interleukin 18 Binding Protein | Interleukin 18 (IL18) is a cytokine that plays an important role in inflammation as well as host defense against microbes. Mammals encode a soluble inhibitor of IL18 termed IL18 binding protein (IL18BP) that modulates IL18 activity through a negative feedback mechanism. Many poxviruses encode homologous IL18BPs, which contribute to virulence. Previous structural and functional studies on IL18 and IL18BPs revealed an essential binding hot spot involving a lysine on IL18 and two aromatic residues on IL18BPs. The aromatic residues are conserved among the very diverse mammalian and poxviruses IL18BPs with the notable exception of yatapoxvirus IL18BPs, which lack a critical phenylalanine residue. To understand the mechanism by which yatapoxvirus IL18BPs neutralize IL18, we solved the crystal structure of the Yaba-Like Disease Virus (YLDV) IL18BP and IL18 complex at 1.75 Å resolution. YLDV-IL18BP forms a disulfide bonded homo-dimer engaging IL18 in a 2∶2 stoichiometry, in contrast to the 1∶1 complex of ectromelia virus (ECTV) IL18BP and IL18. Disruption of the dimer interface resulted in a functional monomer, however with a 3-fold decrease in binding affinity. The overall architecture of the YLDV-IL18BP:IL18 complex is similar to that observed in the ECTV-IL18BP:IL18 complex, despite lacking the critical lysine-phenylalanine interaction. Through structural and mutagenesis studies, contact residues that are unique to the YLDV-IL18BP:IL18 binding interface were identified, including Q67, P116 of YLDV-IL18BP and Y1, S105 and D110 of IL18. Overall, our studies show that YLDV-IL18BP is unique among the diverse family of mammalian and poxvirus IL-18BPs in that it uses a bivalent binding mode and a unique set of interacting residues for binding IL18. However, despite this extensive divergence, YLDV-IL18BP binds to the same surface of IL18 used by other IL18BPs, suggesting that all IL18BPs use a conserved inhibitory mechanism by blocking a putative receptor-binding site on IL18.
| Interleukin 18 (IL18) is an important cytokine in inflammation and immunity. Mammals and poxviruses encode homologous inhibitory proteins of IL18, named IL18BPs, which regulate IL18 activity and, in the case of the viral proteins, contribute to virulence. Previous structural and functional studies revealed residues at IL18:IL18BP interface that are critical for the high-affinity binding, including a phenylalanine on IL18BPs, which is conserved among nearly all IL18BPs with the notable exception of yatapoxvirus IL18BPs. To understand the mechanism by which yatapoxvirus IL18BPs neutralize IL18, we solved the high-resolution crystal structure of the Yaba-Like Disease Virus (YLDV) IL18BP:IL18 complex. The structure revealed a 2∶2 bivalent binding complex, which has not been observed in any other IL18BPs. Through mutagenesis and functional studies, we found a set of interacting residues that are unique for the association of YLDV-IL18BP and IL18, likely compensating for the lack of the interactions involving the conserved phenylalanine. Despite this extensive divergence, however, YLDV-IL18BP binds to the same surface of IL18 used by other IL18BPs. Our study suggests that all IL18BPs use a conserved inhibitory mechanism by blocking a putative receptor-binding site on IL18 but the interface on IL18 is malleable by a broad and diverse family of mammalian and poxvirus IL18BPs.
| Poxviruses are a family of large, complex DNA viruses, infecting a variety of organisms including insects, reptiles, birds and mammals [1]. The poxvirus family is further subdivided into genera based on shared characteristics such as host range, morphology, antigenicity, and sequence similarity [2]. Four genera of poxviruses are known to be pathogenic to humans, including molluscipoxvirus, orthopoxvirus, parapoxvirus, and yatapoxvirus. As an immune evasion strategy, poxviruses encode an assortment of decoy receptors for chemokines and cytokines [3]. One such strategy for evasion of the host immune response is through modulation of the interleukin 18 (IL18) signaling pathway. IL18 is a pro-inflammatory cytokine belonging to the interleukin 1 superfamily and plays an important role in both innate and acquired immune responses by inducing interferon-γ (IFN-γ) production from T lymphocytes and macrophages while also enhancing the cytotoxicity of natural killer cells [4]. IL18 activity is modulated in vivo by a negative feedback mechanism involving a naturally occurring IL18 inhibitor, the IL18 binding protein (IL18BP) [5]. Homologues of IL18BPs are also encoded by many poxviruses including molluscum contagiosum virus and orthopoxviruses [6], [7] such as variola virus, the causative agent of smallpox.
Yaba-Like Disease Virus (YLDV) along with Yaba Monkey Tumor Virus (YMTV) are members of the yatapoxvirus genus of poxviruses. These viruses produce a very distinct disease in primates that is characterized by epidermal histiocytomas and vesicular lesions of the head and limbs [8]–[10]. Although their exact host reservoir is not well established, it is presumed that the immunomodulatory proteins expressed by these viruses can at least partially cope with the primate/human immune system. Analysis of YLDV and YMTV genome revealed yatapoxviruses encode a predicted IL18BP family member, designated as 14L [11], [12]. YLDV and YMTV 14L proteins share approximately 54% sequence identity between each other but less than 14% with IL18BPs of the orthopoxviruses such as the variola virus and the mousepox ectromelia virus (ECTV). Despite this low sequence similarity, YMTV 14L was previously shown to be a functional IL18 inhibitory protein with comparable affinity as orthopoxvirus IL18BPs [13].
The high-resolution crystal structure of the ECTV-IL18BP in complex with human IL18 revealed the structural basis by which orthopoxvirus IL18BPs antagonize IL18 signaling through direct competition with IL18 cognate receptor for binding [14]. The crystal structure along with mutagenesis studies identified a set of conserved residues from IL18 and IL18BPs as key to complex formation. In particular, a phenylalanine (F67 in ECTV-IL18BP) residue that is highly conserved in IL18BPs was found indispensible for IL18 binding [14]–[17]. Mutations of this site in all IL18BPs examined to date significantly decreased or even completely abolished the binding to IL18. In addition, a residue on IL18, K53, was shown as a ‘hot spot’ for binding IL18BPs, since mutations at this site drastically decreased binding affinity [18]. A strong π-cation interaction between IL18 K53 and the conserved phenylalanine residue (F67) of ECTV-IL18BP was revealed in the structure of ECTV-IL18BP:IL18 complex, explaining its important role in binding. Surprisingly, phylogenetic analysis and sequence alignment revealed the presence of a threonine (T64) in yatapoxvirus IL18BPs at the position equivalent to the conserved phenylalanine (Figure 1) [19]. Furthermore, mutation of K53 on IL18 only modestly affected the binding with YMTV 14L [13]. To understand how yatapoxvirus IL18BPs bind IL18, we determined the high-resolution crystal structure of YLDV-IL18BP:IL18 complex. This structure along with the functional analysis through mutagenesis and Surface Plasmon Resonance (SPR) provide new insights into the mechanism by which IL18BPs inhibit IL18. The result provided here could be helpful for developing inhibitors for IL18 or IL18BP.
Recombinant human IL18 and mature YLDV-IL18BP (residue 20–136) proteins were individually purified from E. coli, and were subsequently used to reconstitute a complex of IL18:YLDV-IL18BP. Initially, wild-type (WT) IL18 and YLDV-IL18BP were used, but no quality crystals were obtained. In efforts to improve crystallization, we mutated some non-essential cysteines and additional surface residues in IL18. Substitution of four cysteines with serines in IL18 was previously shown to increase the stability of IL18 without affecting IL18 activities [20]. We found that the use of an IL18 mutant [IL18 (8S), see Materials and Methods], containing substitutions of the four-cysteines with serines and substitutions of four surface residues opposite to the IL18BP binding interface with alanines, greatly improved crystal quality and reproducibility. The crystal structure of IL18 (8S) in complex with YLDV-IL18BP was determined to 2.7 Å. Furthermore, a crystal from the complex of IL18 (8S) and a three-cysteine mutant of YLDV-IL18BP (ΔC21, C87S, C132S) diffracted to 1.75 Å (see Materials and Methods, Table 1). The two structures are essentially identical to each other with a root mean square deviation (r.m.s.d) of less than 0.4 Å, and the three mutated cysteines in YLDV-IL18BP are distant from the IL18 binding interface. Therefore, for discussion of protein:protein interactions between YLDV-IL18BP and IL18, we will mainly focus on the higher resolution structure of the mutant YLDV-IL18BP.
The structure of the complex shows that the YLDV-IL18BP forms a homo-dimer in a back-to-back fashion, with each protomer binding to a molecule of IL18, forming a hetero-tetramer complex with a stoichiometry of 2∶2 (Figure 2A). The complex displays an elongated v-shaped architecture with the IL18BP homo-dimer at the center and IL18 at the ends.
As seen in the ECTV-IL18BP:IL18 complex [14], IL18 adopts the same β-trefoil fold, which is comprised of 12 β-strands (β1–β12) with one short α-helix and one 310-helix. The IL18 molecules in the two structures show very little conformational changes, with an r.m.s.d. of only 0.5 Å from 134 aligned IL18 Cα backbones (Figure 2B). The two IL18 molecules in the current complex structure are also nearly identical, having only a 0.1 Å r.m.s.d from 150 aligned cα backbone residues.
Each protomer of YLDV-IL18BP adopts a canonical h-type immunoglobulin (Ig) fold [21] comprised of mainly β-sheets as observed previously for ECTV-IL18BP (Figure 3A). However, there is an r.m.s.d. of 2.9 Å over the 106 aligned cα backbone residues between the two viral IL18BPs. ECTV-IL18BP has an extended β-sheet architecture with predominantly shorter loops connecting the β-sheets and β-strands, while YLDV-IL18BP has comparatively shorter β-sheets with extended connecting loops. The two IL18BP structures differ mostly at one sheet of the β-sandwich fold (βA, βB, βE and βF) that is not involved in binding IL18 (Figure 2B). There is a major rigid-body movement on this β-sheet, especially at strands βA, βE and βF where twists are estimated at about 30 to 45 degrees. Compared to ECTV-IL18BP, there are two additional α-helices located between β-strands F and G (H2), and at the C-terminus (H3). YLDV-IL18BP contains five cysteine residues, forming two intra-molecular disulfide bonds (SS) (C21–C87, C43–C111) and one inter-molecular SS bond (C132–C132), in contrast to only two intra-molecular SS bonds in ECTV-IL18BP. C43–C111 is the conserved SS bond among many Ig-fold proteins, which connects the two β-sheets and plays a key role in maintaining overall integrity of the Ig-fold structure. C21–C87 SS bond connects the very N-terminus prior to βA with the loop between βE–βF. In contrast, the very N-terminus prior to βA of ECTV-IL18BP is SS bonded to the neighboring βB. In the triple-cysteine mutant YLDV-IL18BP:IL18 structure, the βE–βF loop is not visible in the electron density map, suggesting that C21–C87 SS bond stabilizes the local structure, particularly the βE–βF loop (Figures 2, 3A). However, the C21–C87 SS bond is not critical for the overall structure of YLDV-IL18BP or its binding with IL18 (shown later).
YLDV-IL18BP dimerizes back-to-back, abutting on one edge of the β sandwich, exposing the opposite edge for binding IL18 (Figures 2,3). The homo-dimer interface involves mainly βA, βB, βH and the C-terminal helix H3. It involves extensive hydrophobic interactions in the center (I29, H30, V31, P32 and V33 from βA, the aliphatic side chain of E122, V124 from βH and I129 from H3) flanked by hydrogen bonding and charge-charge interactions, burying approximately 1,700 Å2 and 1,465 Å2 solvent accessible surface area (ASA) for the wild-type complex and the triple-cysteine mutant complex, respectively (Figure 3). The lack of the C132-C132 SS bond in the triple-cysteine mutant complex caused the disorder of four residues at the C-terminus beyond S132, resulting in a slightly smaller ASA. At the dimer interface, the imidazole ring of H30 from one protomer stacks on the P32 from the other (Figure 3C), while V33 is forming favorable van der Waals interactions with a hydrophobic platform comprised of I29, V31, V124 and I129 from the other protomer (Figure 3D). E42 from one protomer appears to be protonated, forming favorable hydrogen bonds with E42 from the other protomer. E42 also forms both intra-chain and inter-chain hydrogen bonds with R44 (Figure 3B). The inter-chain C132-C132 SS bond covalently links the two promoters. Indeed, a non-reducing SDS-PAGE showed that YLDV-IL18BP proteins expressed in E. coli or in mammalian cells formed disulfide-bonded dimers (Figure 4). However, the triple-cysteine mutant of YLDV-IL18BP displays nearly identical structure as the WT protein, and it exists as a dimer in solution judging by size exclusion chromatography (Figure 5) and dynamic light scattering analysis (data not shown). Therefore, the dimerization is not solely dependent on the C132-C132 inter-chain SS bond.
We performed additional mutagenesis studies to verify the importance of the residues at the homo-dimerization interface. E42R/C132S (EC) mutant remained a dimer in solution as the WT (data not shown), while mutants bearing H30A/V33R/C132S (HVC) substitutions or H30A/V33R/E42R/C132S substitutions (HVEC) appeared as monomers in solution based on size exclusion chromatography (Figure 5) and dynamic light scattering analysis (data not shown). Therefore, hydrophobic interactions as well as the disulfide bonding together contribute to the dimerization. When measuring the binding affinity with IL18 by Surface Plasmon Resonance (SPR), we found that the monomeric YLDV-IL18BP (HVEC) had a 3-fold decrease (student t-test P-value<0.05) in binding affinity than the dimeric WT YLDV-IL18BP (Table 2).
Sequence analysis shows that the residues key to YLDV-IL18BP homo-dimerization (H30, P32, V33, C132) are conserved only in yatapoxvirus IL18BPs but not in other IL18BPs (Figure 1). Therefore, dimer formation and bivalent binding of IL18 seems to be unique to yatapoxvirus IL18BPs. In fact, ECTV-IL18BP and human IL18BP were reported to be monomers in solution [14], [22].
YLDV-IL18BP binds IL18 by using the same edge of the β sandwich as observed in ECTV-IL18BP:IL18 complex structure. Specifically, the following regions on YLDV-IL18BP are observed at the interface: loop connecting βB–βC, βC, the short βD, helix H1, βG, and loop connecting βG–βH (Figures 2,3). Similar to what was observed in the ECTV-IL18BP:IL18 complex structure, YLDV-IL18BP molecule sits atop the opening of the IL18 β-barrel and binds the cytokine through extensive hydrophobic and hydrogen bonding interactions (Figure 6), covering about 1,957 Å2 of ASA, which is comparable to the ECTV-IL18BP:IL18 complex at 1,930 Å2 ASA as identified by the program AreaIMol of the CCP4 suite [23]. The numbers of residues involved at binding interfaces are also comparable between the two inhibitory complexes. YLDV-IL18BP contributes mainly 19 residues to the complex interface while IL18 contributes 23 residues, in comparison to 17 residues from ECTV-IL18BP and 25 residues from IL18 using NCont of the CCP4 suite [23]. To assess the energetic contributions to binding by residues at the binding interface, we performed site-directed mutagenesis on both IL18 and the monomeric YLDV-IL18BP (HVEC) and assessed the effects of the mutations on the binding affinity by SPR (Figures 7,8,9,10). In addition, to probe the difference in IL18 binding by IL18BPs of YLDV and ECTV, we performed binding studies of various IL18 mutants with the two IL18BPs simultaneously (Figures 9,10). We will describe the results from these functional studies in context of our depiction of the YLDV-IL18BP:IL18 complex interface. As we described in the previous ECTV-IL18BP:IL18 complex structure, we will continue to use the three identified binding sites, labeled as A, B and C on IL18 here (Figure 6).
Site A of IL18 contains key residues Y1, L5, K53, D54, S55 and P57, making extensive interactions with YLDV-IL18BP (Figure 6A,B). The detailed interactions are similar to those observed in the previous ECTV-IL18BP:IL18 complex structure except for the loss of one of the ‘hot spot’ interactions involving a phenylalanine (F67 in ECTV-IL18BP, described below). Substitution of S55 of IL18 with alanine was previously shown to decrease its binding affinity to orthopoxvirus IL18BP by 7-fold [18]. As observed in the ECTV-IL18BP:IL18 complex, the side chain hydroxyl of S55 is tethered to the main chain amino group of Y1 via a hydrogen bond in the YLDV-IL18BP:IL18 complex. Y56 on βC of YLDV-IL18BP occupies nearly identical position as observed in the ECTV-IL18BP complex without any conformational changes. Its phenolic group is tethered to the D54 main chain of IL18, while its aromatic side chain together with the methyl group of T64 from YLDV-IL18BP, and L5 of IL18 constructs a hydrophobic wall, entrenching the aliphatic side chain of K53 from IL18 (Figure 6B). We found mutation of Y56A on YLDV-IL18BP completely abolished its binding to IL18 (Figures 7,8, Table 2), which is consistent with previous functional analysis of other IL18BPs [16].Therefore, a tyrosine residue at this location in IL18BPs seems to be a conserved ‘hot spot’, as the most important point to anchor the inhibitory protein to IL18.
As predicted, T64 of YLDV-IL18BP, located on the tip of βD, indeed occupies the same location of F67 from ECTV-IL18BP. This phenylalanine residue is highly conserved in IL18BPs of various species including human, all orthopoxviruses and MCV. Mutations at this location of various IL18BPs were shown to dramatically reduce the binding with IL18 [15]–[17]. In the ECTV-IL18BP:IL18 complex, F67 is inserted into an induced hydrophobic pocket, forming strong interactions with IL18 residues located on the surface while forming a strong π-cation interaction with the charged head group from the side chain of K53 of IL18 [14]. These interactions are absent in YLDV-IL18BP complex due to the presence of a threonine instead of phenylalanine. Interestingly, T64F substitution in YLDV-IL18BP did not increase the binding of IL18 (Table 2, Figures 7,8), while T64A substitution only caused a 2.5-fold decrease (student t-test P-value<0.05) in binding affinity. Consistent with the lack of π-cation interaction, IL18 K53 contributes less to the binding with YLDV-IL18BP than with the orthopoxvirus IL18BPs. While K53A mutation drastically reduced the binding affinity to orthopoxvirus IL18BPs [i.e., more than 100-fold decrease for variola IL18BP [18], Figures 9 and 10], this mutation had less impact on binding affinity with YLDV-IL18BP (about 30-fold decrease, Table 3, Figures 9, 10). K53 of IL18 nevertheless remains important for binding to YLDV-IL18BP, because polar interactions involving K53 are preserved in the current structure. Specifically, the positively charged amino head group on the side chain of K53 forms salt bridges with D66 and E76 of YLDV-IL18BP, similar to the interactions of K53 with two glutamate residues of ECTV-IL18BP in the ECTV-IL18BP:IL18 complex structure.
Site A differences also include the side chain rotation and repositioning of Q67 on YLDV-IL18BP (equivalent to H70 of ECTV-IL18BP) and Y1 on IL18, creating a novel interaction that was absent in the ETCV-IL18BP:IL18 structure. Q67 is located in close vicinity to T64 and rotates about 90 degree (vs. H70 of ECTV-IL18BP) forming bifurcated hydrogen bonds with the hydroxyl group and the main chain amide nitrogen of T64. Y1 of IL18 rotates about 80 degrees and stacks on the aliphatic portion of the Q67 side chain, forming favorable van der Waals interactions (Figure 6B). Y1A substitution of IL18 and Q67A substitution of YLDV-IL18BP decreased the binding affinity by 20- and 4-fold, respectively (student t-test P-value<0.05, Table 2, 3, Figures 7,8,9,10). In contrast, Y1A mutation did not affect the affinity with ECTV-IL18BP (Table 3, Figures 9,10). Since T64A substitution of YLDV-IL18BP showed very little effect on binding to IL18 (Table 2, Figures 7,8), the hydrogen bond between the side chain of Q67 and the main chain of T64 seems to be more significant than its interaction with the side chain. It is likely Q67 further stabilizes the local structure, including helix H1 where D66 locates, allowing correct positioning of this acidic residue for interacting with K53 of IL18.
Site B is a large, predominantly hydrophobic cavity spatially adjacent to Site A on the surface of IL18 (Figure 6A). As observed in the ECTV-IL18BP complex, three non-contiguous residues, Y54, I114 and P119 (Y51, T113 and V118 in ECTV-IL18BP) from YLDV-IL18BP βC, βG and G-H loop reside but do not fully occupy the pocket. Y54A substitution in YLDV-IL18BP had negligible effects on IL18 binding, while I114A mutation caused only a 1.8-fold decrease in binding affinity (statistically not significant, student t-test P-value>0.05) (Figures 7,8, Table 2), similar to substitutions of the equivalent positions of other IL18BPs [16], [17]. Therefore, site B interactions appear not essential for binding of YLDV-IL18BP, similar to observations for other IL18BPs [14].
Site C of IL18 is next to Site B and mainly comprised of 10 IL18 surface residues involving a mixture of charged and hydrophobic interactions. Similar to what was observed in the ECTV-IL18BP complex, the loops connecting βB–βC and βG–βH of YLDV-IL18BP interact with Site C on IL18 predominantly through hydrophobic interactions. YLDV-IL18BP F52 adopts nearly identical conformation as ECTV-IL18BP F49, which is inserted into the large hydrophobic pocket of Site C [14]. Surprisingly, F52A mutation of YLDV-IL18BP had negligible effect on binding with IL18 (Figures 7,8), in contrast to the mutation at this location in other IL18BPs significantly decreasing binding affinity to IL18 (83–fold decrease for human IL-18BP:human IL18, 8-fold decrease for ECTV-IL18BP:human IL18 and 138-fold decrease for ECTV-IL18BP:murine IL18) [16], [17]. This difference can be explained by several unique interactions of YLDV-IL18BP at Site C. Residue P116 of YLDV-IL18BP is situated in a hydrophobic groove formed by aliphatic side chains from M60, Q103 and M113 of IL18 and is stabilized by a stair-wise hydrophobic stacking by side chains from F49, Y48 and K23 of YLDV-IL18BP (Figure 6A,C). In addition, P116 main chain is hydrogen bonded with the hydroxyl group from the side chain of IL18 S105, further stabilizing the complex interface. Indeed, P116A substitution reduced binding of YLDV-IL18BP to IL18 by approximately 6-fold (student t-test P-value<0.005, Figures 7,8, Table 2). Mutation of the equivalent residue in human IL18BP (P153) showed negligible effect on its binding affinity to IL18 [15], so P116 appears to be specific for YLDV-IL18BP in binding to IL18. Y48 and K23 of YLDV-IL18BP are hydrogen bonded with IL18 D110 through side chain to side chain interactions. K117 of YLDV-IL18BP forms a bifurcated hydrogen bond with side chains from IL18 S105 and D110 (Figure 6C). D110A of IL18 decreased affinity with YLDV-IL18BP by 6-fold (student t-test P-value<0.05) but had no effect on binding with ECTV-IL18BP (Figures 9,10). Similarly, S105R of IL18 caused more than 30-fold decrease (student t-test P-value<0.005) in binding affinity with YLDV-IL18BP but had no impact on binding with ECTV-IL18BP (Figures 9,10). Therefore, D110 and S105 of IL18 are specifically required for the binding of IL18 with YLDV-IL18BP but not for ECTV-IL18BP. The specificity of S105 towards binding of YLDV-IL18BP is further signified by a double mutant, S105R/P57R of IL18. This double mutant completely abolished the binding with YLDV-IL18BP, while it had a much smaller impact on binding with ECTV-IL18BP (Table 3, Figures 9,10).
We previously determined the crystal structure of the ectromelia virus IL18BP in complex with IL18, revealing the structural basis for the binding and inhibition of IL18 by the IL18BPs [14]. Despite an overall low sequence homology between the diverse viral and host IL18BPs, the key residues of ECTV IL18BPs at the IL18 binding interface are highly conserved. Mutagenesis studies on human and viral IL18BPs also showed that these key residues are almost universally critical for the binding of IL18BPs to IL18. Thus it was enigmatic that functional yatapoxvirus IL18BPs lack a key phenylalanine residue (Figure 1) that has been identified to be essential for many other IL18BPs to bind IL18 [14]–[17]. It is similarly puzzling that a residue of IL18 (K53) that is critical for binding orthopoxvirus IL18BPs only played a modest role in binding with the YMTV-IL18BP [13]. In this report, we resolved these questions by determining the crystal structure of YLDV-IL18BP:IL18 complex and by performing extensive mutagenesis and SPR studies. We revealed two unique signature features of YLDV-IL18BP that distinguish yatapoxvirus IL18BPs from the rest of IL18BP family members. First, YLDV-IL18BP forms a homo-dimer and interacts with IL18 in a 2∶2 binding mode. Second, the binding of YLDV-IL18BP and IL18 does not rely on two of the ‘hot spot’ interactions that were shown to be essential for the binding of all previously studied IL18BPs, including a phenylalanine (F67 in ECTV-IL18BP) at site A and another phenylalanine at site C (F49 in ECTV-IL18BP). Instead, yatapoxvirus IL18BPs evolved interactions with some IL18 residues (Y1, D110, S105) that are specifically important for binding with YLDV-IL18BP. It appears that YLDV-IL18BP shifts and disperses the binding energy across the IL18-binding interface rather than concentrating the binding energy on a few hot spots as is the case for all other IL18BPs examined to date.
It was previously reported that ECTV-IL18BP and human IL18BP are monomeric in solution [14], [22]. In contrast, YLDV-IL18BP forms a disulfide-bonded dimer, which was demonstrated not only in the crystal structure but also in solution by non-reducing SDS-PAGE and gel filtration analysis. The dimer interface is quite large (about 1,700 Å2) and involves extensive hydrophobic interactions in addition to the intermolecular disulfide bond, indicating that the YLDV-IL18BP dimer is intrinsically stable in solution. The dimer could only be separated into monomers by mutations that disrupt both the hydrophobic interactions as well as the inter-chain SS bond. Analysis of the monomeric YLDV-IL18BP (HVEC) showed that the dimerization was not essential for binding IL18 but enhanced the binding affinity by 3-fold in our in vitro assay. Although this enhancement in binding affinity as measured by SPR is modest, it is possible that the dimerization may be more important for the function of YLDV-IL18BP during infection of the host, perhaps by increasing the half-life of the protein in the infected tissue or by increasing the avidity of binding to IL18 at low protein concentration. In fact, divalent or multivalent binding is an important, inherent feature of many biological systems to enhance the effectiveness of binding of ligands to receptors and of antibodies to antigens [24]–[28]. More specifically, this has been a feature for quite a few poxvirus cytokine binding proteins. For example, ectromelia virus IFN-γ binding protein forms a tetramer, which is required for efficient IFN-γ antagonism [29]. Myxoma virus T2 protein, a Tumor Necrosis Factor (TNF) Receptor homolog, is secreted as both monomer and dimer, and the dimeric T2 is a more potent TNF inhibitor [30]. Because residues of YLDV-IL18BP involved in dimer formation are only conserved in yatapoxviruses, yatapoxviruses IL18BPs may be unique among IL18BPs in that they use bivalent binding to increase the affinity and avidity for IL18.
Another difference between YLDV-IL18BP and all other IL18BPs is the lack of two of the ‘hot spot’ interactions at the binding sites A and C on the surface of IL18. The structure of ECTV-IL18BP:IL18 complex showed that a conserved phenylalanine (F67) is engaged in hydrophobic and strong π-cation interactions with K53 of IL18 at binding site A [14]. Alanine substitutions of K53 of IL18 significant decreased binding with orthopoxvirus IL18BPs [18], while alanine substitutions of the conserved phenylalanine (equivalent to ECTV-IL18BP F67) in human, MCV and orthopoxviruses IL18BPs significantly decreased or completely abolished binding of IL18 [15]–[17]. The current structure of YLDV-IL18BP:IL18 complex showed that a threonine residue (T64) is present at the position equivalent to the phenylalanine, indicating the π-cation interaction with K53 is not important for YLDV-IL18BP: IL18 complex. Indeed, T64F or T64A substitution of YLDV-IL18BP had negligible or minor (2.5-fold decrease in affinity for T64A, student t-test P-value<0.05) effect on the binding with IL18, while K53A of IL18 had a more modest effect on binding with YLDV-IL18BP than with orthopoxvirus IL18BPs. A similar loss of ‘hot spot’ interaction was also observed at binding site C. A phenylalanine residue on IL18BPs that binds to site C of IL18 was previously shown to be important for binding IL18 in orthopoxvirus, MCV and human IL18BPs [15]–[17]. Although a phenylalanine (F52) is present at the equivalent position in YLDV-IL18BP, it is not important for binding to IL18 (Table 2, Figures 7,8). Through structural and mutagenesis studies, we have identified contact residues that are unique to the YLDV-IL18BP:IL18 binding interface. This includes Q67 of YLDV-IL18BP and Y1 of IL18 at site A, P116 of YLDV-IL18BP and S105 and D110 of IL18 at site C. Our data are in agreement with the conclusion of a more delocalized energy distribution for binding of IL18 to YMTV-IL18BP [13]. The structural and functional studies of two different IL18BP complexes suggest that there is a degree of plasticity in the IL18BP:IL18 interface that could accommodate certain mutations in IL18BPs without compromising their binding affinity to IL18.
Despite the differences in several key residues for binding IL18, the current YLDV-IL18BP:IL18 complex structure showed that YLDV-IL18BP targets the same surface of IL18 as ECTV-IL18BP does in the previous complex structure. This suggests that all IL18BPs inhibit IL18 function by blocking a putative receptor-binding site on the surface of IL18. Similar to previous findings on human and poxvirus IL18BPs, Y56 of YLDV-IL18BP (interacting with site A of IL18) was found to be absolutely essential for binding to IL18, indicating that this conserved tyrosine residue is an obligatory ‘anchor’ for binding of all IL18BPs to IL18. The conservation and variation in functional residues and their specific interactions with IL18 suggest that IL18BPs share a common ancestor but may have undergone significant evolution through different selection pressures, resulting in a conserved inhibitory mechanism albeit with mutations of interface residues.
The biological activity of IL18 is determined in part by its relative affinities for IL18 receptors and IL18BP. The binding of IL18 to its receptors triggers multiple cellular responses vital to immunity, but excessive IL18 activities are associated with many autoimmune and inflammatory diseases [4], [31]–[37]. Functional IL18BPs are present in many poxviruses including variola virus and vaccinia virus, providing a key strategy of poxvirus immune evasion by inhibition of IL18 cytokine activity. Therefore the studies on IL18BP:IL18 inhibitory complexes could serve dual purposes by providing important clues on how to develop functional inhibitors targeting either IL18 or poxvirus IL18BP. These inhibitors could potentially modulate IL18 and poxvirus IL18BP activities, which may benefit efforts in developing treatments against some autoimmune and inflammatory diseases and in developing treatments for potential pathogenic outbreaks associated with poxvirus infections.
Mature IL18 and YLDV-IL18BP (residues 20–136) were individually cloned into a modified pET vector as SUMO fusion proteins with N-terminal 6×His tags and expressed in E. coli BL21 (DE3) gold (Stratagene) or Rosetta-Gami 2 (Invitrogen) strains, respectively. An IL18 mutant (C38S, C68S, C76S, C127S, K67A, E69A, K70A, I71A) with substitutions of four nonessential cysteines [20] and four additional surface residues opposite to the IL18BP binding interface, IL18 (8S), and the triple-cysteine mutant of YLDV-IL18BP (residues 22–136, C87S, C132S) were cloned and expressed in the same way as WT proteins. The individual proteins were purified using the similar double Ni-nitrilotriacetic acid (Ni-NTA) procedure as described [14]. Briefly, the his-tagged fusion proteins were first purified from cell lysate by Ni-NTA affinity column (Qiagen) and then co-dialyzed with ULP1 protease to remove the SUMO moiety, exposing the authentic N-terminus for both proteins. The cleaved protein mixtures were subsequently passed through a second subtracting Ni-NTA column and further purified by size exclusion chromatography on a Superdex s200 column. The YLDV-IL18BP and IL18 (8S) proteins were mixed together and the complexes were subsequently purified from size exclusion chromatography and each concentrated to 9 mg/ml. The complex of wild-type YLDV-IL18BP:IL18 (8S) crystallized in a condition containing 18% PEG3350, 0.1 M Bis-Tris Propane/Citric Acid, pH 6.5, while the crystallization condition for the complex of triple-cysteine mutant YLDV-IL18BP:IL18 (8S) is 12% PEG3350, 0.1 M Tris, pH 8.0. 25% ethylene glycol was added step-wise to the mother liquid as cryoprotectant.
To make biotinylated IL18, mature human IL18 (residues 37–193) was cloned into a modified pET vector containing a C-terminal 6×His tag along with the coding sequence (GLNDIFEAQKIEWHE) for biotinylation. This plasmid and a plasmid encoding biotin ligase (Avidity) were co-transformed into E. coli BL21 (DE3) gold (Stratagene) for expression. One step Ni-NTA affinity purification was used to purify biotinylated and his-tagged IL18.
For mammalian expression of YLDV-IL18BP, a mammalian expression plasmid for YLDV 14L was constructed as described previously for the construction of expression vector for human IL18BP [16]. Briefly, YLDV 14L was amplified by PCR from genomic DNA of YLDV and cloned into pYX45 with NheI and BamHI sites, so that 14L ORF was appended with a C-terminal biotinylation tag and 6-His tag. 293T cells were transfected with the expression plasmid and then infected with vTF7.3, a vaccinia virus expressing T7 polymerase. 3 days later, the medium was harvested and incubated with Ni-NTA resin (Qiagen). The resin was then washed and added with E. coli biotin holoenzyme synthetase (Avidity). After the biotinylation reaction, the protein was eluted with phosphate-buffered saline containing 300 mM imidazole. The ECTV-IL18BP was expressed in HEK293T cells, purified from the culture medium and biotinylated essentially as described previously [18].
An initial set of data for the tripe-cysteine mutant YLDV-IL18BP:IL18 (8S) was collected at beamline X29 of National Synchrotron Light Source, Brookhaven National Laboratory. Initial phases were determined by molecular replacement using Phaser [38] of the CCP4 suite [23] and a search model containing IL18 along with a trimmed poly-alanine model of the ECTV-IL18BP (PDB ID 3F62). A subsequent data set from a crystal of WT YLDV-IL18BP:IL18 (8S) complex was collect at beamline 19-ID of the Advanced Photon Source, Argonne National Laboratory. The structure of the complex containing WT YLDV-IL18BP was solved similarly as the structure of the complex containing the triple-cysteine mutant YLDV-IL18BP. All datasets were processed with HKL3000 [39]. PHENIX [40] was used for refinement and Coot [41] was used for iterative manual model building. Translation, libration and screw-rotation displacement (TLS) groups used in the refinement were defined by the TLMSD server [42]. The structure of the triple-cysteine mutant YLDV-IL18BP:IL18 (8S) complex was refined to 1.75 Å resolution with Rwork and Rfree of 19.0% and 23.1% respectively. The structure of the WT YLDV-IL18BP:IL18 (8S) complex was refined to 2.7 Å resolution with Rwork and Rfree of 21.9% and 27.0% respectively. The final models are of good refinement statistics for both complexes as shown in Table 1. All molecular graphic figures were generated with PYMOL [43].
The SPR analysis was done essential as described previously [15], [16]. Briefly, biotinylated IL18BP [∼300 resonance units (RU)] or IL18 (∼250 RU) was captured onto a BIAcore CM5 chip coated with streptavidin. Various concentrations of IL18 (from ∼1 nM to ∼40 nM) or IL18BP (from ∼2 nM to ∼60 nM) were injected at a flow rate of 20 µl/min. The chip coated with IL18BP was regenerated with a 10-µl injection of 1 M NaCl, 50 mM NaOH, while the chip coated with IL18 was regenerated with a 10-µl injection of 10 mM glycine (pH 2.5). The sensorgrams were analyzed with BIAEVALUATION software (BIACORE). The binding data from the injection of at least five different concentrations of analyte were globally fitted to a 1∶1 binding model. Analyses with the same concentration series were done twice.
Protein Data Bank: structure factors and atomic coordinates for the WT YLDV-IL18BP:IL18 (8S) and the triple-cysteine mutant YLDV-IL18BP:IL18 (8S) complexes have been deposited with accession codes 4EEE and 4EKX, respectively.
|
10.1371/journal.pcbi.1002454 | Quantifying Type-Specific Reproduction Numbers for Nosocomial Pathogens: Evidence for Heightened Transmission of an Asian Sequence Type 239 MRSA Clone | An important determinant of a pathogen's success is the rate at which it is transmitted from infected to susceptible hosts. Although there are anecdotal reports that methicillin-resistant Staphylococcus aureus (MRSA) clones vary in their transmissibility in hospital settings, attempts to quantify such variation are lacking for common subtypes, as are methods for addressing this question using routinely-collected MRSA screening data in endemic settings. Here we present a method to quantify the time-varying transmissibility of different subtypes of common bacterial nosocomial pathogens using routine surveillance data. The method adapts approaches for estimating reproduction numbers based on the probabilistic reconstruction of epidemic trees, but uses relative hazards rather than serial intervals to assign probabilities to different sources for observed transmission events. The method is applied to data collected as part of a retrospective observational study of a concurrent MRSA outbreak in the United Kingdom with dominant endemic MRSA clones (ST22 and ST36) and an Asian ST239 MRSA strain (ST239-TW) in two linked adult intensive care units, and compared with an approach based on a fully parametric transmission model. The results provide support for the hypothesis that the clones responded differently to an infection control measure based on the use of topical antiseptics, which was more effective at reducing transmission of endemic clones. They also suggest that in one of the two ICUs patients colonized or infected with the ST239-TW MRSA clone had consistently higher risks of transmitting MRSA to patients free of MRSA. These findings represent some of the first quantitative evidence of enhanced transmissibility of a pandemic MRSA lineage, and highlight the potential value of tailoring hospital infection control measures to specific pathogen subtypes.
| Different strains of hospital pathogens may differ in their ability to spread between patients and respond differently to control measures. Attempts to quantify such between-strain variation are lacking in high prevalence settings. We analysed data from concurrent outbreaks with different MRSA strains in two adult intensive care units. MRSA is usually carried by patients asymptomatically, and most of our data came from routine screening swabs used to detect such carriage. We divided strains into two groups: common United Kingdom strains and strains from a type often found in Southeast Asia. We developed a new method to estimate how transmission changes over time and compared results with those from an adaptation of a previously described approach. An advantage of the new method is that it makes weaker assumptions about the process generating the data. The methods gave broadly similar results: the introduction of daily antiseptic bodywashes for all patients was the only intervention associated with a substantial fall in transmission, but this intervention was less effective for the Asian strain. This work should be useful for assessing the between-strain variation in the transmission of other hospital pathogens, and for assessing the impact of interventions on patient-to-patient transmission.
| Methicillin-resistant Staphylococcus aureus (MRSA) is responsible for a high burden of morbidity and mortality worldwide [1]–[4]. While community-associated MRSA is becoming increasingly important globally [5], [6], in many countries, including the United Kingdom, MRSA remains predominantly a nosocomial pathogen [7], [8]. The dominant sequence type (ST) in Asia is ST239, and recent analysis of whole-genome sequence data has shown that this ST has distinct lineages in Asia, Europe and South America which probably share a European ancestor [9]–[11]. Little is known about what has enabled this ST to be so successful, or whether its propensity to transmit between hosts differs from other MRSA types in certain settings. A recent concurrent outbreak due to an ST239 MRSA strain (ST239-TW, subsequently referred to as TW) and the two dominant endemic UK MRSA types (ST22 and ST36, which we refer to as non-TW) in two linked adult intensive care units in a London teaching hospital provided a rare opportunity to compare the transmissibility of different MRSA types in the same clinical setting [12].
The transmissibility of a potentially emerging pathogen (the rate at which it spreads from an infected host to exposed susceptible hosts) is an important factor in determining its success and, in the case of an established pathogen, for estimating how effective interventions must be to bring an epidemic under control [13], [14]. Quantifying the degree to which strains of a nosocomial pathogen differ in their transmissibility in a particular setting could lead to a better understanding of why major clonal replacements occur. Measuring how such transmissibility changes in response to interventions would allow us to quantify the value of specific control measures, which may vary according to the strain [15]. This could lead to better resource use by allowing us to choose control measures appropriate for the specific strain. Such an analysis has greatest relevance for predominantly clonal organisms, such as S. aureus, where distinct lineages cocirculate over extended periods of time [10].
A fundamental measure of the overall transmission potential of a pathogen in a given setting is the basic reproduction number, . This is defined as the mean number of secondary cases generated by a typical case in a fully susceptible population [16], [17]. If the transmissibility of each infected host remains constant throughout its infectious period, and if each infected host has an equal chance of infecting each susceptible host, then is simply the product of the mean rate at which an infected host generates secondary infections and the mean infectious period (provided the two are not correlated).
The self-sustaining chain reaction that constitutes a major epidemic is possible only if is greater than one. If it is less than one, although there may be some self-limiting chains of secondary transmission following the introduction of an index case (and quite large clusters become possible as approaches one), this will not lead to a sustained increase in cases and, in a large population, only a small proportion of susceptible hosts will be infected [18]. An important related number is the net (or effective) reproduction number, . This is defined as the average number of secondary cases generated by a case infected at time , accounting for incomplete host susceptibility to infection and control measures in place. If is greater than one at time , the epidemic will (on average) be growing. If is less than one, it will be declining [16], [19].
These reproduction numbers are central to a mechanistic understanding of infectious disease epidemiology, and a number of methods for estimating them from different types of surveillance data have been devised [16], [19]–[23]. However, epidemics that predominantly affect hospitalized patients require some special considerations. First, unlike the community setting, the population of those exposed to infection changes rapidly over time as patients are admitted and discharged. Second, most common nosocomial pathogens are bacteria which can be carried asymptomatically over long periods, during which time colonized hosts may have several hospital admissions. This can give rise to distinctive dynamics: in addition to the usual explosive outbreaks, we also see epidemic patterns characterized by a sequence of self-limiting clusters of transmission which, over time, become more frequent and eventually coalesce into an exponentially growing epidemic [24], [25].
The concept of the single admission reproduction number, , can help in the understanding of these features of hospital epidemics [25], [26]. is defined as the mean number of secondary cases caused by a typical infectious patient during a single admission to a particular hospital or ward otherwise free of the pathogen. Necessarily, is less than or equal to . However, if and then every outbreak will be locally controlled in the short term, but, with repeated challenges to the hospital, long-term control failure will be inevitable. This results from the persistence of carriage following discharge which, over time, leads to a gradual increase in numbers colonized on admission. To account for changing numbers of susceptibles, we can also define a net single admission reproduction number, . This is analogous to and represents the average number of secondary cases generated during a single hospital/ward admission where not everyone is necessarily susceptible.
Direct ascertainment of and would be possible if we could reliably assess who infected whom during a hospital outbreak. In practice, even with detailed surveillance and molecular typing data, there is almost always considerable uncertainty about the true transmission tree. Instead, computationally-intensive approaches based on fitting mechanistic mathematical models to data which account for uncertainty in transmission routes and screening data represent the state-of-the art for analysing nosocomial transmission dynamics [27]–[32]. However, such approaches require detailed data on both susceptible and colonized or infected patients, and an assumption that temporal changes in the transmissibility can be described parametrically by some standard functional form (most commonly, piecewise constant). As currently implemented they do not allow direct estimates of the number of transmission events associated with each patient.
The aims of this paper are twofold: to describe a new approach (method 1) for estimating using hospital surveillance data; and to use it to analyse MRSA data from concurrent outbreaks with different MRSA types (TW and non-TW) in two linked adult intensive care units (ICUs). The method is simple to use and enables us to track how changes over time without the assumption that changes in transmissibility follow a fixed functional form, and without requiring data on susceptible patients. The method extends techniques for the probabilistic reconstruction of epidemic trees developed for analyzing foot and mouth disease and SARS data [21], [33]–[35]. We contrast results using this approach with that from a fully parametric mechanistic model (method 2), which represents an adaptation of previously described parametric models for nosocomial infection to a multistrain system [31], [32]. This second approach allows to be estimated. It requires more detailed data and stronger assumptions, but allows us to explicitly test hypotheses about how transmissibility is affected by interventions, and how it varies between different wards and subtypes of MRSA. While between-clone and between-ward differences in single admission effective reproduction numbers, , calculated using method 1 may be caused by differences in transmissibility, number of susceptibles, and lengths of stays, with method 2 we assume all MRSA positive patients have the same length of stay distribution and explicitly adjust for different numbers of susceptible patients when calculating .
Under baseline assumptions, on ICU1 there were 282 MRSA importation events (episodes where patients were assumed to be MRSA positive when admitted to the ICU) and 132 acquisition events. These comprised of 12 importations and 23 acquisitions with TW MRSA and 270 importations and 109 acquisitions with non-TW MRSA. On ICU2 there were 285 importations (25 with TW) and 166 acquisitions (43 with TW) (figure 1). Importations with non-TW to the respective ICUs decreased from 0.20 and 0.19 per day in phase 1 to 0.11 and 0.13 per day in phase 4. In contrast, importations with TW MRSA peaked in phase 2 in both ICUs (at 0.03 and 0.12 per day) and were at or below 0.01 per day in phases 1 and 4. Amongst patients who were MRSA positive on admission the median length of stay was 12 days (inter quartile range [IQR]4, 18) for TW-positive patients and 6 days (IQR 3, 13) for non-TW patients (, Wilcoxon rank sum test with continuity correction). For patients who acquired MRSA the corresponding numbers were 26.5 (13.25, 42.5) for TW patients and 19 (12, 29.25) for non-TW patients (). There was no evidence that length of stay differed by ward (), or by study phase (, Kruskal-Wallis rank sum test).
Over the study period (January 2002 to April 2006) there were three interventions (referred to as A, B and C) and these define four study phases. Estimated net single admission case reproduction numbers (expected number of secondary cases per case during a single ward admission) associated with each MRSA-positive patient episode are shown in figure 2 (bottom panel) together with histograms of case reproduction numbers for each ward and study phase (top panel). These highlight wide between-patient variability which decreases in the second half of phase 4 when transmission is reduced and the TW clone is eliminated. While most patients have a very low expected number of secondary cases, 22 out of 103 patients (21%) with TW MRSA are expected to transmit to at least one other patient. Corresponding numbers for non-TW MRSA are 40 out of 762 (4%). This proportion was consistently higher for TW MRSA in all four study phases: 25, 11, 18 and 31% versus 7, 0, 9 and 2% for non-TW MRSA.
Aggregating these reproduction numbers into four-week intervals highlights the temporal trends, differences between wards and impact of interventions (figure 3). In ICU1 these suggest similar patterns of transmission for the different MRSA types for the period prior to intervention C (a surface antiseptic protocol). In contrast, there were marked differences between MRSA types in ICU2 and throughout the study period the four-week averaged reproduction numbers for the TW clone usually exceeded those for non-TW clones when both types were present. These differences are also seen when reproduction numbers are averaged over study phases (table 1); the TW clone had a higher reproduction number than the non-TW MRSA in each phase in ICU2 but not in ICU1. Reproduction numbers for TW MRSA were also more volatile than those for non-TW MRSA in ICU2. There was evidence from both units to suggest differences between the MRSA types in their response to infection control interventions: while the net reproduction number of non-TW MRSA fell to a low level following intervention C in both ICUs, this was not the case for the TW clone which continued to transmit for several months at pre-intervention levels. Eventually, the TW outbreak came to an end after all patients with TW MRSA were treated empirically with systemic antibiotics (linezolid) from 1st September 2004 [12], [15]. After this intervention, although patients with TW MRSA continued to be imported into the ICUs, only three isolated apparent transmission events occurred (figure 1). When reproduction numbers for the two ICUs combined were estimated (allowing for cross transmission between ICUs) the results suggested the reproduction number of the TW clone was consistently higher than that for the non-TW MRSA and varied little throughout the study period (table 1). Reproduction numbers for the non-TW clones, in contrast, fell in phase 2 and 4. These results were not highly sensitive to the assumed strength of coupling between the two ICUs or to the MRSA acquisition assumptions (supplementary table S2).
Results from method 2 showed broad agreement with these findings, but in contrast to method 1 made a priori assumptions about the timing of the changes in transmissibility (tables 2–3, figure S1). On both wards, averaging over all phases, there was about a 1 in 400 chance of a given susceptible patient acquiring MRSA from a particular MRSA-positive patient on a particular day (table 2). When estimates of daily transmission probabilities from a single MRSA positive patient were constrained to take the same values for the TW and non-TW clones but were allowed to vary by ward and study phase, we found clinically significant variation between the four study phases in both ICUs (table 2). In particular, while estimates were similar in phases 1 to 3, there was a marked reduction in phase 4. There was no strong evidence that these joint estimates (for all MRSA clones) varied by ICU in any of the study phases (table 2). These findings were robust to the assumptions made about acquisition events and times (supplementary table S3).
Extending this analysis to allow transmission probabilities to vary with the MRSA type enabled differences between MRSA clones and wards to be quantified (table 3) and allowed hypothesis tests about whether the daily transmission probabilities differed between strains (thus allowing us to test whether the observed differences in transmissibility found using method 1 could be entirely explained by the longer length of stay of the TW patients). In ICU1 no consistent differences were seen. In ICU2 the TW clone had a higher daily transmission probability to susceptible patients in each of the four study phases under the baseline assumption of complete bacterial interference, though confidence intervals were wide and showed considerable overlap. Differences between TW and non-TW MRSA transmission probabilities reached statistical significance at the 5% level for both wards combined and for ICU1, but not for ICU2 alone, and in only one of the four phases (phase 4) using combined data from both ICUs. This phase corresponded to the introduction of the surface antiseptic bodywash protocol, which was associated with a more than halving of the transmission probability from a patient with non-TW MRSA compared to earlier phases. The fall in the transmission probability for TW MRSA in phase 4 was smaller, and likely to be confounded by the use of linezolid for TW carriers in this phase. Large differences in transmission probabilities for the two MRSA types were also seen in phase 2 (corresponding to the introduction of hand hygiene promotion), but in this case confidence intervals were wider reflecting the short duration of this phase. In other phases differences between TW and non-TW estimates were much smaller. The magnitude of the differences depended on which patients were assumed to be susceptible. Under the baseline assumption that patients colonized with one strain were not susceptible to acquiring another (complete bacterial interference) the differences were larger than in the sensitivity analysis where no bacterial interference was assumed (table 3). This can be explained by the higher prevalence of non-TW MRSA clones; under the assumption of no bacterial interference all non-TW MRSA positive patients would be considered susceptible to infection or colonisation with TW MRSA and vice versa. Changing from complete interference to no interference therefore results in a greater increase in the number of susceptibles available for the TW clones to infect than it does for the non-TW clones. To accommodate these changes, a larger reduction in the daily transmission probability for TW clones is required. Overall, combining data from both wards, the TW clone was estimated to have a daily transmission probability that was between 63 and 100% higher than the non-TW clones in phase 4, and between 53 and 94% higher in phase 2 (the lower numbers corresponding to the no bacterial interference assumption) though the differences only reached significance at the 5% level in phase 4 and only under baseline interference assumptions. Transmission probabilities were broadly similar in the two other study phases. Estimates of the single-admission reproduction number () from the model without background transmission (and assuming TW and non-TW patients have the same length of stay distribution) are reported in the supplementary material (figure S1). Results were robust to the assumptions made about the number and timing of MRSA acquisition events (supplementary table S4), but fitting a more complex model allowing patient-to-patient transmission and transmission from background sources suggested that the relative importance of patient-to-patient and background transmission could not be reliably identified in such hyperendemic settings without additional data (supplementary table S5).
Common bacterial nosocomial pathogens have distinct dynamics from typical community pathogens and call for different analytical approaches. Important features of hospital epidemics with such organisms include: i) a host population that changes rapidly over time in comparison with the timescale of epidemic dynamics; ii) a high proportion of infected (or colonized) hosts who are already infected when they enter the population (the hospital or ward); iii) a dominant role for asymptomatic infection so infected hosts can usually only be identified using screening swabs, leading to large uncertainty in the timing of transmission events; iv) a lack of a well-defined serial interval or generation time (since asymptomatic carriage can persist for months or years, but transmission is only intermittently observed during hospital admissions). The probabilistic tree reconstruction approach described above (method 1) overcame these limitations by using a hazards-based approach applied to patient screening data to assign probabilities to potential source patients for observed acquisition events. Using hazards in this way to reconstruct epidemic trees and estimate reproduction numbers appears to have first been suggested by Kenah et al [36]. Results using this method were supplemented with a maximum likelihood approach (method 2) where the timing of cross-infection events was assumed to be known but which allowed estimation of the daily transmission probability, enabling us to study effects related to study phase and MRSA type while controlling for differences in length of stay.
These methods were applied to data from two adjacent general ICUs in which admission and weekly MRSA screens and culture results from clinical samples identified patients admitted with and acquiring MRSA over a four year period. During that time there was sustained transmission with endemic MRSA and a newly introduced TW variant.
Both analytical methods supported the hypothesis that intervention C (the surface antiseptic protocol) was associated with a sustained reduction in MRSA transmission, and both indicated a reduced effect for the TW clone. Both methods gave point estimates that indicated elevated transmission of TW MRSA compared with endemic strains in all four study phases in ICU2 but not ICU1. There were, however, some differences: the ward-level reproduction numbers (method 1) tended to indicate greater increased transmission for the TW compared to non-TW MRSA than was seen using method 2. This reflects the fact that the two methods are quantifying different things: method 1 estimates secondary cases per case, which depends both on transmissibility and the length of ICU stay while carrying MRSA; method 2, in contrast, estimates only the daily transmission probability from one MRSA carrier to one susceptible patient. This will not be affected by length of stay. Indeed, there was some evidence that patients colonised with TW MRSA (particularly those colonised on ICU admission), had a longer length of stay than those colonised with non-TW MRSA. This may reflect the link between MRSA infection and excess length of stay in this cohort [37], and the increased virulence of the TW strain which was over four times more likely to cause blood stream infection in colonised patients compared to non-TW MRSA strains in the same ICUs [12]. Even in the absence of an increased rate of transmission to other patients, increased length of stay would lead to a higher single-admission reproduction number. It is possible that such differences in length of stay reflect underlying differences in the characteristics of patients most vulnerable to acquiring the different MRSA types. For example, because the TW outbreak was centred on the two ICUs, patients carrying TW on ICU admission might be more likely than patients carrying non-TW MRSA to have had recent ICU admissions. The TW clones showed a far broader range of antibiotic-resistance than endemic MRSA clones and have previously been shown to preferentially colonise vascular catheters but not carriage sites compared with endemic strains [12]. Taken together, these observations suggest that the TW MRSA could represent a phenotype particularly adapted to transmission in settings, such as ICUs, with high levels of antibiotic usage and patient catheterisation, perhaps at the expense of persistence outside these areas. There is some evidence that such adaptation results from both increased persistence in the ICU (perhaps by targeting long-stay patients, and causing infections that increase length of stay) and from an increased daily transmission probability (particularly in the presence of widespread antiseptic use). Caveats, of course, apply: differences in lengths of stays between TW and non-TW colonized/infected patients could be confounded by exposure history (the recent arrival of the TW clone rather than its biological properties may account for the different patient characteristics). Differences in daily transmission probabilities could also be subject to such confounding and could also have arisen by chance (in all phases – even phase 4, where the effect size was largest – confidence intervals were wide).
The mechanisms underlying variations in transmissibility of different MRSA (and S. aureus) strains are poorly understood. Reasons for the differences in the two ICUs are also unclear. Chance variation cannot be ruled out, as the formal investigation of transmission potential of different MRSA types was, in part, motivated by perceived differences in transmissibility (using the same data), and the usual limitations of post hoc analyses therefore apply. Also, although the analyses accounts for demographic stochasticity, there may also be important sources of environmental stochasticity which are not accounted for. It seems unlikely that the difference in TW transmission in the two ICUs can be explained by colonized staff: a universal staff screening programme failed to detect the TW clone during the outbreak [12]. Differences in infection control practice also seem unlikely but cannot be ruled out: the two wards share the same infection control policies and staff pool, with medical and nursing staff rotating between units at 3–6 monthly intervals, though only physiotherapy, radiology and pharmacy staff worked across both units at the same time. It is possible that the built environment influences MRSA transmission. ICU2 was last refurbished in 1969, retaining a mixture of original materials including wood, and has much less open space, only eight sinks, and one side room, whereas ICU1 was refurbished in 1999 to an open plan configuration with better space utilization, 19 sinks and three side rooms. The reduced availability of sinks, side rooms and space to circulate may have adversely affected the ability to carry out infection control practice or cleaning, although it is unclear why this should only affect TW MRSA, which was not detected on environmental screening during the outbreak [12].
Despite anecdotal reports that some lineages of S. aureus strains have an enhanced epidemic potential in hospital settings [38], objective assessments of between-strain variation in transmissibility are largely lacking. Such variation is nonetheless to be expected given the large degree of phenotypic variation in different S. aureus and MRSA clones, and the dominance of a small number of MRSA lineages [39]. One of the few instances where the nosocomial transmission potential of different subtypes of the same nosocomial pathogen have been quantified comes from a comparison of the onward transmission from patients admitted to hospitals in the Netherlands carrying MRSA [40]. In this case, because MRSA introductions were infrequent (as MRSA prevalence in hospitals in the Netherlands is below 1%) and contact tracing extensive, the secondary cases could be assigned to distinct clusters of transmission following identified introductions. This allowed the authors to use methods based on a branching process model to estimate the single admission reproduction number, [41]. It was found that newly admitted ST398 MRSA strains (which are commonly associated with livestock production) had a greatly reduced propensity to spread compared with other MRSA sequence types, with an value (95% CI) of only 0.16 (0.04–0.40), about one sixth of the corresponding value for non-ST398 MRSA. The authors concluded that less stringent control measures were likely to be sufficient to control ST398 MRSA clones than those needed for non-ST398 MRSA types.
Such methods would not have been applicable for our data, and the first method used here to quantify the transmissibility of different strains (method 1) instead built on recent approaches to estimate reproduction numbers by probabilistically reconstructing epidemic trees. Such tree-reconstructions have used simple rule-based methods, for example assigning sources from a candidate list based on proximity data [33], more formal semi-parametric methods using partial likelihoods and assuming a known serial interval distribution [21], [34], and, most recently, semi-parametric hazard-based approaches [36]. Hazard-based approaches have some advantages over the first two methods: they avoid some of the arbitrary assumptions of the rule-based approaches, do not require knowledge of the serial interval distribution, and can avoid biases that arise from the fact that the serial interval distribution changes over the course of an epidemic. Advantages over approaches based on fitting a full transmission model include fewer assumptions, in particular with regard to the functional form of changes in the transmission potential over time. In this respect, tree reconstruction approaches have some similarities with other semi-parametric approaches that make use of survival analytical methods, such as the approach adopted by Wolkewitz et al., who derived non-parametric estimates of a time-varying transmission rate changed over time using a Martingale-based method [42]. An important difference in the current approach is that we are specifically interested in estimating how the distribution of the number of secondary cases resulting from each case changes over time. The method 1 approach described here also makes relatively low demands for data (with no information required for patients who do not become colonized or infected), has a low computational burden, and can be easily adapted to cope with co-circulating subtypes as in the application here. This approach is appropriate when the daily probability of a patient acquiring MRSA is small, as in this case reconstructed epidemic trees will be approximately independent of this probability. This approximation is likely to be reasonable for all but the most explosive outbreaks. For example, using the exact formula we found that changing this probability from a baseline of 0.005 to 0.001 and 0.025 changed the estimated mean reproduction numbers for each phase and MRSA type by less than 3%.
Two assumptions underlying the analytical approaches used here are i) that new MRSA acquisitions can be explained by patient-to-patient spread within the units (which is likely to be mediated by contacts with transiently colonized healthcare workers) and ii) that risk of transmission increases in line with colonization pressure (the number of patients with MRSA on the ward). While these assumptions are supported by observational and quasi experimental studies [43], [44], it would be desirable to more rigorously challenge them. Unfortunately, unpublished simulation studies and analysis here with a more complex model allowing different transmission routes (table S5) both suggest that the ability to identify the relative importance of background and patient-to-patient transmission may be limited in hyper-endemic settings in the absence of more discriminatory typing data. The inability of our typing methods to reliably distinguish between non-TW MRSA types, or to identify genetic variants of the TW clone therefore represent important limitations of this work. High resolution genotyping data would enable more definitive assessments of who infects whom, and therefore allow us to quantify the risks of transmission of different MRSA subtypes in different wards at different times with greater certainty.
Figure 1 confirms that not all acquisition events can be explained by transmission from a known MRSA positive patient from the same ward. The combined-ICU analysis, allowing for between-ward transmission, is able to account for some MRSA acquisitions where no known source was present on the same ward, and this explains why combined ICU estimates of the reproduction number are sometimes outside the range of individual ICU estimates (table 1), but unknown MRSA sources are also likely to be present in the patient population [32]. A full model-based analysis using data augmentation (which estimates model parameters and latent parameters that represent “unobserved” - or augmented - data, typically using Markov chain Monte Carlo methods for fitting) could account for such unknown sources. Such an approach retains some important advantages for analysing typical surveillance data. These include the ability to account for imperfect swab sensitivity and for uncertainty in the number and timing of acquisition events, circumventing the need to make arbitrary assumptions about which patients were colonized on admission to a ward. In the present context such an analysis would allow us to explicitly account for the change in the screening protocol in November 2004. Since this involved screening more body sites, it is likely to have increased screening sensitivity and led to increased detection of MRSA, potentially biasing the estimated effect of intervention C. To date, however, no published work has adapted such approaches to cope with multiple co-circulating subtypes. The method 2 used here can be thought of as a simplified version of such an approach (in that it is based on a fully-specified mechanistic transmission model) but it avoids the complexities of data augmentation by assuming the epidemic process is perfectly observed. An important area for future work will be to extend data augmentation methods to cope with carriage of multiple types. Such approaches have been developed for the sequential carriage of community pathogen subtypes [45]. Addressing issues of co-colonisation with different subtypes may be particularly important for some nosocomial pathogens, and neglecting such effects is a potential source of bias. In our analysis here we considered two possibilities – complete bacterial interference (where one strain completely inhibits the acquisition of another), and no bacterial interference. The reality may lie somewhere between these two extremes. Such an analysis will be complicated by the fact that routinely-used laboratory methods are not well-suited to detecting the simultaneous carriage of multiple types [46], and sensitivity for detecting a second type will not, in general, be the same as sensitivity for detecting a single type.
Ethical approval for this research was granted by the NHS National Research Ethics Service, South East Research Ethics Committee. All data were analyzed anonymously.
Anonymised data from two 15-bed adult general intensive care units (ICU) within a 1050-bed teaching hospital in London, United Kingdom, were collected between 1st January 2002 and 30th April 2006 as described elsewhere [12], [15]. Dates of admission and discharge and MRSA culture results from screen and clinical samples were analysed for all 4,570 consecutive patient admissions to both ICUs. Infection control policies were in place including specifying hand hygiene between patient contacts and use of contact precautions for known MRSA colonized patients throughout. On this background three main new MRSA control interventions were introduced: intervention A (introduced on 15th July 2003) was an education campaign to promote hand hygiene and barrier nursing; intervention B (introduced on 15th October 2003) was isolation of known MRSA colonized patients in side rooms or in patient and nursing cohort pairs; intervention C (introduced on 26th April 2004) was a surface antiseptic protocol which included daily chlorhexidine bodywashes for known MRSA positive patients, and daily triclosan bodywashes for other patients. The three interventions defined four study phases for analysis: phase 1 from 1st January 2002 to 14th July 2003; phase 2 from 15th July to 14th October 2003; phase 3 from 15th October 2003 to 25th April 2004; and phase 4 from 26th April 2004 to 30th April 2006. Patients were swabbed for MRSA carriage on admission and every Monday morning. Swabs were taken from nose, axillae and perineum until 1st November 2004, when additional rectal and throat samples were included (a change associated with an approximate 30% increase in the proportion of patients identified as carriers on admission to ICU) [47]. Clinical samples were collected when infection was suspected. S. aureus colonies were identified using a combination of catalase positivity, Staphaurex (Remel Europe Ltd., Dartford, England) and/or salt mannite positivity with confirmation by a tube coagulase test. Methicillin resistance was determined by disc testing. Screen samples were identified using a selective mannitol broth technique [47].
TW MRSA was defined initially by its distinctive and extensive antimicrobial resistance pattern, sequence typing and microarray analysis [12]. More extensive typing of available admission and acquisition isolates has shown all antimicrobial resistance patterns defined TW isolates to belong to CC8/239 and non-TW isolates to be ST22 and ST36 [15]. When TW and non-TW MRSA isolates were recovered from the same patient, only the first type recovered was considered. This was, however, rare: two patients had both types recovered from pooled screening sites; nine had both types from sputum; and seven had both types from wounds. Thirteen patients had both types recovered from different sites. Further details of patient characteristics, interventions, swabbing sites and microbiological procedures have been described elsewhere [12], [15].
We analyse the data using two separate approaches which we refer to as method 1 and method 2. In both analyses we define a new MRSA acquisition to have occurred if a patient has a negative admission screening swab, a subsequent MRSA positive screen or clinical sample while in the ICU and more than 48 hours after being admitted to the ward, and no prior MRSA positive isolate in the 90 days preceding ICU admission. Patients with any MRSA positive samples taken within 48 hours of admission are assumed to be positive on admission (MRSA importations). A patient who is believed to be neither colonized nor infected on a given day is assumed to be susceptible to becoming colonized or infected by either MRSA type (see supplementary material for further details). In the first approach (method 1), which probabilistically reconstructs the epidemic tree, we assume that the acquisition occurred one or more days before the first positive screening swab. In the second approach (method 2) we assume a new acquisition to have occurred on day if a patient has his or her first MRSA positive swab on day , following a negative MRSA admission screening swab during the same ward admission.
We also assume i) that once MRSA positive, a patient remains so until ward discharge (hence no information from swab results after the first positive is used), and ii) MRSA-positive patients only become potential sources for transmission to other patients after their first positive swab, unless they are assumed to be positive on admission, in which case they are potential sources from their date of admission. For patients readmitted to one of the wards following ward discharge, we apply the same criteria that we use for first time admission to determine admission colonisation status. We use a time unit of one day, and take dates of admission and discharge to represent the first and last whole days of a patient admission.
|
10.1371/journal.pntd.0005219 | CD4/CD8 Ratio and KT Ratio Predict Yellow Fever Vaccine Immunogenicity in HIV-Infected Patients | HIV-infected individuals have deficient responses to Yellow Fever vaccine (YFV) and may be at higher risk for adverse events (AE). Chronic immune activation–characterized by low CD4/CD8 ratio or high indoleamine 2,3-dioxygenase-1 (IDO) activity—may influence vaccine response in this population.
We prospectively assessed AE, viremia by the YFV virus and YF-specific neutralizing antibodies (NAb) in HIV-infected (CD4>350) and -uninfected adults through 1 year after vaccination. The effect of HIV status on initial antibody response to YFV was measured during the first 3 months following vaccination, while the effect on persistence of antibody response was measured one year following vaccination. We explored CD4/CD8 ratio, IDO activity (plasma kynurenine/tryptophan [KT] ratio) and viremia by Human Pegivirus as potential predictors of NAb response to YFV among HIV-infected participants with linear mixed models.
12 HIV-infected and 45-uninfected participants were included in the final analysis. HIV was not significantly associated with AE, YFV viremia or NAb titers through the first 3 months following vaccination. However, HIV–infected participants had 0.32 times the NAb titers observed for HIV-uninfected participants at 1 year following YFV (95% CI 0.13 to 0.83, p = 0.021), independent of sex, age and prior vaccination. In HIV-infected participants, each 10% increase in CD4/CD8 ratio predicted a mean 21% higher post-baseline YFV Nab titer (p = 0.024). Similarly, each 10% increase in KT ratio predicted a mean 21% lower post-baseline YFV Nab titer (p = 0.009). Viremia by Human Pegivirus was not significantly associated with NAb titers.
HIV infection appears to decrease the durability of NAb responses to YFV, an effect that may be predicted by lower CD4/CD8 ratio or higher KT ratio.
| Yellow Fever (YF) vaccine is considered one of the most effective vaccines ever produced. However, previous studies suggest that HIV impairs YF vaccine response. In this study, we assessed if HIV infection impacts the risk of adverse events and could reduce antibody response to YF vaccine. We explored if laboratory markers of persistent inflammation, frequently present among HIV-infected patients, could predict antibody response to YF vaccine in this population. We found that HIV had no significant effect on adverse events or levels of antibodies through 3 months after vaccination, but this may be limited by the small sample size of 12 HIV-infected and 45-uninfected participants in the study. However, we were able to show that, compared to HIV-uninfected participants, HIV–infected patients had lower antibody titers 1 year following YF vaccine even after statistical adjustment for the potential effects of sex, age and prior vaccination. Persistent inflammation seems to reduce YF vaccine antibody response in HIV-infected participants. In conclusion, HIV-infected individuals have impaired antibody response to YFV due to a poorer persistence of antibodies, despite a seemingly normal initial response. HIV-infected patients at permanent or recurring risk of YF infection may benefit from a booster dose of YF vaccine.
| Effective antiretroviral treatment (ART) drastically improved clinical outcomes for people living with HIV. However, these patients still present increased risk of death, higher prevalence of comorbidities, and impaired responses to vaccines [1–6]. Prior studies have shown impaired Yellow Fever vaccine (YFV) immunogenicity among HIV-infected persons is associated with detectable HIV viral load (VL) [7–12] and lower CD4 T cell counts [11]. However, it is still unclear whether reduced YFV antibody response among HIV-infected individuals is caused by a blunted initial response, decreased persistence of antibodies, or both. Moreover, predictors of YFV immunogenicity among patients with effective and early ART are not well known.
More recently, studies including patients with early initiation of ART have suggested a negative effect of persistent immune activation on responses to Influenza vaccine [13, 14], Neisseria meningitis vaccine [15] and YFV [16, 17] in both HIV-infected and–uninfected individuals. This is consistent with previous studies that demonstrate excessive immune activation and inflammation predict residual morbidity and mortality in treated HIV-infected patients [18–20]. A range of biomarkers has been used in different settings to quantify persistent immune activation [20]. One increasingly appraised indirect biomarker is the ratio of CD4 to CD8 T lymphocytes, or CD4/CD8 ratio. Previous studies have shown that CD4/CD8 ratio correlates with markers of CD8 T cell activation, and a lower CD4/CD8 ratio predicts higher risk of non-Aids events and mortality among ART-treated HIV-infected patients [21–23]. Furthermore, a low CD4/CD8 ratio is also strongly associated with the activity of Indoleamine 2,3-dioxygenase-1 (IDO), an enzyme expressed by activated myeloid cells in HIV and other inflammatory conditions that causes adaptive immune defects. IDO catabolizes tryptophan (T) to kynurenine (K) and other metabolites that may contribute to proliferative lymphocyte defects, regulatory T cell expansion, microbial translocation and immune activation in treated HIV infection [24]. Therefore, elevated IDO activity (as measured by plasma KT ratio) may also indicate adaptive immune dysfunction in this population. Finally, chronic co-infection with Human Pegivirus has been associated with reduced innate and adaptive immune activation among HIV-infected patients in prior studies [25–27].
An additional relevance of YFV in HIV-infected patients concerns the theoretic higher risk of YFV-associated severe adverse events (AE) in this population [28, 29]. YFV is produced from the 17D or 17DD attenuated viral strains, and although mechanisms for YFV-associated AE are not completely elucidated, immunosuppressed patients and persons at extremes of age are considered at increased risk [28, 30]. It is hypothesized that immune response fails to contain the vaccine virus replication, typically seen only in the first 3–5 days after vaccination, leading to uncontrolled viral dissemination and clinical disease [31].
In this study, HIV-infected patients and controls referred to receive YFV were enrolled and prospectively followed for 1 year after vaccination. We addressed clinical and laboratory AE and viremia by the YFV virus. In addition, we investigated if HIV status was associated with titers of Yellow Fever (YF)-specific neutralizing antibodies (NAb) in the first 3 months following vaccination (henceforth defined initial YFV immunogenicity) and one year following vaccination (henceforth defined persistence of YFV immunogenicity). Finally, we explored if correlates of immune activation including CD4/CD8 ratio, KT ratio, and Human Pegivirus co-infection predict YFV response among HIV-infected subjects.
Subjects aged 18 years old and older who were referred to receive YFV at Clinics Hospital in Sao Paulo, Brazil, between October 2011 and April 2014 were screened for participation.
Potential participants were evaluated by an attending physician who determined whether YFV was indicated based on risk of exposure to wild YF and YFV contraindications as defined by National Guidelines. The Guidelines do not recommend YFV to pregnant and breastfeeding women, subjects under immunosuppressive medications and patients with conditions such as cancer and thymus dysfunction. HIV-infected patients with a CD4 T cell count above 350/ml measured in the previous 4 months were considered eligible for vaccination. At enrollment, HIV-negative persons underwent a rapid HIV-test.
For both groups, participants with immunosuppressive conditions other than HIV infection were excluded. These included diabetes, chronic liver or kidney diseases, any type of cancer (except resolved Kaposi Sarcoma), and use of systemic immunosuppressive therapy in the last 3 months. Female participants in reproductive age underwent a pregnancy test at enrollment.
At enrollment, medical history and date of previous YFV was obtained, if applicable. A blood sample was collected for assessment of baseline complete blood count and liver enzymes, CD4 and CD8 T cell counts, plasma KT ratio and Human Pegivirus viremia. HIV-infected participants had HIV VL measured at baseline and all subsequent visits. Participants were followed on days 3, 5, 7, 14, 28, 56, 84, and 365 after vaccination. We measured viremia by the YFV virus on days 3, 5, 7, and 14 after vaccination, and measured titers of YF-specific NAb at baseline and on days 7, 14, 28, 56, 84 and 365 after vaccination. We also collected data on spontaneous and solicited clinical AE, as well as laboratory AE on visits 3, 5, 7, 14 and 28 after vaccination. We measured CD4, CD8 T cells and CD4/CD8 ratio in all visits (Fig 1).
Clinical and laboratory AE were assessed as binary variables, defined as positive if the participant had any clinical AE, or any clinically significant laboratory AE at visits 3, 5, 7, 14 or 28, and negative otherwise. Laboratory AE were considered clinically significant if graded ≥2 according to the National Institutes of Allergy and Infectious Diseases’ Division of AIDS Table for Grading the Severity of Adult and Pediatric AE [32]. Viremia by YFV virus was assessed both as numeric and binary variable. The binary variable for YFV viremia was defined as positive if the participant had a detectable measurement (>200 copies/ml) on days 3, 5, 7 or 14 after vaccination, and negative otherwise.
The HIV VL was determined by reverse-transcriptase (RT)-PCR using Amplicor HIV-1 Monitor Test (Roche Diagnostic Systems, NJ, USA), with a lower detection limit of 40/mm3. CD4 and CD8 T cell counts were determined by flow cytometry (FACSCalibur, BD Biosciences, CA, USA) using Multitest reagent (BD Biosciences).
Kynurenine and tryptophan were quantified on cryopreserved plasma samples by liquid chromatography–tandem mass spectrometry as previously described [33].
Human Pegivirus RNA was extracted from 140μl serum samples using QIAamp Viral RNA Mini Kit (QIAGEN Inc., California, USA), according to manufacturer’s instructions. A 5μl aliquot of extracted RNA was used to perform qRT-PCR with SuperScript III Platinum One-Step Quantitative RT-PCR System with ROX kit (Life Technologies), with primers and a TaqMan probe that amplified and quantified a fragment of 72-bp of the 5' untranslated region (5'UTR). The reaction was made with 0.5μl of SuperScript III RT/Platin Taq Mix, 12.5μl of 2X Reaction Mix with ROX, 0.75μl of 10μM Forward primer RTG1 (5’GTGGTGGATGGGTGATGACA3’) (Sigma), 1.25μl of 10μM Reverse primer RTG2 (5’GACCCACCTATAGTGGCTACCA3’) (Sigma), 0.4μl of 25 μM TaqMan probe ([6’FAM]CCGGGATTTACGACCTACC [TAMRA-6-FAM]) (Life Technologies), and reaction final volume of 25μl was completed with DEPC-treated water. cDNA synthesis was performed during the first 15 minutes at 50°C. After 2 minutes at 95°C, amplification and quantification were performed during 40 cycles with the following times and temperatures: 95°C, 15 seconds; 60°C, 30 seconds. The reading of FAM fluorescence was made during annealing period at 60°C.
For measurement of YFV viremia, total RNA was extracted from 140μl of plasma using QIAamp RNA Blood Mini Kit (Qiagen, Hilden, Germany) and eluted in 60μl of elution buffer. cDNA was obtained through RT reaction using 10μl of extracted RNA, 300ng of random primer (Amersham Biosciences, Piscataway, NJ, USA); 10U/μl of Super Script II RT (Invitrogen, Carlsbad, CA, USA) in a buffer solution with 0.25U/μl of ribonuclease inhibitor (Invitrogen) and 0.5mM deoxyribonucleotide triphosphates (Invitrogen), at final volume of 20μl. The reaction was incubated at 45°C for 90 minutes. Five μl of cDNA was added to 20μl of TaqMan Master Mix (Applied Biosystems, Foster City, CA, USA) and amplified by RT-PCR using the following primers and probe: (YF-NS5_F) 5’-GCACGG ATGTAACAGACTGAAGA-3’; (YF-NS5_R) 5’-CCAGGCCGAACCTGTC AT-3’ and (YF-NS5Probe) 5’-FAM-CGACTGTGTGGTCCGGCCCATC-3’–TAMRA [34]. The product was amplified using optical detection system layout of BioRad ICycler for 45 cycles at the following settings: 10 min, 95°C; 45 cycles of 15s for 94°C and 60s for 60°C.
NAb titers against YF virus were measured by Plaque Reduction Neutralization Test (PRNT) performed at Virologic Technology Laboratory of Bio-Manguinhos (LATEV, FIOCRUZ, Rio de Janeiro) as previously described [16].
The study was approved by the Ethics Committee at Clinics Hospital in University of Sao Paulo. Upon participation, all participants signed an informed consent form. HIV tests were performed with pre and post-test counseling, and all individual identifiable information was maintained in secured cabinets and electronic files.
Groups were compared using Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables. Titers of YF-specific NAb, CD4 and CD8 T cell counts were log-transformed to approximate normal distribution, and antilog transformation was required for model interpretation.
The effect of HIV status on levels of YFV viremia, and on initial YFV immunogenicity were investigated using mixed models with robust standard errors, adjusted for age, sex, previous YFV and interaction between HIV and visits. The effect of HIV status on persistence of YFV immunogenicity was investigated using a linear regression model adjusted for age, sex, previous YFV and baseline values of YF-specific NAb titers.
The effects of T CD4 and T CD8 cell count, detectable HIV VL, CD4/CD8 ratio, KT ratio, and Human Pegivirus viremia on YF-specific NAb titers among HIV-infected patients were investigated using mixed models adjusted for age, baseline NAb titers and HIV VL. Correlations between NAb titers and CD4/CD8 ratio or KT ratio were explored using Spearman rank correlation.
For all analysis, we assumed a two-sided alpha error of 0.05. All analyses were performed in Stata version 13.1 (StataCorp. College Station, TX: StataCorp LP).
We calculated sample size based on the impact of HIV status on titers of YF-specific NAb in the first 3 months following vaccination, using estimates of effect size and standard deviation from a prior study published by our group [16]. Since the analysis plan included mixed models for repeated outcomes, we assumed a 20% reduction in error variance and estimated a final sample of 33 participants per group using conventional means comparison.
Between October 2011 and April 2014, 63 participants were enrolled. Enrollment of a greater number of participants was compromised due to high refusal rates, mainly because most potential participants were referred to receive the YFV for a scheduled trip to an YF endemic area, and were therefore planning to be out of town and unable to attend the study visits. In addition, many potential participants refused participation due to the busy visit schedule, particularly in the first 2 weeks of follow-up. After exclusion of five controls and one HIV-infected participant due to missing visits, our final cohort included 12 HIV-infected and 45 HIV-uninfected individuals. All participants in the HIV-infected group were men, 11 (92%) were under ART and 8 (67%) had undetectable (<40 copies/ml) HIV VL. Participants under ART with detectable HIV VL had a range of 43–1,982 HIV copies/ml, and the only untreated participant had 110,531 HIV copies/ml at baseline. HIV-infected participants were more likely than -uninfected participants to be male and tended to be younger; 4 (33%) reported a prior YFV at a median of 14.5 years before enrollment (interquartile range [IQR] 10.5–21.5); none had received more than one YFV shot. Among HIV-uninfected participants, 16 (36%) reported a prior YFV at a median of 12 years before enrollment (interquartile range [IQR] 11–20); 10 had received a single shot and 6 (13%) had received two YFV shots in lifetime. Groups had similar YF NAb titers at enrollment, whether all participants or only participants with a prior YFV were considered (Table 1). Baseline CD4 T cell count was high among HIV-infected participants (median 722 cells/mm3, [IQR] 526–795), although still significantly lower than controls (median 941 cells/mm3, IQR 807–1470; p = 0.003). As expected, CD4/CD8 ratio was lower among HIV-infected participants (median 0.7, IQR 0.5–0.8) compared to controls (median 1.6, IQR 1.3–2.6; p<0.0001). KT ratio was also higher in HIV-infected group, but the difference did not reach statistical significance (median 35.9 versus 31.3 nM/μM, p = 0.06). Viremia by Human Pegivirus was more prevalent in HIV-infected (67%) than in -uninfected group (27%, p = 0.016, Table 1).
During study visits, HIV-infected subjects had no substantial change in CD4 T cell count compared to baseline values (Fig 2A). HIV-infected participants who had detectable HIV VL at baseline had no meaningful change in HIV VL across visits (Fig 2B). Among the 8 participants with undetectable HIV VL at baseline, only one had a detectable value (51 copies/ml) at visit 84. HIV-infected participants had no change in ART status during follow-up.
Any clinical AE was reported by 6 (50%) participants in the HIV-infected group, and 22 (48.9%) controls. All reported clinical AE (local pain, tenderness and redness; nausea, myalgia, fatigue, dizziness and fever) were mild and self-limited, as were laboratory AE–anemia, neutropenia, lymphopenia, thrombocytopenia and liver enzymes elevation–which were detected in 3 (25%) individuals in the HIV-infected group, and 13 (28.9%) controls.
Viremia by the YFV virus was detected in at least one visit in 40% of HIV-infected participants and 34% of controls. Maximum detected viremia was 11210 copies/mL in one HIV-uninfected participant at day 5 after vaccination; in the HIV-infected group, highest measured viremia was also observed at day 5 (4197 copies/mL). HIV status was not statistically associated with levels of viremia by the YFV virus (p-value = 0.99 for the overall effect of HIV on YFV viremia on days 3, 5, 7 and 14).
At baseline, 4 HIV-infected participants (33%) and 17 controls (38%) had levels of NAb considered seropositive for a cutoff of 794 mUI/mL as defined by the referent laboratory [35]. If only participants with a previous YFV were considered, 3 HIV-infected (75%) and 15 controls (94%) had seropositive NAb titers. At 28 days after vaccination, all participants in both groups were seropositive for YF, and at one year following vaccination, 11 HIV-infected participants (92%) and 43 controls (96%) maintained seropositive YF-specific NAb.
We failed to find a statistically significant difference between groups defined by HIV status on initial YFV immunogenicity in either visit individually or overall in a mixed model adjusted for age, sex and previous YFV (Table 2). The model predicted lower YF-specific NAb titers for women; in average, women had 0.33 times the titers observed for men (95% CI 0.17–0.66, p = 0.002). As expected, compared to individuals without previous YFV, those who reported a previous YFV had higher NAb titers (fold change 13.69, 95% CI 7.12–26.30, p<0.001).
Persistence of YFV immunogenicity was significantly lower in HIV-infected participants compared to controls in a mixed model adjusted for age, sex, previous YFV and baseline NAb titers. In average, HIV-infected individuals had 0.32 times the NAb titers observed for HIV-uninfected participants at one year after vaccination (95% CI 0.13–0.83, p = 0.021). We found no statistically significant effect of age, sex or previous YFV on persistence of NAb (Table 3).
In the exploratory analysis restricted to HIV-infected patients, higher CD4/CD8 ratio and lower KT ratio predicted higher YF-specific NAb titers; in average, for each 10% increase in CD4/CD8 ratio, post-baseline NAb titers were 21% higher (95% CI 3–38%, p = 0.024), and for each 10% increase in KT ratio, post-baseline NAb titers were 21% lower (95% CI 5–37% lower, p = 0.009) after adjustment for age, baseline NAb titers and HIV VL. There was no evidence for an association between CD4 or CD8 T cell count and YFV immunogenicity (multiplicative effect per 10% increase 1.05, 95% CI 0.91–1.20, p = 0.469, and 0.95, 95% CI 0.86–1.04, p = 0.295, respectively) or between Human Pegivirus co-infection and YFV immunogenicity among HIV-infected individuals (fold change 0.65, 95% CI 0.09–4.47, p = 0.659, Table 4). Adjusted for age and baseline NAb titers, having detectable plasma HIV was associated with 60% lower YF-specific NAb titers (fold change 0.40, 95% CI 0.21–0.75, p = 0.004).
As to confirm the effects of CD4/CD8 ratio and KT ratio on NAb titers, we performed simple non-parametric correlation tests; as expected, CD4/CD8 ratio correlated positively with NAb titers in all time-points, with statistically significant correlation in visit 28 (Rho = 0.74, p = 0.0139) and visit 365 (Rho = 0.9, p = 0.0374). Similarly, KT ratio correlated negatively with NAb titers in all time-points, with statistically significant correlation in visit 56 (Rho = -0.76, p = 0.0171).
In this prospective cohort of individuals receiving YFV, those with HIV had similar levels of YFV viremia and AEs as HIV-uninfected controls. Compared to controls, HIV-infected participants also had similar initial immunogenicity to YFV, measured by YF-specific NAb titers at 7, 14, 28, 56, and 84 days after vaccination, adjusted for age, sex and previous YFV. However, HIV status was independently associated with lower persistence of YF-specific NAb titers one year after vaccination. In the analysis of predictors of immunogenicity among HIV-infected participants, lower CD4/CD8 ratio, higher KT ratio and detectable HIV VL were associated with lower YF-specific NAb titers. There was no evidence for an association between viremia by Human Pegivirus, CD4 and CD8 T cell counts and YF-specific NAb titers among HIV-infected individuals.
Earlier studies of YFV immunogenicity including HIV-infected patients in the pre-ART period or in the initial phases of ART had demonstrated that CD4 T cell count and HIV VL predicted YF-specific NAb titers [7–12]. However, in the current era of early ART initiation, more HIV-infected patients are expected to have undetectable HIV VL and higher CD4 T cell counts. In our study, despite the elevated CD4 T cell count and high proportion of ART-suppressed individuals, HIV status was still associated with lower persistence of YF-specific NAb titers following an apparently adequate initial immunogenicity. Our results are consistent with prior publications suggesting that HIV-infected subjects still present lower responses to vaccines [1, 6, 16]. In addition, while the START and TEMPRANO trials demonstrated that earlier ART initiation dramatically reduces the risk of infectious outcomes, there was still a substantial risk of infectious outcomes in the immediate ART arms [36, 37]. Thus there are likely to be subtle immune defects that persist despite early ART initiation. Our study provides potentially important insights into mechanisms that might contribute to this persistent risk of infectious complications as well as point of care diagnostics that might identify patients at highest risk. For example, higher plasma KT ratio–a marker of IDO activity–strongly predicted lower YFV Nab titers after vaccination. While our observational study cannot assess causality, the fact that IDO-generated tryptophan catabolites suppress lymphocyte proliferation and function provides a plausible mechanistic pathway of its detrimental effect on vaccine responsiveness and, more broadly, adaptive immunity. Indeed, higher IDO activity has already been shown to predict increased mortality in several cohorts of ART-suppressed HIV-infected individuals [38–40]. While higher IDO activity might simply be a surrogate of other immunologic defect causally associated with impaired B cell function (e.g., the extent of T follicular helper cell infection and/or dysfunction in lymphoid tissues), a potential causal role for IDO activity in impairing vaccine responsiveness is plausible. Interestingly, high KT ratio was one of the strongest immunologic correlates of low CD4/CD8 ratio in another recent study of ART-suppressed individuals [21], suggesting that this biomarker–already obtained as part of routine clinical care—might help identify individuals with highest risk of impaired vaccine responses and adaptive immune defects.
Collectively, our findings further encourage the development of therapeutic interventions to reduce immune activation in ART-treated HIV-infected individuals [18, 20]. Early ART initiation is a well-recognized, yet insufficient strategy to normalize persistent immune activation [41], and additional strategies including inhibition of IDO activity are currently under study [42].
While effective interventions to inhibit immune activation are not available for clinical use, another important implication of our findings is the potential to substantiate recommendations for a booster dose of YFV for HIV-infected individuals at permanent or recurring risk of wild YF. In a recent publication, the Advisory Committee on Immunization Practices from Centers for Disease Control and Prevention published recommendations for YFV, suggesting HIV-infected individuals may benefit from a booster vaccination, which would not be recommended in routine circumstances due to the high immunogenicity and durability of YFV in the general population [43]. Our results suggest that a booster YFV may be beneficial even for HIV-infected individuals with high CD4 T cell counts. Since lower persistence of NAb was observed one year after YF vaccination, and AE following a booster dose of YFV are rare [28, 29], either periodic monitoring of YF-NAb or administration of a booster YFV dose could be recommended for HIV-infected individuals at permanent or recurring risk of wild YF as early as one year after primary vaccination. Additional studies are necessary to determine the durability of immunogenicity after a booster vaccination in this population.
Because all included participants received YFV as indicated due to potential risk of exposure to wild-type virus, we cannot rule out that natural exposure, rather than YFV alone, partially accounted for the observed NAb titers. However, most participants were residents in non-endemic areas and received YFV due to temporary visits to endemic regions with low risk of natural exposure. Furthermore, this potential competing immune stimulus would likely occur non-differentially regarding HIV status. Consequently, we do not believe our results are substantially compromised by natural exposure to wild YF virus.
Due to the small sample size, our results must be interpreted with caution. The model addressing predictors of YFV immunogenicity among HIV-infected participants included only 12 individuals followed longitudinally with 7 repeated outcome measurements. Although statistical methods for longitudinal analysis typically reduce error variance and improve power, this exploratory analysis needs confirmation in larger samples and different settings. In addition, our study was likely underpowered to provide definitive conclusions regarding the effect of HIV status on initial YFV response, on risk of AE (in particular rare severe AE) and on viremia by the YFV virus. Sensitivity for detection of AE was enhanced in our study by the measurement of solicited clinical AE, laboratory assessment of potential hematological or biochemical abnormalities, and systematic measurement of YFV viremia. Therefore, although not definitive, our findings provide important information on YFV clinical and laboratory adverse events, as well as vaccine virus kinetics among HIV-infected participants.
Our study may also be underpowered to detect significant effects of viremia by Human Pegivirus, CD4 and CD8 T cell counts on YF-specific NAb titers among HIV-infected participants. Our study included HIV-infected individuals with a very high range of CD4 T cell count, and we cannot rule out that CD4 T cell count would still predict YFV immunogenicity among patients with wider CD4 T cell count variability. Finally, we used a single measurement of RT-PCR to determine Human Pegivirus co-infection, and were unable to discriminate recent unresolved infections from chronic infections.
In conclusion, HIV-infected individuals have impaired NAb response to YFV due to a poorer persistence of antibodies, despite a seemingly normal initial response. Immune activation seems to reduce YFV immunogenicity, consistent with the observation that immune activation markers are useful predictors of clinical outcomes in the current era of HIV care [21, 22, 38]. A booster dose of YFV, although not recommended in routine circumstances, may be beneficial for HIV-infected individuals at permanent or recurring risk of wild YF.
|
10.1371/journal.pcbi.1002905 | Viral Capsid Proteins Are Segregated in Structural Fold Space | Viral capsid proteins assemble into large, symmetrical architectures that are not found in complexes formed by their cellular counterparts. Given the prevalence of the signature jelly-roll topology in viral capsid proteins, we are interested in whether these functionally unique capsid proteins are also structurally unique in terms of folds. To explore this question, we applied a structure-alignment based clustering of all protein chains in VIPERdb filtered at 40% sequence identity to identify distinct capsid folds, and compared the cluster medoids with a non-redundant subset of protein domains in the SCOP database, not including the viral capsid entries. This comparison, using Template Modeling (TM)-score, identified 2078 structural “relatives” of capsid proteins from the non-capsid set, covering altogether 210 folds following the definition in SCOP. The statistical significance of the 210 folds shared by two sets of the same sizes, estimated from 10,000 permutation tests, is less than 0.0001, which is an upper bound on the p-value. We thus conclude that viral capsid proteins are segregated in structural fold space. Our result provides novel insight on how structural folds of capsid proteins, as opposed to their surface chemistry, might be constrained during evolution by requirement of the assembled cage-like architecture. Also importantly, our work highlights a guiding principle for virus-based nanoplatform design in a wide range of biomedical applications and materials science.
| Viruses are increasingly viewed not as pathogens that parasitize all domains of life, but as useful nanoplatforms for synthetic maneuvers in a wide range of biomedical and materials science applications. One of the most well-known examples of virus-based nanotools developed so far features viral capsules as therapeutic agents, which protect and deliver drug molecules to targeted disease sites in the human body before the drug molecules are released. In order to optimize these nano-designs to best fulfill their purposes, we first have to understand properties of the constitutive building blocks of these viral containers, so as to rationalize and guide the synthetic modification attempts. Based on the observation that viral shells are functionally unique to viruses, we hypothesize that the structure of the building blocks must also be distinct from generic proteins, given that function follows form. Our computational modeling and statistical analysis support this novel hypothesis, and recognize the folded topology of these ‘Lego’ proteins as a differentiating factor to ensure correct geometry, and consequently, proper tiling into the large complex architecture. Our findings highlight an important design principle: efforts on imparting new functionalities to virus templates should restrain from disrupting the fundamental protein fold.
| Viral capsid proteins protect the viral genome by forming a closed protein shell around it. Most of currently found viral shells with known structure are spherical in shape and observe icosahedral symmetry [1]. Comprised of a large number of proteins, such large, symmetrical complexes assume a geometrically sophisticated architecture not seen in other biological assemblies. Here we make a distinction between protein cages in viral capsid shells that have sizes ranging from about 10 nm to about 90 nm in radius (Figure 1A), and other oligomeric containers of a much smaller scale, such as ferritins and chaperones. In the simplest form, 60 identical copies of an icosahedral asymmetric unit (IAU) are assembled with 5∶3∶2 symmetry, by positioning three IAUSs on each of the 20 triangular faces of the icosahedron [2]. The triangulation-number, or T-number, can be used to describe the number of proteins in each icosahedral asymmetric unit and therefore the size of the virus. Thus the number of capsid proteins in each shell is a multiple of 60, such as 180 proteins for a T = 3 virus and 240 proteins for a T = 4 virus. While T = 1 viruses can place each protein in an identical environment, other viruses having multiple proteins per IAU achieve the symmetry by following the ‘quasi-equivalence’ principle proposed by Caspar and Klug [3]. Also worth noting is that large viruses, such as double-stranded RNA (dsRNA) viruses, deviate from this principle, while preserving a rigid icosahedral symmetry nonetheless [2].
Geometry of the complex architecture aside, another striking feature of viral capsid proteins lies in the folded topology of the monomers, with the canonical jelly-roll β barrel appearing most prevalent (but not sole) as a core structural motif among capsid proteins that make up these viral shells of varying sizes [4]. Traditionally, this fold has also been termed as a wedge shape [5], an RNA virus capsid domain [6], a β-barrel [7], a β-sandwich [8], and an eight-stranded antiparallel β-barrel fold with a β-roll topology [9], all of which are consistent with the overall morphological characteristic of the fold (Figure 1B). Remarkable diversity in the loop regions connecting the β strands has been observed across different viruses, with variations in length and in inserted segments ranging from secondary structural elements to complete domains [10]. This signature fold of capsid proteins has been extensively studied [11], [12], and has also been compared with non-viral proteins in many separate works, most of which aimed to investigate the evolutionary relationship between viruses and their hosts. Other than the jelly-roll β barrel, there are also the Greek key β barrel with six strands [13], the helix bundle [14] and the immunoglobulin-like fold [15].
Given the unique geometry of the complex formed by viral capsid proteins, one interesting question arises as to whether the structural folds of capsid proteins that assemble into this distinct architecture are also unique to viruses. By comparing the structural topology of capsid proteins that form the icosahedral shells and generic proteins that interact to form other types of complexes, we can potentially establish a link between capsid fold and capsid architecture, or the lack thereof. The answer to this question can lend novel insights to protein-protein interactions, in terms of how folds of protein monomers, as opposed to their surface chemistry, might be related to the assembled multimer complex architecture. Furthermore, the ability of many viral capsid proteins to self-assemble spontaneously makes them an attractive platform for synthetic manipulation across the fields of biomedical applications and nanosciences [16]. Understanding how much influence viral capsid folds place on the assembled architecture is likely to provide guiding principles in the design of drug delivery systems and nanomaterials.
In this work, we present, to the best of our knowledge, the first attempt to examine whether the structural folds of viral capsid proteins set them apart from generic proteins, and with how much statistical significance. We recognize that a general assumption is that any class of proteins with a unique function is expected to be found in exclusive folds, which may or may not hold, given that folded topology is a coarse description of structural characteristics. Thus in addition to testing our hypothesis in the specific case of viral capsid proteins, we perform similar analysis for a few representative classes of proteins with diverse functions. At a finer level of granularity, i.e., the superfamily level, Abroi and Gough also surveyed the classification of all viral proteins and the other superkingdoms to study their genetic interactions in evolutionary history [17]. We distinguish our work by restricting our analysis to viral capsid proteins, which are functionally unique in viruses, in order to establish the link between topology of the building block and the assembled complex architecture. In another related work, Janin and coworkers provided an extensive analysis of physicochemical characteristics of protein-protein interfaces in icosahedral viruses, and compared them with generic protein-protein interfaces [18]. Rather than adopting the same approach of enumerating what's similar and what's different between the two classes, we will employ a direct comparison metric to evaluate whether there is significant statistical evidence supporting our conjecture that viral capsid proteins are structurally unique.
To test our hypothesis that viral capsid folds are not commonly found in generic proteins, we proceed to evaluate if the proportion of non-viral capsid proteins that share similar structural folds with viral capsid proteins is significantly small (Figure 2), based on a well-defined quantitative measure.
We chose the Template Modeling-score (TM-score) [19] as our structural comparison metric, for the following reasons. This structure-alignment-based scoring function using the fr-TM-align algorithm [20] is very fast to compute and suits our large-scale comparison; it is normalized, or protein size independent, making the comparison between pairs of domains with complex topology and pairs with simpler ones fair; it has been established in large scale benchmark studies that most of the pairs of proteins with a TM-score of more than 0.5 have the same fold classification, and most of those with a TM-score of less than 0.5 are in different fold classes [21]. In addition, a TM-score of 0.4 has also been extensively used as a criterion to decide if a pair of structures are similar or not [22]. Given that many proteins within the same SCOP fold can have a TM-score of 0.4 and higher, we chose the TM-score of 0.4 as the threshold to validate our hypothesis.
Briefly, the structural alignment score is defined aswhere LN and LT are the lengths of the two peptides being compared, di is the distance between the Cα atoms of the structurally equivalent residues, and d0 is a normalization score to make the alignment length-independent. The term Max stands for an optimal superimposition between the two structures to minimize distances between structurally-equivalent residues. We define structural distance between a pair of proteins by (1–TM-score), which ranges from zero to one.
In our work, we included all of capsid, nucleocapsid and envelope proteins for analysis, which we collectively call capsid proteins, because of their common structural role in forming the viral shell despite differentiated functions in a few cases. We collected the viral capsid protein set from the VIrus Particle ExploreR (VIPERdb) [23], which is a database of icosahedral virus capsid structures, with 319 entries in total. Altogether 1174 protein chains having at least 80 residues were extracted from these entries, as short peptides are known to assume very simple topologies. These 1174 were further cut into domains; while 452 proteins have domain annotations in SCOP, 637 proteins have homologues (sharing a sequence identity of at least 40%) that are well-annotated by SCOP. The remaining 85 were examined visually and dissected into individual domains. Lastly, the non-compact domains (extended structure with little secondary structure content) are removed, leaving 1447 domains in total.
We used the non-redundant set of 10569 proteins covering 1195 folds from the database Structural Classification Of Proteins (SCOP) 1.75 [24] filtered at 40% sequence identity, available from the ASTRAL compendium [25], to constitute our total protein set. This set was further reduced to 8921 proteins covering 1047 folds after removal of short peptides with fewer than 80 residues. The viral capsid protein set was then subtracted from the total protein set to yield the non-capsid protein set. In addition, 24 capsid proteins in the total protein set that were originally not deposited in VIPERdb were added to the capsid set and removed from the non-capsid set (Table S1). A sequence filter of 40% identity was then applied to the domains of the capsid set, which resulted in 151 domains that are sequence-wise non-redundant.
As viruses across the same family are known to share limited sequence identity despite remarkable structural resemblance, a further structural filter was applied to the capsid set of 151 domains by clustering analysis. We performed hierarchical clustering via the average linkage method, and selected the cluster medoids of the resulting N clusters as our structurally non-redundant capsid set. Optimal partitioning of the data from hierarchical clustering was obtained by choosing the minimal number of clusters such that all intra-cluster distances are less than 0.6, using our structural distance measure. This criterion is based on the rationale that we would like to sort out the most representative capsid structures, without their repeating one another resulting in unfair comparison with the permutation test that we will describe shortly.
A preliminary survey of the two sets revealed differences in the sizes of domains. As shown in Figure 3, a typical capsid domain (blue) has approximately 180 residues, compared to about 150 residues for a typical non-capsid domain (pink). This size comparison is purely based on existing structural data of viral capsid proteins, but we do see a larger proportion of complex topologies in certain capsid domains, as opposed to the under-representation of longer folds in generic proteins. In order to preclude the possibility of concluding that capsid and non-capsid proteins have different folds that are in fact largely a result of the difference in length, we performed an additional separate analysis by removing domains having longer than 600 residues in both datasets.
After obtaining the non-redundant viral capsid set and the non-capsid set, we quantify the extent to which the structural space of the non-capsid set overlaps with that of the capsid set in the following manner. We performed an all-against-all structural comparison between the non-capsid set and the capsid set. For each member in the non-capsid set, we select its nearest neighbor in the capsid set, and use the distance between the two to represent how far structurally this particular non-capsid protein is to viral capsid proteins. With the structural distances between all non-capsid proteins and their nearest neighbors in capsids in hand, we then filter the non-capsid set by retaining only proteins that are less than 0.6 away from capsid proteins. We thus obtain the final distribution of distances between viral capsid proteins and those non-capsid proteins that structurally resemble capsid proteins. Among these ‘relatives’ of viral capsid proteins, we count the number of folds covered by them, following the fold classification in SCOP. This defines our test statistic, which we term as ‘shared folds’ in the rest of the paper.
To estimate the statistical significance of the number of shared folds between capsids and non-capsid proteins, we calculated the probability of observing at most the same number of shared folds by random chances by running a permutation test on the total protein set. The total set of proteins was randomly partitioned into set A and set B, with set A consisting of an equal number of proteins as that in the capsid set, and set B being their complement in the total set. The same procedure as described above was carried out to obtain the number of shared folds between this particular set A and their non-self counterparts. To avoid finding ‘relatives’ in set B that are evolutionarily closely related to (i.e. belonging to the same family) the proteins in set A, we further excluded ‘self folds’ from the shared folds found, as an approximation to, or a lower bound of, folds shared with non-self proteins. Here ‘self fold’ is defined as the fold annotation by SCOP of a particular structural analogue found in the large protein set that is already covered by any protein in the small set of proteins. Altogether 10,000 independent permutations were done to give rise to the estimated distribution of shared folds, based on which the p-value of our test statistic can be evaluated.
To examine if unique function generally implies unique folds, we chose a few functional classes of proteins to perform the same analysis described above for capsid proteins. Seven classes were chosen, namely kinases, globins, dehydrogenases, DNA/RNA polymerases, chaperones, antigens and muscle proteins, with functions ranging from catalysis, to transport to signal transduction. The total protein set which is filtered at 40% sequence identity level was partitioned into two sets based on SCOP annotations at the domain level; one being the functional class and the other being the complementary set, and the statistical significance of shared folds is again estimated by permutation tests.
We found 56 clusters for the viral capsid set, using the criterion described in the Materials and Methods section. These clusters are fairly compact, with all members within each cluster being less than 0.6 apart from one another. Furthermore, the clusters are maximally separated, with only 26 pairs of proteins (0.24%) from two different clusters being closer than 0.4. In Figure 4, we show the statistics demonstrating a good separation between clusters that are reasonably homogeneous. The resulting 56 cluster medoids thus represent the distinct domain architecture adopted by capsid proteins.
Figure 5 illustrates these 56 clusters with all members in each cluster superimposed on one another. The alignment shows high structural similarity across the same cluster, while different clusters display mostly different folding topologies, in agreement with our quantitative assessment. There are a fairly large number of singlet clusters that are unlike one another, mostly because the structural data for these few viral families are lacking. The few most populated clusters correspond to the canonical jelly-roll fold, with variations in the terminal ends.
By comparing the viral capsid set and the non-capsid set, we found altogether 2078 generic proteins sharing similar topology with viral capsid proteins, based on a distance cutoff of 0.6. These 2078 proteins cover 210 folds in total. If we disregard marginally similar capsid-like proteins by looking at those within a distance 0.5 of capsid proteins only, we find altogether 600 proteins covering 21 folds (Table 1). A further inspection of the distribution of shared folds for randomly sampled sets of 56 proteins and their non-self counterparts immediately reveals that viral capsid proteins are structurally separated from generic proteins. Referring to Figure 6, the cumulative fraction of non-self proteins across the entire structural distance spectrum from viral capsid proteins is clearly shifted to the right compared to those of the 10,000 permutation tests. Through this plot, we expect to arrive at the answer that capsid proteins are different from generic proteins regardless of the distance cutoff used in defining similar folds.
The distribution of shared folds, estimated from the 10,000 permutation tests, is plotted in Figure 7. The number of capsid-like folds shared by non-capsid proteins hence lies on the extreme left tail of the distribution, demonstrating that viral capsid folds are far less populated in structural fold space compared to generic proteins (Figure 7). The one-tailed p-value of our test statistic is less than 0.0001, and we thus conclude that there is significant statistical evidence against the null hypothesis that viral capsid folds span the protein fold space. We also show in Figure S1 (supporting information) that the p-value of our test statistic, based on the datasets containing domains of comparable sizes only, is 0.0002, therefore excluding size as a compounding factor contributing to the difference in fold. In conclusion, viral capsid folds are unique to viruses.
The seven other functional classes of proteins we examined range in size from 18 to 297 in the total set of 8921 proteins. When compared with their complementary set, the number of shared folds with non-self proteins is found to be statistically insignificant, with a one-tailed p-value greater than 0.05 in all cases (Table 2). This is not surprising, given that cellular proteins have evolved over a relatively shorter period of time, and therefore their folds are more similar to one another as compared to viruses, similar being defined by having a TM-score of greater than 0.4. We thus showed that it is not always true that unique function implies unique structural folds. Without making this assumption, we further proved that viral capsid proteins are segregated in structural fold space, which is remarkable.
In this work, our major interest is to compare the independently folded domains of capsid proteins with generic protein domains, so as to reveal their relationship with the higher order of structural organization. Domains defined in this work therefore refer to integral structural units that are connected by single peptide to neighboring domains, although in a few cases these criteria are not fully met. We followed strictly the definition of domains in SCOP to make fair comparison with generic proteins collected from the same database. Our work does not focus on a finer granularity of structure such as subdomains, or motifs, which might have been called ‘domains’ in certain literature for the interpretation of their evolutionary origin. While our choice of domain definition addresses our question of interest adequately, we also note that the question of whether viral folds and generic proteins are evolutionarily segregated can be answered by comparing subdomains or structural motifs, which is outside the scope of discussion here.
Prior to our work, several studies have reported that certain classes of cellular proteins also share similar topologies or structural cores with certain capsid proteins. These include the tumor necrosis factor superfamily [8], the serine proteases [13], the superantigen class [26], the concavalin A class [11], and the CUB-like domains [27]. All of the above classes of proteins were among the generic proteins that we found to share similar folds as capsid proteins, as expected. In addition, analysis of our set of 600 non-viral relatives of capsid proteins revealed that many virus proteases, certain hydrolases, transcription regulators and histone chaperones also shared close topological characteristics with viral capsid proteins (Table 1).
We first examined the structural relatives that are highly similar to capsid proteins (within a distance 0.4 or less). Many of these structural relatives possess the typical jelly-roll topology, with some variations in each case. The tumor necrosis factor superfamily is characterized by 10 strands in two sheets, with the core eight strands having identical connectivity as that of a standard capsid jelly-roll. Truncation in one strand and addition of two extra strands make them slightly different in shape compared to capsid proteins. The CUB-like domains in spermadhesins display a particular variation of the jelly-roll topology in terms of connectivity, including reversed β-strands, two disulphide bridges and two additional β-strands. They thus share a minimal structural core with capsid proteins (specifically the bean pod mottle virus capsid protein), but have shorter β-strands and overall smaller shape as a distinction. Superantigen Ypm is yet another class that overlaps significantly in structure with capsid proteins, especially satellite tobacco necrosis virus capsid proteins. Other than an additional disulphide bond connecting the C terminus with one β-strand that differentiates itself, superantigen Ypm also has a much more compact structure compared to capsid proteins, owing to its shorter loops connecting the β-strands. The supernatant protein factor protein consists of two domains, and the C-terminal domain also follows the jelly-roll topology that resembles satellite tobacco necrosis virus most, with minute differences in the concavity of the two β-sheets. The histone chaperone proteins are characterized by the same topology as capsid proteins, with some of them having one or two additional strands. Remarkably, all of these proteins discussed occur naturally (as opposed to crystal packing) as heterodimers (the monomers having identical topology), trimers, pentamers or hexamers, although their mode of interaction differ from that of capsid proteins in many cases. This suggests that the β-sandwich formed by proteins with varying connectivity generally facilitates aggregation, presumably because of the greasy, flat surfaces presented by their wedge-like shapes to promote monomer association.
In addition to these structural analogues found naturally in oligomeric states, we also identified quite a few proteins in the immunoglobulin fold and the methyltransferases fold that are highly similar to capsid proteins; however, they typically occur as part of some multi-domain proteins, such as the N-terminal binding fragment of the human polymeric immunoglobulin receptor. It thus might not be feasible to simultaneously arrange all domains on a shell in such cases, which may explain why we are not observing multimeric complexes for these proteins. We omit here discussion on the remaining types of protein domains, mainly for the reason of their limited structural similarity to capsid proteins (distance-wise more than 0.4 apart). These proteins typically either appear smaller in size or are tightly coupled with other domains, and consequently significantly different in shape, and have not been observed to form symmetric complexes in general.
Given the above interesting observations, we need to highlight that the structural relatives of capsid proteins only marginally resemble capsid proteins to the extent of their common structural core, as evident from the large structural distances (majority are greater than 0.4) between the two classes. Decorations on top of this level of similarity directly differentiate the exposed edges of the proteins, such that geometrical complementarities along multiple symmetry axes are easily satisfied by repeating units of the same monomers in the case of capsid proteins but not in the other. In other words, the positions in which monomers interact with one another are also fine-tuned by geometric and physicochemical factors of protein-protein interfaces. We thus do not observe any protein cages assembled from these cellular proteins despite their sharing similar structural topologies with capsid proteins. Lastly, we speculate that the structural but not functional close relationship between these few classes of proteins and capsid proteins resulted from ancient genetic interactions between viruses and their hosts, although further investigation is needed to support this view.
An important aspect that cannot be overlooked is that we have drawn our conclusion in this work based solely on existing structural data of capsid proteins taken from icosahedral viruses. We cannot exclude possibilities of identifying novel viral capsid folds that span a larger subspace of protein folds in future, as predicted in several recent publications [16], [17], [28] given the diversity of the virosphere. This is especially so when we take into account the current challenges in determining the structure of viral proteins embedded in lipid membranes for enveloped viruses. In addition, experimental limitations in determining the structure of large assemblies place a heavy bias in highly symmetrical viral particles, and thus statistics for irregularly shaped viruses such as HIV are missing in our analysis. Given all structural data available up to this date, we have derived our conclusion with rigor and confidence, but we remain open to potential changes should abundant novel discoveries be made.
Our study provided support for the hypothesis that viral capsid proteins, which are functionally unique in viruses in constructing protein shells, are also structurally unique in terms of their folding topology. This implies that protein-protein interactions, in the case of viral capsids at least, confer evolutionary constraints on capsid proteins, specifically on their folds. Bhadur and Janin [29] found that residues making up capsid cores are more conserved than interface residues and surface residues, which highlights a greater selective pressure on capsid structural core. Interpreted together, the characteristic folds (and therefore fundamental shapes) of capsid proteins are most likely a consequence of geometric requirements of the building block so as to form the cage-like macromolecular assembly, which corroborates the theory proposed by Mannige and Brooks [30]. From a more general point of view, core residues of cellular proteins have also been known to be evolving at a slower rate compared to interface and surface residues [31], with a 25%–35% higher conservation score compared to surface residues. Most studies that investigated the degree to which proteins are subject to constraints due to their interactions with other proteins mainly focused on interface residues [31]–[33], and it remains to be established whether the greater conservation of structural cores of generic proteins is similarly affected by the interaction with their partners during evolution. Our work sheds light on this missing link by studying the particular case of viral capsid proteins, and it will be interesting to verify whether this evolutionary constraint is true in general.
Additionally, virus-like particles (VLPs), which are self-assembling capsid shells without the infectious viral genetic materials encapsulated, are already a popular choice among a variety of nanoparticle platforms for wide applications both in the biomedical arena and in material science [34]–[40]. For a comprehensive review, readers may refer to this paper [41]. Compared to other nanoparticle materials, VLPs offer several advantages, including the full range of protein templates they provide that adapt to diverse environmental conditions including extreme thermal environments [42], their proteinaceous nature which makes them biodegradable [43], and their plasticity to a wide range of synthetic manipulations [44]–[46]. For biomedical applications, VLP design has been formulated for targeted delivery of drug molecules [47], tissue-specific imaging reagents [45], as well as novel vaccine development [48]. VLPs have also been extensively explored as nanocontainers [49] and nanotubes [50] in materials science. In order to fulfill their desired purposes, VLPs are introduced new functional modules, to facilitate specific interactions with the intended biological sites or nonbiological surfaces, to alter the overall architecture and stability [51], and to package various cargos as well as directing the cage assembly [52]. Our work laid out the fundamental principle in such tailored design of VLP platforms; in order to preserve the assembled architecture of viral capsid shells, it is important for the newly formulated protein subunits to adhere to the library of viral capsid folds. In other words, significant adaptations that result in unfolding or misfolding of capsid proteins are undesirable. Where human creativity has no bound in exploring all synthetic possibilities, feasibility has its bound; decorations on VLPs should minimally disrupt the folded topology and geometry of the building block to make it work.
Having established that viral capsid proteins possess distinct folds, we would like to take one step further by examining whether the protein-protein interfaces in viral capsid assemblies are also unique to viruses. Because differences in monomer structure do not imply differences in protein-protein interfaces [22], our conclusion of the uniqueness of capsid fold cannot be directly extended to capsid interfaces. The results of this second comparison will again have interesting implications. Should capsid interfaces resemble those of generic ones, the mode of capsid-capsid interation is then governed purely by physicochemical laws, and evolution merely plays a part in dictating the building block structure for their proper tiling. If, on the other hand, we learn that viral capsid interfaces are quantitatively different from interfaces formed by their cellular counterparts, we can then tap on this difference and design pathogen-specific antiviral drugs targeted at disintegrating the protection shells, without disrupting normal cellular activities. Work in this direction is in progress.
In summary, our comprehensive analysis of the viral capsid proteins and their cellular counterparts revealed the segregation of capsid proteins in structural fold space. This provides important clues to requirements of the building blocks for the distinctive viral shell architecture; the unique folds of viral capsid proteins present favorable geometry to allow effective packing and assembly into the right complex architecture. With this in mind, the design of gene therapy delivery agents as well as nanoparticles, both targeted at making packing tools, can be tailored to satisfy geometric constraints by following closely the viral capsid templates nature has created for us.
|
10.1371/journal.ppat.1003643 | Inherited Prion Disease A117V Is Not Simply a Proteinopathy but Produces Prions Transmissible to Transgenic Mice Expressing Homologous Prion Protein | Prions are infectious agents causing fatal neurodegenerative diseases of humans and animals. In humans, these have sporadic, acquired and inherited aetiologies. The inherited prion diseases are caused by one of over 30 coding mutations in the human prion protein (PrP) gene (PRNP) and many of these generate infectious prions as evidenced by their experimental transmissibility by inoculation to laboratory animals. However, some, and in particular an extensively studied type of Gerstmann-Sträussler-Scheinker syndrome (GSS) caused by a PRNP A117V mutation, are thought not to generate infectious prions and instead constitute prion proteinopathies with a quite distinct pathogenetic mechanism. Multiple attempts to transmit A117V GSS have been unsuccessful and typical protease-resistant PrP (PrPSc), pathognomonic of prion disease, is not detected in brain. Pathogenesis is instead attributed to production of an aberrant topological form of PrP, C-terminal transmembrane PrP (CtmPrP). Barriers to transmission of prion strains from one species to another appear to relate to structural compatibility of PrP in host and inoculum and we have therefore produced transgenic mice expressing human 117V PrP. We found that brain tissue from GSS A117V patients did transmit disease to these mice and both the neuropathological features of prion disease and presence of PrPSc was demonstrated in the brains of recipient transgenic mice. This PrPSc rapidly degraded during laboratory analysis, suggesting that the difficulty in its detection in patients with GSS A117V could relate to post-mortem proteolysis. We conclude that GSS A117V is indeed a prion disease although the relative contributions of CtmPrP and prion propagation in neurodegeneration and their pathogenetic interaction remains to be established.
| Prions are infectious agents causing incurable brain disease in humans and animals. Prion diseases are by definition transmissible, which means that it should be possible to experimentally transfer disease from patient brain tissue to laboratory animals by inoculation. While many forms of prion disease have been shown to be experimentally transmissible, some inherited forms, in particular, Gerstmann-Sträussler-Scheinker syndrome (GSS) associated with the substitution of valine for alanine at amino acid position 117 (GSS A117V) of the human prion protein gene have not. This has led to the suggestion that such syndromes are not true prion diseases and are better designated non-transmissible proteinopathies. Since prions may transmit more efficiently when the host's normal prion protein amino acid sequence matches that of the infecting prion, we generated transgenic mice expressing human prion protein with the same amino acid sequence found in A117V GSS. We found that brain tissue from GSS A117V patients could transmit disease to these mice, producing the typical brain lesions associated with GSS A117V. We therefore conclude that GSS A117V is an authentic prion disease.
| According to the widely accepted “protein-only” hypothesis [1], an abnormal isoform (PrPSc) of host-encoded cellular prion protein (PrPC) is the principal, and possibly the sole, constituent of the transmissible agent or prion [2]. Prions exist in multiple strains which are thought to represent distinct polymeric forms of misfolded PrP which faithfully propagate by recruitment of host PrPC onto pre-existing seeds or fibrils (for review see [3]). Human prion diseases may occur sporadically, be acquired by infection with environmental prions, or be inherited as autosomal dominant conditions as a result of one of more than 30 different coding mutations in the human PrP gene (PRNP) [4]. The cause of neuronal dysfunction and death in prion disease is unclear but neurotoxicity may be uncoupled from infectivity suggesting that prions themselves may not be directly neurotoxic and other PrP species might be involved in mediating toxicity [3], [5], [6].
By definition, prion diseases are transmissible, and while all the sporadic and acquired human prion diseases have been transmitted to laboratory animals, not all of the inherited forms have. It has been suggested therefore that some of these inherited neurodegenerative syndromes are prion proteinopathies with a distinct pathogenesis that may not involve production of infectious prions. One inherited prion disease (IPD) in particular, associated with an alanine to valine substitution at residue 117 of PrP (A117V), has been proposed to cause neurodegeneration in the absence of PrPSc, with pathogenesis mediated by aberrant production of CtmPrP, a transmembrane form of PrP [7]. It has also been proposed that PrPSc accumulation in other forms of prion disease may cause pathology by inducing the synthesis of CtmPrP de novo [8]. This aberrant topologic form of PrP has been hypothesised to cause neurologic dysfunction by disrupting the function of mahogunin, a cytosolic ubiquitin ligase whose loss causes spongiform neurodegeneration [9]. However, a recent study using mice lacking the Mahogunin Ring Finger 1 (MGRN1) E3 ubiquitin ligase concluded that disruption of MGRN1-dependent pathways does not play a significant role in the pathogenesis of prion diseases [10].
A117V is one of the IPD mutations associated phenotypically with the Gerstmann-Sträussler-Scheinker syndrome (GSS) which usually presents clinically as a progressive cerebellar ataxia with dementia occurring later in a clinical course usually far more protracted than that of Creutzfeldt-Jakob disease (CJD) [11]. Pathologically, GSS is characterised by the presence of multicentric PrP amyloid plaques. However, in common with other IPD's, A117V has a wide phenotypic diversity at both the clinical and pathological level even within the same kindred [11]. This disease was originally misdiagnosed as Alzheimer's disease [12] before the advent of PrP immunohistochemistry [13] and the subsequent identification of the A117V mutation by PRNP sequencing [14].
A major determinant of phenotypic heterogeneity in prion diseases of humans and animals is prion strain diversity, with distinct prion strains producing characteristic clinical and pathological phenotypes [15]. Prion strains can be distinguished by biochemical differences in PrPSc, referred to as molecular strain typing [16]. In a number of inherited prion diseases, distinct PrPSc types have been reported associated with the same PRNP pathogenic mutation and this may in part explain phenotypic heterogeneity (for review see [17], [18]). In this regard while classical CJD is typically characterised by proteinase K-resistant PrP fragments of ∼21–30 kDa on immunoblots [19] most GSS cases show additional low molecular mass fragments of 7–15 kDa [19]–[27]. Notably, the major protease-resistant peptide extracted from brains of GSS A117V patients is a ∼7–8 kDa PrP fragment [26], and to-date it has not been possible to detect proteolytic fragments of molecular mass 21–30 kDa in these samples. The pathogenic role of the PrP species from which the 8 kDa fragment is generated is not clear because, inocula containing this fragment induced conversion of murine 101L-PrP into amyloid but did not induce spongiform neurodegeneration in the recipient mouse brains [28]. These facts coupled with the negative experimental transmission data have led to the suggestion that GSS A117V may not be an authentic prion disease and would be more accurately described as a non-transmissible proteinopathy [29], [30].
That most sporadic and acquired CJD occurred in individuals homozygous at PRNP polymorphic codon 129 supported the view that prion propagation proceeded most efficiently when the interacting PrPSc and PrPC were of identical primary structure [31], [32]. It has been demonstrated that the species barrier may be abrogated in transgenic mice expressing PrP homologous to that of the exogenous PrPSc [33]. This is also the case with transmission of human prion diseases. Classical CJD transmits rarely if at all to wild type mice but highly efficiently (indeed without a transmission barrier) to mice expressing human (and not mouse) PrP [34], [35]. However, prion strain type may also play a key role in transmission barriers, which are thought to be mediated via conformational selection where a given PrP primary structure has a preferred subset of disease-associated conformations it can adopt [3], [36]. While therefore it is possible that some naturally occurring human prion strains could transmit more efficiently for example to wild type mice rather than to mice transgenic for a particular human PrP polymorph (as is the case for vCJD for example [37]), it is logical to test for transmissibility of GSS A117V using transgenic mice expressing only human PrP 117V.
Here we present the first evidence that IPD A117V cases produce transmissible prions; previous transmission attempts may have failed from use of inappropriate experimental models. Furthermore, we show that the previous failure to detect PrPSc in GSS A117V patient brain may have been due to its unusual instability with consequent loss by post-mortem proteolysis in human brain samples.
We produced transgenic mice homozygous for both a human PrP 117V, 129V transgene array and murine PrP null [38] alleles (Prnpo/o), designated Tg(HuPrP117V,129V+/+ Prnpo/o)-31 (hereafter referred to as 117VV Tg31), with human PrP expression levels three times that of pooled normal human brain (data not shown). We studied an ageing cohort of 20 mice for evidence of spontaneous neurodegeneration, however all of these uninoculated mice died of intercurrent illnesses or old age between 460 and 904 days without developing neurological disease. In addition, three out of a further control group of five mice mock-inoculated with PBS buffer lived to between 344 and 735 days post-inoculation without developing neurological signs. One mouse was scored clinically sick at 303 days post inoculation (Table 1) but this and two other PBS-inoculated mice had no evidence for pathological PrP in brain by either immunoblotting or immunohistochemistry (IHC). One mouse, culled at 582 days post-inoculation due to intercurrent illness, although negative for PrPSc by immunoblotting was found to have minor PrP immunoreactivity in the anterior commissure by IHC (Figure S1A). This finding in a single sample was not studied further.
To assess the susceptibility of these novel transgenic lines to prion infection, we first inoculated them with well characterised isolates of classical CJD with proven transmissibility to mice expressing wild-type human PrP [35], [37], [39], [40] although recognising that the presence of the A117V mutation may introduce a transmission barrier to prions generated from wild-type PrP. Sporadic CJD isolate I022 (PRNP 129VV with type 2 PrPSc) caused clinical disease in 2/6 mice with relatively short incubation periods of 263 and 303 days post-inoculation (Table 1). Although clinical attack rate was low, most mice (5/6) were subclinically infected and showed positivity for abnormal PrP by IHC and/or immunoblotting (Table 1). In contrast, other inocula comprising sporadic CJD and iatrogenic CJD with different PRNP codon 129 status and PrPSc types, and also vCJD and vCJD passaged in 129VV Tg152 mice (which contained type 5 PrPSc [37], [40]) transmitted very poorly or not at all (Table 1). Collectively these findings show that at 3-fold expression level of human PrP, 117VV Tg31 mice can replicate human prions, although this varies with prion strain and codon 129 genotype effects.
117VV Tg31 mice were inoculated with three isolates of GSS A117V and remarkably all resulted in transmission with clinically affected mice (Table 1). To our knowledge this is the first time these isolates have been shown to have transmissible prions. However, it should be noted that clinical transmissions were associated with extremely long incubation periods, ranging from 609 to 673 days post-inoculation (Table 1). It is therefore unsurprising that previous attempts to transmit this disease, into animals expressing endogenous levels of a PrP of different primary structure, were completely unsuccessful [41], [42].
We investigated all the clinically unaffected mice challenged with brain homogenates from GSS A117V patients or classical CJD prions for evidence of subclinical infection [5], [39], [43] by both PrP immunohistochemistry and immunoblotting. We found that 6/8 of the mice inoculated with GSS A117V prion isolate I514 were positive by immunohistochemistry (Figure 1 A–D), although only 2/9 showed clinical signs (Table 1). In contrast to sporadic CJD inoculum I022 which produced only synaptic type PrP deposits (Figure 1E and G), all three IPD A117V inocula resulted in intense deposition of PrP plaques in cerebral cortex, hippocampus, thalamus and cerebellum (Figure 1A and C). There was neuronal loss (Figure S1B and Figure 2) and spongiosis, more pronounced in the white matter (Figure 1B), and gliosis (Figure 1D and Figure 2) that reflected the extent of the PrP plaque load. Sub-clinical infection was also prominent in 117VV Tg31 mice challenged with sporadic CJD inoculum I022, with 5/6 mice being positive by immunohistochemical analysis (Table 1, Figure 1 E–H) despite a low clinical attack rate of 2/6.
The classical proteolytic PrPSc fragments of ∼21–30 kDa have to-date not been detected in brain from A117V patients. We analysed the brains of all clinically affected mice and those that died of inter-current illnesses by immunoblotting for the presence of PrPSc. We first confirmed that 117VV Tg31 mice are capable of producing stable PrPSc by analysing brains of mice inoculated with sporadic CJD prions. In order to adequately digest PrPC in these mice, we used stringent PK digestion conditions of 100 µg/ml incubated at 37°C for 1 hour, and demonstrated the presence of PrPSc in the brains of A117V Tg31 mice inoculated with sporadic CJD inoculum.
Immunoblots show clear evidence that 117V PrPC is convertible to PrPSc in 117VV Tg31 mice challenged with sporadic CJD inoculum I022 (Figure 3A, lanes 3, 4 and 7) and is present at similar levels in control mice expressing wild type human PrP-129MV challenged with the same prion inoculum (Figure 3A, lane 1). Transmission of iatrogenic CJD prion isolate I1477 to 117VV Tg31 mice shows a low intensity positive signal associated with the brain of a subclinically infected mouse culled 804 days post-inoculation (Figure 3B lane 4), which when compared with the absence of signal in a second mouse culled relatively early at 294 days post-infection (lane 3) probably reflects the relative abundance of 117V PrPSc accumulated over the respective survival periods.
Having established that 117V PrPC would support the propagation of conventional PrPSc in our transgenic mice, we analysed brains of 117VV Tg31 mice that were inoculated with GSS A117V prions for the presence of disease-related PrP. Encouraged by the confirmation that these 117VV mice can replicate human prions, and to adequately digest PrPC in these mice which express at higher levels, we used relatively harsh PK conditions (100 µg/ml PK at 37°C for 1 hour) and found that five brain samples analysed showed variable PK resistance (Figure 3C, lanes 3–7). Brain samples appear to have achieved only partial digestion even under these conditions, and displayed concurrently the presence of PrPC, PrPSc 21–30 kDa fragments and extra fragments of about 7–8 kDa (Figure 3C lanes 4–6). The presence of multiple PK digestion products seen on immunoblots was not due to inadequate PK digestion parameters because under the same conditions an inoculated 117VV Tg31 mouse that was killed due to intercurrent illness at 188 days post-inoculation (Figure 3C, lane 3), and a PBS-inoculated control mouse killed due to intercurrent illness at 569 days post-inoculation (Figure 3C, lane 8) showed only residual PrPC signal that was only visible after long exposure.
Notably, one brain sample (Figure 3C, lane 7) achieved complete digestion with 100 µg/ml PK at 37°C for 1 hour, and clearly shows the presence of PrPSc at a level comparable to the positive control sample in lane 1. Interestingly, the 8 kDa PrP fragment was not detected in this sample.
As this was the first demonstration of detectable classical PrPSc (generating PK-resistant fragments equivalent to PrP27–30 [44]) associated with GSS A117V, we sought to reproduce the immunoblotting results but we were surprised to find that after freeze-thawing of the brain homogenates, we were unable to demonstrate PrPSc under the same harsh proteinase K (PK) conditions of 100 µg/ml at 37°C for 1 hour (Figure S2 A and B). Figure 3D shows the same sample in Figure 3C lane 7 that on repeat western blotting and exposure for the same length of time showed only a weak PrP27–30 signal at much reduced PK concentration of 10 µg/ml digested at 37°C for 1 hour. Repeat immunoblotting of the three other 117VV Tg31 brain homogenates shown in Figure 3C lanes 4–6, even at drastically reduced PK concentrations were negative for disease-associated PrP bands and showed only residual non-digested bands corresponding to PrPC (data not shown). Of note, immunoblotting of the same samples in the absence of PK digestion showed that PrPC remained relatively stable in these samples (data not shown). These results strongly suggest that 117V PrPSc is significantly more labile than that seen in CJD and other human prion diseases. The remarkable difference in migration patterns between classical CJD-challenged (Figure 3A lanes 3, 4, and7) and those of IPD A117V-challenged Tg31 mice (Figure 3C lanes 4–5) is a further reflection of the unique properties of A117V prions that set them apart from those of classical CJD prions.
To corroborate these novel findings, we also inoculated a second transgenic line expressing HuPrP 117V PrPC, called Tg(HuPrP117V,129V+/+ Prnpo/o)-30 (designated 117VV Tg30), with the same three IPD A117V prion isolates in addition to one case of classical CJD and the same case of vCJD (Table 2). The 117VV Tg30 mice were produced similarly to 117VV Tg31 mice but have a level of human PrP expression two-fold higher than a pooled normal human brain standard (data not shown), as compared with the 3-fold PrP overexpression in the 117VV Tg31 line. Consistent with the low rate of clinical disease in 117VV Tg31 mice, the 117VV Tg30 mice did not show a single case of clinical disease from any of the inocula administered (Table 2). However, as seen with prion-inoculated 117VV Tg31 mice, evidence of sub-clinical prion infection as measured by positive immunohistochemistry was seen in the majority of inoculated mice (Table 2 and Figure 1, I–L). Additionally, immunohistochemical analysis of the brains of 117VV Tg30 mice inoculated with GSS A117V prion isolate I514 all showed pathological lesions characterised by gliosis (Figure 1L) and spongiosis (Figure 1J) that reflected the level of PrP plaques (Figure 1K) deposited in a similar pattern to 117VV Tg31 mice described above. Spongiosis was more pronounced in white matter and neuronal loss was prominent (Figure S1B and Figure 2). Two other A117V prion inocula (I1321 and I1322) produced neuropathologically similar patterns to that of I514, though the plaque load was slightly less (Figure 2). In all GSS A117V prion-infected 117VV Tg30 mice only the 8 kDa PrP fragment was detected (Figure 3E lanes 7 and 8).
Interestingly, and in contrast to the Tg31 mice with higher levels of expression of the mutant protein, we observed spontaneous clinical disease in three mice at between 476 and 742 days in an ageing cohort of 20 uninoculated mice. This was associated with PrP plaque deposition in the anterior commissure (data not shown). We are currently investigating whether this pathology is transmissible on sub-passage.
We also investigated the pattern of neuropathology produced in vCJD-inoculated 117VV Tg30 mice. One of 3 positive samples showed abundant plaques in the cerebral cortex, hippocampus, thalamus and cerebellum (Figure 1M and O). Although neuronal loss was present, there were no florid plaques, consistent with the propagation of vCJD in the PRNP 129VV genotype [37], [40]. However, while only a few non-florid plaques are typically seen with vCJD transmission to the wild-type human PrP 129VV genotype [37], [40], the abundance of non-florid plaques associated with vCJD transmission to 117V mice is remarkable and clearly suggests a modifying effect of the mutation.
Transmission of vCJD prions to transgenic mice homozygous for human PrP valine-129 invariably results in a strain shift from the characteristic type 4 PrPSc molecular signature to type 5 PrPSc [37], [40]. The presence of type 5 PrPSc fragment size in vCJD-inoculated 117VV Tg30 mouse brain (Figure 4A lane 4) compared to type 4 PrPSc propagated in the vCJD-inoculated 129MM Tg45 control brain (lane 1) clearly shows that the 117V mutation on the valine-129 allele does not influence the previously established strain shift phenomenon. Interestingly, a truncated PrP peptide of about 8 kDa that is associated with GSS A117V mutation was also seen on longer exposure in PK-digested (Figure 4A lanes 3 and 4) and PK-titrated samples (Figure 4B lanes 1 to 4). Notably, the 8 kDa fragment was not seen without PK digestion (lane 5 Figure 4A and 4B respectively), thus confirming this PrP fragment as disease specific. Indeed, in some vCJD-challenged 117V Tg30 mouse brains that were positive by immunohistochemistry, only the GSS-associated 8 kDa PrP fragment was detectable (Table 2).
We have demonstrated that GSS A117V is indeed a transmissible condition and properly designated an inherited prion disease rather than simply a prion proteinopathy without generation of prions. Additionally, we report that classical PrPSc is detectable in PrP 117V transgenic mouse brain using suitable conditions. The inability to detect classical PrPSc in patient brain had led to the proposal that the A117V mutation may cause pathology principally via an alternative pathway, namely through an increase in C-terminal transmembrane PrP, designated CtmPrP, to the total exclusion of PrPSc [8]. It has also not been shown whether or not 117V-PrPC is convertible to PrPSc. Using appropriate transgenic models challenged with classical CJD prion isolates, we have demonstrated that, despite the observed transmission barrier to clinical disease which can be explained by the 117V mutation producing a partial transmission barrier, 117V PrPC is a competent substrate for conversion to PrPSc. Notably, the newly generated PrPSc assumes the stable strain properties of the exogenous PrPSc and is therefore readily detectable on immunoblots.
Similarly, although transmission properties of GSS A117V prions in these mice were not typical of prion transmission to transgenic mice expressing the homotypic substrate, our detection of classical PrPSc is unprecedented and confirms that experimental conditions in our 117VV transgenic mice were favourable for replication of PrPSc. However, in contrast to the stable PrPSc propagated from classical CJD prion transmission to these mice, the observation that PrPSc generated from GSS A117V prions in vivo was inherently unstable may in part explain the low clinical attack rates observed in the present study and the failure of previous transmission attempts. It is reasonable to infer that because the A117V-derived abnormal PrP is labile, prion replication and the probability of a sustained prion infection in these mice would have been greatly enhanced by the 2–3 fold over-expression of the substrate, 117V PrPC.
Given that the only protease-resistant PrP fragment found to-date in A117V patients' brains is the characteristic 8 kDa PrP fragment [26], our 117VV Tg30 line in which only 8 kDa PrP fragment was detectable has recapitulated the GSS A117V disease phenotype. Since the 8 kDa peptide was only seen as a proteinase-K resistant truncated fragment, it represents a GSS-specific PrP degradation product, the detection of which can be taken as a reliable surrogate marker for confirming prion disease in GSS A117V patients [26]. The possibility of classical PrPSc being present at low and undetectable levels in GSS A117V patient brain homogenates cannot be ruled out. It therefore remains to be determined whether the parent PrP conformer that generates the 8 kDa protease resistant PrP, is capable of initiating and sustaining prion infection or that transmissibility remains associated with classical PrPSc present below the threshold of detection. In this regard, even a successful serial passage of GSS A117V-challenged Tg30 mouse brains apparently propagating only the 8 kDa fragment and resulting in the propagation of classical PrPSc, may not resolve this issue.
All previous reports of PrP point mutations causing spontaneous neurodegeneration have involved superimposing human PrP pathogenic mutations onto rodent PrP [7], [45], [46], and these studies have invariably reported very high incidences of spontaneous neurological dysfunction. Since destabilising effects measured in a mouse protein cannot be assumed to be equivalent in the human protein [47], [48], we have modelled the A117V mutation directly on human PrP. This difference in approach can explain the contrasting low incidence of spontaneous disease in our 117VV transgenic mice. The development of neurological dysfunction in transgenic mice expressing disease-associated mutations modelled on rodent PrP has been described as disease acceleration [49], because PrPSc has not been detectable and transmissibility has not been demonstrated conclusively. In this regard, transmissibility of spontaneous PrP plaque deposits in aged 117VV Tg30 mice is being investigated and will be reported in a subsequent publication.
The observation that vCJD prions transmit more readily, albeit subclinically, to 117VV Tg30 but not to 117VV Tg31 mice that have higher PrP expression levels was unexpected. However, whereas all vCJD-inoculated 117VV Tg31 mice had a maximum post-inoculation survival period of 547 days (culled in the range 292–547 days), 6/7 117VV Tg30 mice challenged with the same inoculum survived in the range of 627–811 days post-inoculation. These data suggest that very prolonged replication periods may be required for pathological PrP to become detectable in vCJD-challenged 117V transgenic mice by either IHC or immunoblotting. Subpassage of apparently negative brains could be used to explore this, however this is not a central part of this study.
Our results may have wider implications for other inherited prion diseases that have not been shown to be transmissible as yet. Firstly, it is possible that demonstration of transmissibility of such inherited prion diseases would require specific transgenic models with over-expression of the relevant mutant human PrP, rather than endogenous levels of mutant PrP expression, if transmissibility is to be demonstrated within the lifespan of a mouse. The transient detection of PrPSc in our study suggests that A117V-associated PrPSc is labile and readily susceptible to proteases. This results in progressive reduction of PrPSc to undetectable, yet still potentially infectious levels. In this regard, failure to detect low levels of PrPSc in the past from patient brain samples could be due to technical limitations of currently available biochemical techniques, rather than its absence.
Storage and biochemical analysis of human tissue samples and transmission studies to mice were performed with written informed consent from patients or relatives under approval from the Local Research Ethics Committee of UCL Institute of Neurology/National Hospital for Neurology and Neurosurgery and the code of practice specified in the Human Tissue Authority licence held by UCL Institute of Neurology. Work with mice was performed under licence granted by the UK Home Office (Animals (Scientific Procedures) Act 1986 ; Project Licence number 70/6454) and conformed to University College London institutional and ARRIVE guidelines.
The 759 bp human PrP ORF was amplified by PCR with pfu polymerase from genomic DNA prepared from the brain of a patient with the inherited prion disease A117V mutation, using forward primer 5′-GTCGACCAGTCATTATGGCGAACCTT-3′ and reverse primer 5′-CTCGAGAAGACCTTCCTCATCCCACT-3′. Restriction sites Sal I and XhoI (underlined) were introduced in the forward and reverse primers respectively for cloning. The sequence was confirmed and ligated into the cosmid vector CosSHaTet [33]. Microinjection of the purified DNA was carried out according to standard protocol into single cell eggs of Prnp null mice [38] which had been backcrossed onto an FVB/N background. Genotyping was performed by PCR and PrP expression levels estimated by Western blot analysis as previously reported [50]. Two homozygous lines were established for HuPrP 117V described as Tg(HuPrP117V,129V+/+ Prnpo/o)-30 (designated human PrP 117VV Tg30) and Tg(HuPrP117V,129V+/+ Prnpo/o)-31 (human PrP 117VV Tg31) with mutant transgene expression levels of 2 and 3 times respectively, compared to pooled normal human brain levels.
Strict bio-safety protocols were followed. Inocula were prepared, using disposable equipment for each inoculum, in a microbiological containment level 3 laboratory and inoculations performed within a class 1 microbiological safety cabinet. Ten mice per group of 117VV Tg31 transgenic mice were inoculated with prion isolates comprising human brain homogenates from: three separate IPD A117V cases; two sporadic CJD cases; three iatrogenic CJD cases; one case of vCJD and one mouse brain isolate from vCJD passaged once in Tg152 mice expressing wild-type human PrP V129 (containing type 5 PrPSc) [37], [40], as detailed in Table 1. Similarly, the second 117VV transgenic line, 117VV Tg30 mice were challenged with the same three IPD A117V inocula, and 1 inoculum each of iatrogenic CJD and vCJD as detailed in Table 2. All cases were neuropathologically confirmed.
The genotype of each mouse was confirmed by PCR of tail DNA prior to inclusion and all mice were uniquely identified by sub-cutaneous transponders. Disposable cages were used and all cage lids and water bottles were also uniquely identified by transponder and remained with each cage of mice throughout the incubation period. Mice were anaesthetised with a mixture of halothane and O2, and intracerebrally inoculated into the right parietal lobe with 30 µl of a 1% brain homogenate prepared in phosphate-buffered saline (PBS). All mice were thereafter examined daily for clinical signs of prion disease. Mice were killed if they exhibited any signs of distress or once a diagnosis of prion disease was established.
Mice were culled by CO2 asphyxiation. Brain was fixed in 10% buffered formol saline and then immersed in 98% formic acid for 1 hour and paraffin wax embedded. Serial sections of 4 µm thickness were pre-treated by boiling for 10 min in a low ionic strength buffer (2.1 mM Tris, 1.3 mM EDTA, 1.1 mM sodium citrate, pH 7.8) before exposure to 98% formic acid for 5 min. Abnormal PrP accumulation was examined using anti-PrP monoclonal antibody ICSM 35 (D-Gen Ltd, London) on a Ventana automated immunohistochemical staining machine (Ventana Medical Systems Inc., Tucson, Arizona) using proprietary secondary detection reagents (Ventana Medical Systems Inc) before development with 3′3 diaminobenzedine tetrachloride as the chromogen. Harris haematoxylin and eosin staining was done by conventional methods. Appropriate controls were used throughout.
Preparation of brain homogenates (10% w/v in phosphate buffered saline), proteinase K digestion (titration up to 100 µg/ml for 1 h at 37°C), and subsequent western blotting was performed as described previously [51]. For primary screening of both transgenic and wild type mouse brain homogenates, blots were probed with either a monoclonal antibody which detects human, but not mouse, PrP (3F4 ([52])) or a biotinylated anti-PrP monoclonal antibody which recognises both human and mouse PrP (biotinylated-ICSM 35 (D-Gen Limited, London)) in conjunction with an avidin-biotin-alkaline phosphatase conjugate (Dako) and development in chemiluminescent substrate (CDP-Star; Tropix Inc). Primary screening of brain homogenates was performed blind to sample identity.
|
10.1371/journal.pcbi.1000305 | Timing the Emergence of Resistance to Anti-HIV Drugs with Large Genetic Barriers | New antiretroviral drugs that offer large genetic barriers to resistance, such as the recently approved inhibitors of HIV-1 protease, tipranavir and darunavir, present promising weapons to avert the failure of current therapies for HIV infection. Optimal treatment strategies with the new drugs, however, are yet to be established. A key limitation is the poor understanding of the process by which HIV surmounts large genetic barriers to resistance. Extant models of HIV dynamics are predicated on the predominance of deterministic forces underlying the emergence of resistant genomes. In contrast, stochastic forces may dominate, especially when the genetic barrier is large, and delay the emergence of resistant genomes. We develop a mathematical model of HIV dynamics under the influence of an antiretroviral drug to predict the waiting time for the emergence of genomes that carry the requisite mutations to overcome the genetic barrier of the drug. We apply our model to describe the development of resistance to tipranavir in in vitro serial passage experiments. Model predictions of the times of emergence of different mutant genomes with increasing resistance to tipranavir are in quantitative agreement with experiments, indicating that our model captures the dynamics of the development of resistance to antiretroviral drugs accurately. Further, model predictions provide insights into the influence of underlying evolutionary processes such as recombination on the development of resistance, and suggest guidelines for drug design: drugs that offer large genetic barriers to resistance with resistance sites tightly localized on the viral genome and exhibiting positive epistatic interactions maximally inhibit the emergence of resistant genomes.
| The ability of HIV to rapidly acquire mutations responsible for resistance to administered drugs underlies the failure of current antiretroviral therapies for HIV infection. The recent advent of drugs that offer large genetic barriers to resistance, e.g., tipranavir and darunavir, presents a new opportunity to devise therapies that remain efficacious over extended durations. The large number of mutations that HIV must accumulate for resistance to drugs with large genetic barriers impedes the failure of therapy. Further, these drugs appear to exhibit activity against viral strains resistant to other drugs in the same drug class, thereby significantly improving options for therapy. Rational identification of treatment protocols that maximize the impact of these new drugs requires a quantitative understanding of the process whereby HIV overcomes large genetic barriers to resistance. We develop a model that describes HIV dynamics under the influence of a drug that offers a large genetic barrier to resistance and predict the time of emergence of viral strains that overcome the large barrier. Model predictions provide insights into the roles of various evolutionary forces underlying the development of resistance, quantitatively describe the development of resistance to tipranavir in vitro, and suggest guidelines for drug design.
| Current antiretroviral therapies for HIV infection often fail to elicit lasting virological responses in patients because of the emergence of multidrug resistant strains of HIV [1],[2]. The enormous replication rate and the high mutation and recombination rates of HIV [3]–[7] propel the acquisition of mutations that confer upon HIV resistance to administered drugs. The same mutations are often responsible for resistance to multiple drugs belonging to a given drug class [1],[2]. Consequently, treatment options for patients who experience failure of therapy are restricted [8],[9]. The newly approved protease inhibitors (PIs), tipranavir and darunavir, offer large genetic barriers to resistance [10],[11]. The genetic barrier of a drug, n, is the number of mutations that HIV must accumulate to gain high level resistance to the drug [12]. When n is small (e.g., n = 1 for 3TC [1]), drug resistant genomes are likely to exist in patients prior to the onset of therapy [13]. As n increases, the likelihood of the pre-existence of resistant genomes decreases considerably [13],[14]. Resistant genomes must then emerge during therapy through mutation and/or recombination of susceptible genomes. The replication of susceptible genomes, however, is suppressed during therapy. Besides, HIV must undergo a large number of replication cycles to accumulate all the mutations required for resistance to a drug with large n. Consequently, the development of resistance to a drug with large n may be significantly delayed. Indeed, up to 9 months were required for HIV to develop resistance to tipranavir in in vitro serial passage experiments [10].
Current treatment guidelines for HIV infection recommend a combination of 3, but at least 2, active drugs, (i.e., drugs for which resistance has not developed) in order partly to increase the overall genetic barrier of therapy [9]. For treatment naïve patients, a combination of 2 nucleoside/nucleotide reverse transcriptase inhibitors (NRTIs) is typically employed in combination with either a non-nucleoside reverse transcriptase inhibitor (NNRTI), usually efavirenz, or a ritonavir-boosted PI, usually lopinavir [9]. With ritonavir-boosted lopinavir monotherapy, fewer patients achieved plasma HIV RNA levels below detection and more patients witnessed emergence of PI resistance mutations than in patients receiving ritonavir-boosted lopinavir in combination with 2 NRTIs [15]. Similarly, despite comparable times to virological failure, patients receiving a 2 drug combination of efavirenz and lopinavir experienced more frequent emergence of resistance than patients receiving a 3 drug combination of efavirenz or lopinavir and 2 NRTIs [16]. Therapy with 4 NRTIs had a similar response to therapy with efavirenz and 2 NRTIs [17]. Consequently, a 3 drug combination is the current standard of care for treatment naïve patients. When failure did occur with a 3 drug combination, it was typically associated with NNRTI resistance in patients receiving efavirenz but not with PI resistance in patients receiving lopinavir [16], in accordance with the larger genetic barriers offered by PIs than by NNRTIs [18]. The large genetic barrier in conjunction with a superior pharmacokinetic profile may also underlie the high rates of viral suppression despite sub-optimal adherence in patients receiving ritonavir-boosted lopinavir-based therapy [19].
For second-line therapy, which follows the failure of the initial regimen, a drug from a new drug class is recommended in order to minimize the risk of cross-resistance [9]. Thus, among several newly available agents [20], the fusion inhibitor enfuvirtide and the recently approved integrase inhibitor raltegravir present potent options. Both enfuvirtide and raltegravir, however, offer small genetic barriers and are therefore recommended for use in conjunction with a supporting drug such as darunavir [9]. Remarkably, the new PIs, tipranavir and darunavir, elicit responses against viral strains resistant to other PIs [11],[21], increasing options for second-line therapy. The new PIs thus present promising weapons to avert the failure of antiretroviral therapy. Indeed, significant efforts are ongoing to identify treatment protocols that maximize the impact of the new PIs [8],[22]. Identification of improved protocols hinges on our understanding of HIV dynamics under the influence of drugs that offer large genetic barriers to resistance and of the process by which HIV surmounts these large genetic barriers.
Description of the development of resistance to a drug with a large n is complicated for several reasons. First, resistance to such a drug typically develops gradually, increasing progressively with the number of mutations accumulated [10],[23]. As a result, the emergence and the competitive dynamics of a large number of distinct viral genomes carrying different combinations of resistance mutations and possessing various intermediate levels of resistance must be described. For instance, the accumulation of mutations at 6 loci confers high level resistance to tipranavir [10]. Consequently, depending on whether each resistance locus carries a mutation or not, 26, or 64, distinct strains (see below) may emerge in the course of infection. Because HIV is diploid, the 64 strains yield 64 homozygous and 2016 different kinds of heterozygous virions, whose evolutionary dynamics must be followed to describe how the genetic barrier of tipranavir is overcome. Second, the population size of HIV in vivo may be small, especially under the influence of therapy, which implies that the emergence of resistant genomes is likely to be governed by stochastic rather than deterministic effects [24]. Third, in addition to mutation, recombination can play a significant role in the formation of drug resistant strains that carry multiple mutations [25],[26]. The influence of recombination, which is yet to be fully understood, depends on several factors, viz., the frequency of multiple infections of cells, the effective population size of HIV in vivo, and the nature of fitness interactions between resistance mutations, characterized by epistasis [27]–[33]. No models exist that describe HIV dynamics under the simultaneous influence of mutation, multiple infections of cells, recombination, epistatic interactions between multiple resistance mutations, and stochastic effects of finite population sizes. Consequently, timing the failure of antiretroviral drugs with large genetic barriers is currently not possible. Rational identification of improved treatment protocols is therefore precluded.
Here, we develop a model of HIV dynamics that quantitatively predicts the expected waiting time for the emergence of genomes that carry the requisite mutations for resistance to a drug with any given genetic barrier. Extant models of HIV dynamics assume that deterministic forces are predominant in the emergence of drug resistance [34]–[36]. Consequently, extant models predict that drug resistant genomes emerge immediately upon the initiation of therapy, albeit in small numbers. In contrast, especially when the genetic barrier is large, stochastic forces are expected to dictate the emergence of resistant genomes. A key consequence of the predominance of stochastic forces is a delay in the emergence of resistant genomes following the initiation of therapy. Our model accounts for this delay in a deterministic manner by predicting the expected waiting time for the emergence of resistant genomes. Model predictions capture the development of resistance to tipranavir in vitro quantitatively, indicating that our model captures the underlying dynamics of the development of resistance to antiretroviral drugs. Further, model predictions provide insights into the impact of underlying evolutionary forces on the development of drug resistance and suggest guidelines for drug design.
We consider uninfected cells, T, exposed in the presence of a PI with a genetic barrier n to a viral population, V, containing genomes highly susceptible to the PI. The highly susceptible, or wild-type, genomes are assumed to contain no resistance mutations. As infection proceeds, error-prone replication gives rise to mutant genomes. distinct mutant genomes can arise, each with at least one resistance mutation (Figure 1). Our aim is to determine the waiting time for the first formation of the genome that carries all the n resistance mutations and is therefore highly resistant to the drug. We number the different viral genomes 0, 1, 2, 3…S, where genome 0 represents the wild-type (Figure 1). We let Vjh denote the population of virions containing genomes j and h, where j, h ∈ {0, 1, 2…S}. Because virions V10, for instance, are indistinguishable from virions V01, we impose the constraint j≤h [37]. Following the infection of a cell by a virion Vjh, mutation and recombination give rise to a proviral genome i ∈ {0, 1, 2…S} with probability Qi(jh). We distinguish infected cells by the proviral genomes they contain: Cells Ti are infected by a single provirus i and cells Tij by proviruses i and j, where i≤j and i, j ∈ {0, 1, 2…S}. Infected cells produce progeny virions. Drug action causes some of the progeny virions to be non-infectious [3],[34],[35]; we denote the noninfectious virion population by . Cells Ti and Tii infected by a single kind of provirus produce homozygous virions Vii and . Cells Tij infected with distinct proviruses (i≠j) yield homozygous virions Vii, , Vjj and and heterozygous virions Vij and . The resulting infection network is shown in part in Figure 2.
We construct dynamical equations to predict the time-evolution of various cell and viral populations and estimate the average waiting times for the first production of each of the S mutant proviral genomes (Methods). We denote by W the waiting time for the emergence of the provirus that contains all the n resistance mutations and hence overcomes the genetic barrier of the drug.
We solve model equations to describe the development of resistance in in vitro serial passage experiments (e.g., [10]). Here, T0 uninfected cells are exposed to viruses in the presence of a known concentration of the PI. Infection is allowed to progress until time tp (∼3.5 days), the duration of a passage. The resulting viral population is employed to initiate infection of a fresh set of T0 uninfected cells in the next passage. At the start of the first passage, the viral population is assumed to consist of V00 wild type viruses, highly susceptible to the drug. Gradually, genomes with increasing levels of drug resistance emerge.
We apply our model to describe the development of resistance to tipranavir in in vitro serial passage experiments [10]. We let n = 6 because a genome with 6 resistance mutations exhibited >10 fold resistance to tipranavir in these experiments. We choose IC50 values for different intermediate mutants from the ranges determined experimentally (Table S1). Further, we employ actual distances between resistance sites to calculate the recombination probabilities and also assign fitness advantages to genomes containing specific combinations of mutations (Table S1). (In contrast, in our calculations above, the number of mutations and not their specific combinations was assumed to determine the fitness advantage.) We also vary the concentration of tipranavir as in the experiments (Table S2). Further, following the experimental protocol, we employ 90% of the viral population at the end of any passage to initiate infection in the succeeding passage when the drug concentration is maintained constant across the passages and 50% of the viral population when the drug concentration is increased in the succeeding passage. Genomes carrying 2, 3, 5 and 6 resistance mutations were first observed in the experiment in passages 16, 33, 39 and 49, respectively [10]. In close agreement, our model predicts the emergence of these genomes in passages 14, 29, 44 and 49, respectively (Figure 6). (Ignoring the concept of the waiting time, i.e., letting wi = 0 in our model, severely underpredicts the times of emergence of drug resistant genomes (Figure 6). The agreement between model predictions and experiments indicates that our model captures the underlying dynamics of the development of resistance to antiretroviral drugs accurately.
Current models of HIV dynamics successfully predict short-term changes in the plasma viral load in patients undergoing therapy but fail to provide a quantitative description of the emergence of drug resistance [34]–[36]. A key limitation of current models is the underlying assumption that the emergence of resistant genomes is governed by deterministic effects. Deterministic effects predominate when the population of cells in an infected individual is large. In a finite cell population, because the probability of the formation of a resistant genome with many mutations can be small, resistant genomes emerge stochastically. The waiting time for the emergence of resistant genomes can therefore be substantial. In contrast, by assuming that deterministic effects predominate, current models predict that resistant genomes emerge, albeit in very small numbers, immediately upon the onset of therapy. Once resistant genomes emerge, their numbers grow due to viral production from the cells they infect leading to the rapid fixation of resistance. Current models thus underestimate the time for the development of drug resistance (Figure 6).
Simulations of viral evolution, based on models of population genetics, consider finite populations and present descriptions of the stochastic emergence of drug resistant genomes [28],[29],[32]. Importantly, the simulations also enable incorporation of recombination and fitness interactions between multiple loci, which are central to the development of drug resistance but are not easily incorporated in models of HIV dynamics. The simulations, however, make several simplifying assumptions, such as fixed population sizes and discrete generations, which approximate the dynamics of the development of drug resistance and introduce uncertainties in the influence of underlying processes, such as recombination [30],[33]. Besides, simulations are difficult to incorporate in mathematical formalisms for therapy optimization.
Here, we develop a model that employs the deterministic framework of models of HIV dynamics and at the same time captures the influence of stochastic effects associated with the emergence of drug resistant genomes. To accomplish this, we invoke the concept of the expected waiting time. We develop a detailed description of mutation and recombination between multiple loci, which enables calculation of the probability of the formation of resistant genomes in one replication event. Given the viral and cell populations and the efficacy of the drug, the frequency of replication events and hence the rate of formation of resistant genomes is determined. From the rate of formation, we estimate the expected waiting time for the first resistant genome to emerge. Different mutant genomes are assumed to appear first in the viral population at their respective expected waiting times. The limitation of current models of HIV dynamics, which predict the emergence of resistant genomes immediately upon the start of therapy, is thus overcome. Yet, by calculating the “expected” waiting time, our model captures the influence of stochastic effects associated with the emergence of resistant genomes in an averaged sense and retains the dynamical framework of current models. The limitations of population genetics based simulations are also thus overcome.
The waiting time for the emergence of a genome carrying a certain number of mutations depends on the times of emergence and the growth of subpopulations of genomes with fewer mutations. Our model assumes that the latter genomes emerge at their expected waiting times. Consequently, the variation in the waiting times for the emergence of higher mutants due to the variation in the times of emergence of lower mutants is suppressed in our model. Further, following emergence, particularly when the population size is small, the chance that stochastic forces cause the extinction of genomes may be significant. We assume, however, that following emergence, the growth of genomes is deterministic. The extent of the uncertainties introduced in our model predictions by these simplifying assumptions remains to be estimated. Semi-stochastic simulations, where the times of emergence of mutant genomes alone are determined stochastically, and fully stochastic simulations (see, e.g., [42]) of the emergence of mutant genomes would serve as tests of our model. Performing the simulations, however, is beyond the scope of the present study. Here, we compare model predictions with experiments and find that our predictions are in close agreement with experimental observations [10] of the times of emergence of various genomes possessing different degrees of resistance to tipranavir, suggesting that our model captures the underlying dynamics of the development of drug resistance by HIV.
Model predictions indicate that the waiting time, W, for the emergence of the strain that overcomes the genetic barrier of a drug depends on several factors that may be tuned during drug design. A large genetic barrier significantly enhances W. This enhancement of W with the genetic barrier is amplified when fitness interactions between resistance loci exhibit positive epistasis. Recombination, in contrast, lowers W regardless of epistasis or the genetic barrier. If the separation between resistance loci is small, however, the role of recombination is suppressed. Thus, for delaying the emergence of resistant genomes, drugs that offer large genetic barriers with resistance sites localized tightly on the viral genome and exhibiting positive epistatic interactions are desirable. These observations may serve as guidelines for structure-based drug design [43]. The fixation of resistant genomes following their emergence may depend differently on drug characteristics and remains to be fully elucidated.
When distinctions between different viral genomes are ignored, the expected waiting time vanishes and our model reduces to the basic model of HIV dynamics, which successfully captures viral load changes in patients undergoing therapy [3],[34]. Our model may thus be applied to predict drug failure in vivo. Several advances of our model are essential, however, to describe the in vivo scenario accurately. First, the higher frequency of multiple infections [44], possible cell-cell transmission of infection [45],[46], and the existence of resistance mutations prior to the onset of therapy [1] in vivo must be incorporated into our model. Second, during potent drug therapy, viral replication may be suppressed significantly, resulting in a small effective population size of HIV. The variation of the waiting time about the mean may then become large. Consequently, the assumption that mutant genomes emerge at their expected waiting times becomes less accurate. Our model must therefore be advanced to account for the variation of the emergence times of genomes in vivo. Third, our model must be extended to drugs from other drug classes to mimic current combination therapies. With these advances, our model would enable timing the emergence of resistance to drugs in vivo and facilitate the identification of treatment protocols that maximally impede the failure of current therapies.
We present equations below that describe the in vitro dynamics of various cell and viral populations.
We employ the following parameter values based on earlier studies [4],[37],[39],[52]: the birth and death rate of uninfected T cells, λ = 0.624 day−1 and dT = 0.018 day−1; the death rate of infected cells, δ = 1.44 day−1; the viral burst size, = 103; the viral clearance rate, c = 0.35 day−1; the second order rate constants of the infection of uninfected and singly infected cells, k0 = 10−8 day−1 and k1 = 0.7k0; the mutation and recombination rates, μ = 3×10−5 per site per replication, and ρ = 8.3×10−4 crossovers per site per replication.
|
10.1371/journal.pgen.1006335 | Family Based Whole Exome Sequencing Reveals the Multifaceted Role of Notch Signaling in Congenital Heart Disease | Left-ventricular outflow tract obstructions (LVOTO) encompass a wide spectrum of phenotypically heterogeneous heart malformations which frequently cluster in families. We performed family based whole-exome and targeted re-sequencing on 182 individuals from 51 families with multiple affected members. Central to our approach is the family unit which serves as a reference to identify causal genotype-phenotype correlations. Screening a multitude of 10 overlapping phenotypes revealed disease associated and co-segregating variants in 12 families. These rare or novel protein altering mutations cluster predominantly in genes (NOTCH1, ARHGAP31, MAML1, SMARCA4, JARID2, JAG1) along the Notch signaling cascade. This is in line with a significant enrichment (Wilcoxon, p< 0.05) of variants with a higher pathogenicity in the Notch signaling pathway in patients compared to controls. The significant enrichment of novel protein truncating and missense mutations in NOTCH1 highlights the allelic and phenotypic heterogeneity in our pediatric cohort. We identified novel co-segregating pathogenic mutations in NOTCH1 associated with left and right-sided cardiac malformations in three independent families with a total of 15 affected individuals. In summary, our results suggest that a small but highly pathogenic fraction of family specific mutations along the Notch cascade are a common cause of LVOTO.
| Left-ventricular outflow tract obstructions comprise a group of developmental heart disorders that are genetically and phenotypically heterogeneous, with no single gene accounting for the majority of cases. In order to identify mutations contributing to disease, we selected patients from 51 families with a history of congenital cardiac malformations. We interrogated the entire coding sequences of 106 patients and identified a small but highly pathogenic fraction of mutations that are likely to contribute to disease in 12 families (23.5%). Furthermore, we present a strategy for identifying candidate mutations based on familial segregation in a genetically heterogeneous disorder.
| Left-ventricular outflow tract obstructions (LVOTO) comprise a group of cardiac malformations that restrict blood flow in the left portion of the heart. This heterogeneous subclass of cardiac malformations is commonly associated with aortic valve disease and early onset aortopathy, including severe stenosis and aortic dilation. These complications often result in a high disease burden later in life manifested by thoracic aortic aneurysm and valve replacement. Despite recurrent clustering of LVOTO traits among family members and a strong hereditary component [1,2], the underlying genetic cause of disease remains largely enigmatic [3]. Recent whole-exome sequencing and genotyping approaches revealed single nucleotide variants (SNVs) and copy number variations (CNVs) in genes that play an important role during the course of early outflow tract development and in chromatin modifications [4,5]. The list of known CHD disease genes is rapidly expanding and can be broadly defined by three functional groups: transcriptional regulators, cardiac structural proteins and signaling components. Disrupting mutations in critical factors regulating cardiac gene expression, such as NKX2-5, TBX5 and members of the GATA gene family (GATA4-GATA6) have been identified in family based studies for congenital heart disease (CHD) [6–9]. In addition, rare missense mutations in structural proteins of the cardiac muscle, including FLNA and MYH6 have been associated with single gene disorders and isolated cases of CHD [10]. However, the major fraction of patients with CHD cannot be solely explained by mutations in known disease genes [3]. Recent studies suggest that multiple genes in conserved signaling pathways mediating important cues for crucial processes during early embryonic development contribute to disease. Both familial and sporadic occurrences of LVOTO have been associated with mutations in genes of the Notch-signaling cascade [11,12]. Multifaceted Notch signaling plays a crucial role in cardiac cell fate regulation and orchestrates the morphogenesis of cardiac chambers and valves [13–15]. A recent study has suggested that NOTCH1 haploinsufficiency alters specific gene networks affecting valve development and osteogenic factors which in turn result in aortic valve disease [16]. This is consistent with the identification of the NOTCH1 ligands JAG1 and DLL4 which have been shown to cause Alagille syndrome and aortic valve disease in human and animal models [17–19]. However, the high functional redundancy of the Notch signaling cascade and the multitude of overlapping phenotypes between interacting genes suggest a certain level of locus heterogeneity associated with disease. Here, we report our results from a family based approach that aims to decipher the underlying genetic heterogeneity of LVOTO traits in a pediatric cohort including 182 patients from 51 families of French Canadian origin. Central to our approach is the family unit which serves as a reference to identify causal genotype-phenotype correlation.
Whole-exome sequencing was performed in 182 individuals from 51 families with a strong heritable history of congenital heart disease (CHD) on the SOLiD 5000xl and Illumina HiSeq2000 platform (Fig 1a). The average read depth for the targeted platforms was 82x with 78% of targeted regions covered at greater than 20x. The ethnicity of our cohort was evaluated using principal component analysis and revealed that all except 1 patient, who was removed from subsequent analysis, clustered closely with individuals of European descent (Fig 1b) [20]. We evaluated the overall burden of co-segregating disease mutations for protein coding genes in our families instead of performing family-based association analysis due to the modest sample size and phenotypic heterogeneity of our cohort [21]. Among all 182 samples, 72,332 potential deleterious coding SNVs were identified of which 1245 were loss-of-function SNVs, including splice site disrupting variants, stop-gain mutations, frameshift indels and 71087 non-synonymous SNVs (Fig 2). In order to identify candidate deleterious variants, we applied four prediction algorithms for variant pathogenicity (SIFT, Poly-Phen-2, Mutation Taster, Mutation Assessor) and the CADD method [22]. All non-synonymous variants predicted to be damaging by at least two algorithms were included for further downstream analysis. We then filtered for rare variants, based on minor allele frequencies from the Exome Aggregation consortium (ExAC) ExAC Total MAF < 0.1%) and for variants which co-segregated with disease within families. This family-based approach that includes disease-segregation among multiple affected as filtering criteria, significantly reduced the number of potential pathogenic variants per patient in our family cohort (S1 Fig). As a next step, we prioritized candidate genes among the remaining 321 loss of function and 13,501 non-synonymous variants under the assumption that an excess of rare deleterious variants (rdSNVs) in genes intolerant to deleterious mutations would be more likely to affect gene function and contribute to disease in our cohort.
To reduce the noise of our enrichment analysis, we further selected genes based on the Residual Variation Intolerance Score (RVIS) statistic [23]. Fig 3a highlights the enrichment of co-segregating rdSNVs in genes towards the lower and higher end of the genome-wide genic intolerance distribution based on their RVI scores. This unbiased enrichment approach for rare variants segregating with disease revealed 14 candidate genes showing an enrichment of rdSNVs (z-scores > 2) in our cohort among the top 10% most intolerant genes genome-wide (Fig 3a, S1 Table). Among these are two promising candidate genes, NOTCH1 and KMT2D harboring rdSNVS co-segregating with disease in multiple families. NOTCH1 is associated with aortic valve disease [11,12] whereas variants in KMT2D cause the pediatric disorder Kabuki syndrome with multiple congenital malformations including aortic valve disease [24–26]. Overall, we observed a significant excess (p< 0.01, Wilcoxon) of co-segregating variants among highly conserved genes (RVIS < 10th percentile) in our data set (S2 Fig). We used the Gene Set Enrichment Analysis (GSEA) to investigate the potential biological relevance of a higher genetic burden in conserved genes for 50 hallmark gene sets from the Molecular Signatures Database [27–29]. Hallmark gene sets represent well-defined biological states that display coherent expression and were considered significantly enriched with a FDR multiple test correction p value of < 0.05. We found significant enrichment in the two hallmark gene sets, “myogenesis” and “epithelial-mesenchymal transition” (Fig 3b). The highly conserved genes related to “myogenesis” and “epithelial-mesenchymal transition” are likely to play a role in disease etiology due to their crucial role during cardiac development [30,31]. Contrary to the enrichment of co-segregating rdSNVs in genes intolerant to deleterious variants, an excess of rdSNVs was observed in genes with a high genome-wide tolerance of common variants (RVIS > 90th percentile) (S2 Fig). This is to be expected, since genes on this extreme end of the RVIS distribution include frequently mutated loci such as MUC6 and MUC16 (Fig 3a) that accumulated an excess of low frequency variants due to their length and exon number which makes them unlikely contributors to CHD [32–34]. Finally, mutational burden testing was performed on highly conserved candidate genes (RVIS < 10th percentile) in order to determine the contribution of individual candidate genes to disease in our cohort, based on collapsing co-segregating rdSNVs into a single gene score. We compared 106 affected individuals from our family cohort with 193 non-consanguineous population controls. The control samples sequenced at the Baylor-Hopkins Center for Mendelian Genomics were selected from the Mendel cohort without a history of cardiovascular disease and cluster predominantly with our samples of French Canadian origin (Fig 2b) [35]. While we did not observe a significant enrichment of rdSNVs in a single candidate gene, testing for novel mutations revealed a significant burden in NOTCH1 (novel alleles 0/193, novel alleles in LVOTO cohort 3/106, p = 0.0462, Fisher exact test one tailed). This is in line with an earlier association of NOTCH1 with bicuspid aortic valve disease highlighting a significant excess of private mutations in the locus [36]. After gene prioritization, we retained 53 candidate genes based on the following criteria: harboring co-segregating variants observed in at least one family, pathway association with the hallmark pathways “myogenesis”, “epithelial-mesenchymal transition” and “chromatin modification” for downstream analysis (S3 Fig)
In the next step of our family-based approach, we performed a detailed analysis of affected family members in three families sharing novel and deleterious mutations in NOTCH1 that are likely to contribute to disease (Fig 4a, S2 Table). Phenotypic analysis of affected family members revealed the absence of syndromic disease and indicated a high rate of cardiac valve anomalies and vascular obstructions (Table 1). Two novel NOTCH1 nonsense mutations in two independent families were identified, both of which are located in the extracellular domain of the protein that has been associated with aortic valve disease [11] (Fig 4b). Both stop-gain mutations (Family 1: c.C3765A:p.C1255*; Family 2: c.C2439G:p.Y813*) are located in the N-terminal EGF domain repeats 14–36 of NOTCH1, potentially leaving the ligand binding site of the protein intact if escaping nonsense-mediated decay. The affected mother in family 1 (Family1, II-5) (Fig 4) had a bicuspid aortic valve and severe aortic stenosis. She is a mutation (c.C3765A:p.C1255*) carrier and had been remarried after conceiving two siblings who died of Tetralogy of Fallot (TOF) shortly after birth. Unfortunately, DNA was not available for genetic characterization of the variant in these patients. The affected half-brother had valve dysfunction and a bicuspid aortic valve, which indicates a strong penetrance of the Mendelian variant within this family. The inheritance patterns of valve disease in family 1 highlight the importance of different genetic backgrounds for the expressivity of the associated phenotype. While members of family 1 and 2 had primarily left-sided cardiac malformations, affected members in family 3 suffered primarily from right-sided cardiac defects. We identified two highly conserved NOTCH1 missense mutations in cis in family 3, of which one is novel (c.G578A:p.G193D) and one is rare (7/73804 ExAC alleles, c.G3860A:p.R1287H) in the general population. Both mutations segregate with disease on the same haplotype with TOF or ventricular septal defects in five affected family members. The novel mutation c.G578A:p.G193D is predicted to be highly deleterious by all major algorithms (SIFT score = 0, Polyphen-2 score = 1, CADD score = 32) [22,37,38]. Both mutations resided in the EGF repeat domains 4 and 33 and are located in the extracellular NOTCH1 domain which mediates ligand binding to JAG1 and DLL4 [39]. In order to gain additional evidence of disease association in family 3, we performed linkage analysis on the extended family (9 individuals). The analysis supported the whole-exome based association with NOTCH1 with a genome-wide linkage signal of LOD > 1.5 at four genomic loci (chr1, chr9, chr10 and chr15) (S4a Fig). When only considering rare co-segregating variants under the linkage peaks, solely NOTCH1 remains as a potential candidate gene for this family (S4b Fig). While protein disrupting mutations in NOTCH1 are likely to contribute to disease, interpreting the role of additional deleterious mutations in genes of uncertain function that may act epistatically remains challenging in a family based context (S2 Table). Apart from the highly pathogenic variants in NOTCH1, we identified a family trio with a private co-segregating nonsense mutation in ARHGAP31 (Family IV: c.A4222T:p.K1408*) in our cohort (S5 Fig, S2 Table). ARHGAP31, a Rho GTPase and specific regulator of the RAC1/CDC42 axis has been associated with valve disease and is strongly and specifically expressed in the course of early heart development in mouse [40] and in chicken (S6 Fig). The father, carrying the loss of function variant in ARHGAP31 in family 4 (Family 4, I-1) (S5 Fig) was diagnosed with aortic valve disease, manifested as aortic regurgitation, bicuspid aortic valve and a slight dilation of the ascending aorta. His son had been diagnosed with aortic stenosis early in life, associated with a bicuspid aortic valve, and developed a dilation of the ascending aorta later in life. Both patients showed no history of syndromic disease. Notably, NOTCH1 and ARHGAP31 mutation carriers overlap in their phenotypic spectrum, showing a clustering of valvular phenotypes and aortopathies. Furthermore, ARHGAP31 and NOTCH1 have both been associated with Adams-Oliver syndrome (OMIM#100300, OMIM#616028), which is associated with severe cardiac malformations such as valvular defects and TOF [40,41]. These findings suggest a certain degree of locus heterogeneity in our cohort.
To test the hypothesis that multiple candidate loci contribute to disease and eliminate gene coverage biases in our 53 previously identified candidate genes, we performed Nimblegen targeted re-sequencing for 153 patients from our initial whole-exome sequencing cohort in a trio design which included 82 affected individuals and 71 unaffected family members (Fig 1). We achieved an average coverage of 71.4x and 99% of the target exons were covered at 20x. Principal component analysis of our candidate genes against the ExAC dataset revealed that several genes including MYH6, KMT2D, EP300 and EP400 show an enrichment of rare private mutations despite their high intolerance to common mutations (RVIS: MYH6 = -2.78, KMT2D = -5.29) in the ExAC cohort. This is also reflected by an enrichment of rare benign mutations in these intrinsically variable genes (S7 Fig) based on ClinVar annotation. Despite the significant reduction of benign mutations by applying stringent allele filtering thresholds (p = 0.021; Fisher exact test), it remains a challenging task to discriminate between causative and non-causative missense mutations in common and often asymptomatic diseases of the aortic valve, such as BAV. For this reason, we used the recently developed Combined Annotation Dependent Depletion (CADD) framework to characterize and rank co-segregating mutations in our candidate genes [22]. This framework provides a genome-wide estimate of protein altering mutations and can be used as a meta-annotation tool to rank variants according to their pathogenicity scores. Based on the unbiased CADD score statistics, we observed a significant (p< 0.05, Wilcoxon) enrichment of deleterious mutations in patients compared to unaffected family members and 193 population controls from the Mendel dataset (Fig 5a). This enrichment in variant pathogenicity in our family cohort is due to a small fraction of variants exclusively observed in patients exceeding CADD scores >20 and ranking among the top 1% most pathogenic mutations genome-wide (p < 0.05, chi-square, Fig 5b). In contrast to the enrichment of pathogenic variants in patients, we did not observe an excess of rare or novel variants in cases when compared to controls (p > 0.05, chi-square, Fig 5b). These results are consistent with the notion that natural selection removes rare protein altering mutations from the general population and implies that only a small but highly pathogenic fraction of familial transmitted mutations contributes to disease in our cohort.
To assess whether highly ranked pathogenic mutations are enriched in disease relevant gene sets, we compared their pathogenic burden for 50 hallmark gene sets in patients and controls using the Wilcoxon rank sum statistic. We found a significant enrichment of CADD scores in members of the Notch signaling and the previously highlighted myogenesis gene sets (p < 0.05, FDR adjusted Wilcoxon) (S3 Table). This enrichment was due to mutations residing in genes related to both pathways, including NOTCH1, MAML1, JAG1, MYH6, ARHGAP31 and EPHB4 with CADD scores > 20. Finally, by ranking mutations based on their predicted pathogenicity we were able to exclude mutations in intrinsically variable genes and identified the small but highly pathogenic fraction of transmitted missense mutations that are likely to contribute to CHD in 12 families of our 51 families (S3 Table). Notably, our highest ranked disease-segregating mutations clustered predominantly in genes along the Notch signaling axis (Table 2). We identified two families with mutations in MAML1, a transcriptional co-activator of NOTCH1 involved in outflow tract defects [42,43]. Both rare mutations with CADD scores > 20 (c.G2129A:p.R710Q; c.G913A:p.A305T) segregated with aortic valve disease in five affected patients in both families, including a distantly related cousin. In addition, a family trio with a private mutation (c.G1540A:p.G959A) in JARID2 was identified which segregated with bicuspid aortic valve in the index case and his father. JARID2 has been reported to cause outflow tract malformations in a mouse model and regulates NOTCH1 expression via histone modifications in the developing heart [44]. This is in line with the identification of a patient with an overlapping phenotype and a compound heterozygous mutation (maternal: c.C3830T:p.P1277L/paternal: T4211G:p.V1404G) in the Notch signaling repressor SMARCA4 (BRG1). This analysis is however insufficient to conclusively assess novel gene associations with CHD based on family association alone, especially in the light of emerging evidence supporting a role of deleterious mutations inherited from unaffected family members [45].
The CADD based pathogenic ranking revealed a compound heterozygous mutation in MYH6 in a family with pleiotropic congenital heart disease (S8 Fig). The index case in Family 5 (Family 5, III-3) recessive for rare maternal and paternal MYH6 mutations (maternal:c.G5653A:p.E1885K/paternal: c.G1482A:p.M494I) died shortly after birth of hypoplastic left heart syndrome. This is consistent with a previously described disease association of recessive MYH6 mutations in two patients with HLHs [46]. Notably, four affected family members in Family 5 are carriers of the paternal mutation that co-segregated in incomplete penetrance with aortopathy, septal defects and bicuspid aortic valve. This finding overlaps with a study, reporting incomplete penetrant autosomal dominant MYH6 mutations in a family with an identical clustering of LVOTO subtypes [47]. This instance highlights the complexity of assessing causal genotype-phenotype correlations in families with CHD where the deleterious missense mutation with a higher pathogenicity in the index case (maternal CADD = 34/paternal CADD = 17.7) is inherited from the maternal branch of the family without a prevalent history of disease (S8 Fig).
In this study, we performed whole-exome and targeted candidate gene sequencing for a cohort of families with LVOTO in order to dissect the underlying genetic architecture of CHD. We identified a clustering of rare co-segregating mutations in highly conserved genes along the Notch signaling axis.
While both gain and loss of function mutations have been reported for NOTCH1 [11,12,48], identification of families with multiple affected members carrying novel nonsense mutations remains rare in the context of congenital heart disease. In our cohort, we observed a mutational clustering in the extracellular NOTCH1 terminus in three families with 15 affected mutation carriers. This supports the notion that disruption of the cytosolic domain, but not the ligand binding part of the protein contributes to valve disease. Our findings are consistent with recent studies highlighting that the disruption of NOTCH1 downstream signaling can affect epithelial-mesenchymal transition in the course of outflow tract development [30,49]. During this developmental process, that is enriched for rare deleterious mutations in our LVOTO cohort, endocardial cells detach to become a migratory mesenchyme which in turn is crucial for the proper formation of cardiac valves [30]. Notably, we did not observe a stringent genotype-phenotype correlation with NOTCH1 deficiency and LVOTO subtypes across families in our cohort. The presence of right-sided outflow tract obstructions in families 1 and 3 indicates a degree of phenotypic heterogeneity for this locus. This is consistent with multiple disease associations of NOTCH1 with right-sided lesions reporting that up to 18% of NOTCH1 mutation carriers show right sided cardiac malformations [11,50]. The phenotypic expansion in our family cohort is intriguing due to the previous association of NOTCH1 and ARHGAP31 with Adams-Oliver syndrome (AOS) [40,41]. When analyzing the molecular pathways disrupted in AOS and LVOTO, it seems likely that the specific CDC42/RAC1 regulator ARHGAP31, which is highly expressed in the developing mouse heart [40], and the Notch pathway converge in a common signaling route that is critical for outflow tract development. Notably we did not observe characteristics of Adams-Oliver syndrome in patients with NOTCH1 nonsense mutations but one incidence of aortic valve disease.
Furthermore, chromatin remodeling is an important mechanism in neural crest cells in the process of valve cushion formation and required for outflow tract formation. This process is mediated by several of the candidate genes along the Notch signaling axis identified in this study including SMARCA4 JARID2 and MAML1. We speculate that the disruption of chromatin modifiers, such as MAML1, which functions as a transcriptional co-factor for NOTCH1, might predispose patients for the observed phenotype of outflow tract defects in our cohort. Targeted disruption of Notch downstream signaling in neural crest cells by dominant negative MAML1 (DNMAML) results in similar outflow tract malformations in an animal model of disease [43]. This is in line with the tissue specific role of transcriptional Notch repressors JARID2 and SMARCA4 (BRG1). Depletion of Brg1 in endothelial cells frequently results in bicuspid aortic valve in mice and endothelial JARID2 is required for normal cardiac development [44,51,52]. These findings support the role of locus heterogeneity in aortic valve disease where disruption of functionally related members of the same pathway can result in similar phenotypes in mice and men. This is emphasized by the identification of a highly pathogenic and rare disease segregating variant in EPHB4. In a mouse model, Ephb4 loss of function phenocopies the effect of Notch1 mutations resulting in similar aortic malformations [53].
Our study highlights that familial segregation and weighing a variant in favor of pathogenicity can serve as an important resource for discriminating potential disease causing variants in CHD in the absence of large sample sizes. Stringent allele filtering thresholds (ExAC < 0.1% MAF) and CADD score based pathogenicity ranking (scaled CADD scores > 15) are particularly useful in analyzing intrinsically variable genes such as KMT2D and MYH6. Such filtering allowed us to identify two recessive MYH6 mutations in a family with pleiotropic cardiac malformations including HLHs, septal defects and aortopathy. Dominant MYH6 mutations are known to have a variable penetrance and despite the absence of MYH6 expression in the outflow tract, impaired vascular function and decreased blood flow have been reported in an animal model [54]. This highlights how autosomal dominant transmitted missense mutations may create a sensitized genetic background which predisposes a family member to CHD despite being completely penetrant. The presence of a second mutation in the same locus or environmental perturbations would then be required to reach a disease state. This is likely the situation in the index patient with HLHs in family 5 (Family 5, III-3) (S5a Fig), where the maternal MYH6 mutation with a higher pathogenicity is inherited from the branch of the family without a prevalent history of CHD.
While we observed an enrichment of deleterious mutations in highly conserved genes along the multifaceted NOTCH signaling cascade, our study has limitations. First, none of our candidate loci showed a genome wide significant association with disease. This might be due to the sample size of our cohort which provides limited statistical power, the differences in sequencing enrichment kits and sequencing platforms used or more importantly the underlying complex genetic architecture of CHD.
In addition, environmental factors, such as aortic flow, or differences in the genetic background of family members were not considered in our study design. However, these factors are likely to contribute to the phenotypic heterogeneity and incomplete penetrance observed in families with CHD. Further functional assessment of the identified mutations in multiple loci, that may act epistatically, will be needed to unravel their role during the course of disease. Taken together, our study provides a broader insight in the genetic and phenotypic heterogenic landscape of CHD. We identified co-segregating mutations that are likely to contribute to disease in 12 of our 51 families (23.5%) in multiple loci, which is in line with several recent studies highlighting oligogenic inheritance patterns of non-syndromic CHD in families [45,55–57]. While the number of CHD associated genes increases with each exome study, there is also a growing body of evidence that these genes converge in a complex, yet discrete network driving outflow tract and vascular development [16,58,59].
Families were recruited at CHU-Sainte Justine hospital within the French Canadian population. Participants were evaluated by clinical examination, standard 12 lead electrocardiography measurements as well as two-dimensional echocardiography. Written and informed consent was obtained from all patients or from parents in the case of pediatric patients. The study was approved by the CHU-Sainte Justine ethical board commission. The 193 unrelated control samples used for the burden test were selected among 865 samples from the Baylor-Hopkins Center for Mendelian Genomics (BHCMG) that were not from consanguineous family, of North American descent, had no known cardiovascular phenotypes and are unrelated [35].
DNA was extracted from peripheral blood using the Qiagen Gentra Puregene blood kit (QIAGEN, Toronto, Canada). 96 samples (55 affected individuals) were sequenced using standard library preparation protocols with the Agilent SureSelect 50Mb exome enrichment kit and subsequently subjected to 150 base pair, paired-end sequencing on the Illumina HiSeq2000 platform at the McGill Genome Center. Post-quality control reads were aligned to the reference human genome version 19 (hg19) using bwa [60]. A total of 86 samples (51 affected individuals) were sequenced using the TargetSeq enrichment kit developed for the SOLiD5500xl sequencing platform at CHU-Sainte Justine hospital. Sequencing was performed using bar coded multiplex runs of twelve samples on eight flowchips. Mapping was performed using LifeScopeTM 2.0 Genomic Analysis Software, on the hg19 genome builds. Nimblegene target re-sequencing was performed using a custom designed SeqCap panel for candidate genes and sequenced on the Illumina HiSeq2000 platform. Aligned reads were subsequently processed using Picard (http://picard.sourceforge.net) duplicate removal and Genome Analysis Toolkit (GATK v3.2), INDEL realignment, base quality score recalibration and SNP and INDEL discovery using Haplotypecaller according to GATK best practices [61]. The 193 controls from the Mendel Genomics cohort were sequenced on the Illumina HiSeq2000 platform and aligned using the GATK pipeline as described elsewhere [62].
Variants were filtered based on Exome Aggregation Consortium (ExAC) allele filter thresholds for the general population [63]. Variants were prioritized on co-segregation patterns in pedigrees and absence of variants in control individuals within the dataset for which echocardiograms and ECGs were available. Cumulative sums of co-segregating rdSNVs were computed for all protein coding genes based on RefSeq gene annotation and transformed to z-scores. These counts were summed for 10 equal sized successive bins for the Residual Variation Intolerance Scores based on the ExAC population dataset. Significance was assessed under the null hypothesis that there should be no significant relationship between gene-based rdSNVs count and RVI scores (RVIS). In the presence of co-segregating rare mutations in genes that are intolerant or tolerant to mutations, larger counts in the low or high RVI score bins towards the left and right side of the distribution were tested for enrichment using the Wilcoxon test statistic. We further assessed the departure from the null hypothesis by evaluating the significance of the cumulative distribution by transforming z-scores to p-values. Adapted gene set enrichment analysis was performed on z-score based gene ranking for co-segregating rdSNVs with the weighted parameter and 1000 permutations against 50 hallmark gene sets from the Molecular Signatures Database [28,29]. Population stratification and IBD analysis was performed for whole-exome and target enrichment variant calls in order to exclude sample duplications and mismatches within family pedigrees. Principal components analysis on joined variant calls from 1000 genomes, study subjects and control samples from the Mendel cohort was performed using the SNPRelate package (version 1.0.1). Common variants (MAF > 5%). for 2504 individuals from the 1000 Genomes dataset phase 3 release and study subjects were plotted in the R software environment. All variants were LD pruned in SNPrelate with an LD threshold of 0.2 to produce a set of SNPs in approximate linkage equilibrium. PCA plots using the first two principal components were created using the ggplot2 package in R [64]. Prediction on variant deleteriousness for SIFT, Poly-Phen-2, Mutation Taster, Mutation Assessor and CADD scores were obtained through the ANNOVAR [65] annotation toolset for RefSeq gene annotation models.
All prioritized variants for the presented families were validated using PCR-based bidirectional Sanger sequencing using BigDye 3.1 chemistry on an ABI 3130xl genetic analyzer (Applied Biosystems) to confirm mutations detected by whole exome sequencing. PCRs for family members were performed using primer pairs in S4 Table. Electropherograms were aligned to the NCBI build 37 reference sequence of the respective candidate locus and analysed using SeqMan Pro, which is part of the Lasergene suite (DNASTAR).
DNA from whole blood from drawn from family 3 was used for genotyping on the Infinium OmniExpressExome-8 Bead Chip in order to validate the whole-exome sequencing association with NOTCH1. Genotypes were called with 99.8% accuracy. Raw genotypes were filtered using plink [66]. LD based pruning was applied using a window size of 50 SNPs and a total of 24551 markers were used for linkage analysis using Superlink Online [67]. Multipoint analysis was performed in Superlink using an autosomal-dominant mode of inheritance with a Disease Frequency of 0.01 and a MOI of 0, 0.99, 0.99 (S2 Fig). Results from linkage analysis were extracted from Superlink web-based output and subsequently visualized in Rstudio using the ggplot2 package [64].
The URLs for data presented herein are as follows:
1000Genomes, www.browser.1000genomes.org
RVIS, www.genic-intolerance.org/download.jsp
ANNOVAR, www.annovar.openbioinformatics.org
dbSNP, www.cbi.nlm.nih.gov./projects/SNP
Exome Aggregation Consortium (ExAC), www.exac.broadinstitute.org
Gene Set Enrichment analysis (GSEA) tool, www.broadinstitute.org/gsea
Broad derived hallmark gene sets, www.broadinstitute.org/gsea/msigdb/collection
Combined Annotation Dependent Depletion scores, www.cadd.gs.washington.edu
|
10.1371/journal.ppat.1002613 | Structural Basis for Type VI Secretion Effector Recognition by a Cognate Immunity Protein | The type VI secretion system (T6SS) has emerged as an important mediator of interbacterial interactions. A T6SS from Pseudomonas aeruginosa targets at least three effector proteins, type VI secretion exported 1–3 (Tse1–3), to recipient Gram-negative cells. The Tse2 protein is a cytoplasmic effector that acts as a potent inhibitor of target cell proliferation, thus providing a pronounced fitness advantage for P. aeruginosa donor cells. P. aeruginosa utilizes a dedicated immunity protein, type VI secretion immunity 2 (Tsi2), to protect against endogenous and intercellularly-transferred Tse2. Here we show that Tse2 delivered by the T6SS efficiently induces quiescence, not death, within recipient cells. We demonstrate that despite direct interaction of Tsi2 and Tse2 in the cytoplasm, Tsi2 is dispensable for targeting the toxin to the secretory apparatus. To gain insights into the molecular basis of Tse2 immunity, we solved the 1.00 Å X-ray crystal structure of Tsi2. The structure shows that Tsi2 assembles as a dimer that does not resemble previously characterized immunity or antitoxin proteins. A genetic screen for Tsi2 mutants deficient in Tse2 interaction revealed an acidic patch distal to the Tsi2 homodimer interface that mediates toxin interaction and immunity. Consistent with this finding, we observed that destabilization of the Tsi2 dimer does not impact Tse2 interaction. The molecular insights into Tsi2 structure and function garnered from this study shed light on the mechanisms of T6 effector secretion, and indicate that the Tse2–Tsi2 effector–immunity pair has features distinguishing it from previously characterized toxin–immunity and toxin–antitoxin systems.
| Bacterial species have been at war with each other for over a billion years. During this period they have evolved many pathways for besting the competition; one of the most recent of these to be described is the type VI secretion system (T6SS). The T6SS of Pseudomonas aeruginosa is a complex machine that the bacterium uses to intoxicate neighboring cells. Among the toxins this system delivers is type VI secretion exported 2 (Tse2). In addition to acting on competing organisms, this toxin can act on P. aeruginosa; thus, the organism synthesizes a protein, type VI secretion immunity 2 (Tsi2), which neutralizes the toxin. In this paper we dissect the function and structure of Tsi2. We show that although Tsi2 interacts with and stabilizes Tse2 inside the bacterium, the toxin does not require the immunity protein to reach the secretion apparatus. Our structure of Tsi2 shows that the protein adopts a dimeric configuration; however, we find that its dimerization is not required for Tse2 interaction. Instead, our findings indicate that Tse2 interacts with an acidic surface of Tsi2 that is opposite the homodimer interface. Our results provide key molecular insights into the process of T6 toxin secretion and immunity.
| The type VI secretion system (T6SS) is a multifaceted protein export pathway that is widely distributed in Gram-negative Proteobacteria [1], [2]. With a minimal functional requirement for the products of at least 13 genes, this secretion machine rivals the complexity of the more extensively characterized type III and IV systems [3], [4]. Among the conserved components of the T6SS are a AAA+ family ATPase, ClpV, two proteins with sequence similarity to the type IVB secretion proteins IcmF and DotU, TssM and TssL, and several proteins with sequence or structural similarity to non-filamentous phage proteins [5]. The latter group of proteins includes Haemolysin co-regulated protein (Hcp) and Valine glycine repeat protein G (VgrG), which bear structural similarity to the tail protein of lambda phage (gpV) and the puncturing device of T4 phage (gp27/gp5), respectively [6]–[8]. Hcp and VgrG proteins are exported in a co-dependent fashion from nearly all T6SSs characterized to date. In combination, these observations have led to a prominent structure-function model in which the T6S apparatus resembles outward facing phage on the bacterial cell surface [9].
Early investigations of the T6SS focused on its role in modulating bacterial-host cell interactions. These efforts yielded information about the genetic requirements for T6S function and provided evidence that a subset of T6SSs influence pathogenesis by specifically mediating bacterial interactions with eukaryotic cells [10]. In addition to mediating host cell interactions, the T6SS has been shown to regulate gene expression and contribute to biofilm formation [11], [12]. It is not currently understood how the system mediates such diverse phenomena, nor is it known in all cases whether the effects observed are a direct or indirect result of its function.
Recently it has become clear that the T6SS plays a key role in mediating interactions between bacterial cells [2]. This was first observed in P. aeruginosa, where the Hcp secretion island I-encoded T6SS (H1-T6SS) was shown to target an effector protein, Tse2 (type VI secretion exported 2), to other P. aeruginosa cells [13]. Recipient cells lacking a Tse2-specific immunity protein, Tsi2 (type VI secretion immunity 2), were found to be at a competitive disadvantage relative to donor cells possessing Tse2. Although the mechanism of action of Tse2 remains unknown, the fitness advantage bestowed by the protein requires a functional T6SS in the donor cell and close association of donor and recipient cells. The H1-T6SS exports at least two additional effector proteins, Tse1 and Tse3 [14]. These proteins are targeted by the T6SS to the periplasm of recipient cells where they degrade peptidoglycan and thereby provide a competitive fitness advantage for P. aeruginosa donor cells. P. aeruginosa protects its own cells from the action of these toxic proteins by synthesizing cognate periplasmic immunity proteins, Tsi1 and Tsi3.
Tsi2 differs from Tsi1 and Tsi3 in several respects. For instance, Tsi2 is an essential protein in P. aeruginosa, whereas Tsi1 and Tsi3 are dispensable for growth [14]. This reflects the differences in the subcellular sites of action of the associated cognate toxins. Because the T6S export mechanism avoids periplasmic intermediates, immunity proteins for periplasmically-targeted effectors (Tse1 and Tse3) are required only for resisting intercellular self-intoxication. Conversely, Tse2 appears to act in the cytoplasm; therefore, in addition to resisting Tse2 delivered to the cytoplasm via intercellular self-intoxication, Tsi2 is essential for protecting against endogenous cytoplasmic intermediates of Tse2. Owing to their localization in the same subcellular compartment, Tse2 and Tsi2 are able to complex prior to toxin export. In the case of Tse1–Tsi1 and Tse3–Tsi3, the physical separation of the toxins (cytoplasmic) from their immunity proteins (periplasm) prevents such interactions. This likely imparts a unique requirement for Tse2 export – it must be recognized by the secretory apparatus in the context of a protein complex. Since Tsi2 is not exported by H1-T6SS, it must also be dissociated prior to or during the secretion of Tse2. In this way, Tsi2 is analogous to specialized dedicated secretion chaperones involved in the export of effectors from other alternative secretion pathways such as the type III and IV systems [3], [15]. Secretion chaperones are known to function critically both in stabilizing cognate effectors prior to export and in targeting effectors to the secretion apparatus.
It is apparent that bacterial genomes possess an enormous diversity of toxin-immunity modules outside of T6S-associated Tse–Tsi pairs [16]–[18]. Perhaps the most abundant and thoroughly characterized of these are the toxin-antitoxin (TA) systems [19], [20]. Growing evidence supports a general role for TA systems in resistance to stress and persister cell formation [21]. Type II TA systems consist of a cytoplasmic toxin that is maintained–under favorable conditions–in an inactive state by direct binding to a specific cognate antitoxin protein. Upon activation of cellular stress-response pathways, the antitoxin, which is typically less stable than the toxin, is rapidly degraded by cellular proteases including Lon (Long Form Filament), allowing the toxin to act on its target(s). Toxins vary in their mechanism, however most act as either ribosome-dependent or -independent mRNAses [22], [23].
The properties of the Tse2–Tsi2 pair that make it unique among T6S effector–immunity proteins are the same as those that offer analogy to effector–chaperone and toxin–antitoxin systems. In this report, we sought to ascertain the degree of similarity between these systems by interrogating the structure and function of Tsi2. Our results define properties of Tse2–Tsi2 that are shared with both TA and effector-chaperone systems, however we find that the Tse2–Tsi2 system is altogether functionally, structurally, and mechanistically distinct.
We have reported that P. aeruginosa donor cells capable of delivering Tse2 by an active H1-T6SS to P. aeruginosa recipient bacteria lacking tsi2 have a pronounced competitive fitness advantage [13]. However, absolute colony forming units (CFU) of competing bacteria were not determined in these experiments, which precluded defining whether Tse2 causes cell death or stasis in recipient cells when delivered by the T6SS. Lacking this information, the physiological role of Tsi2 – the subject of our current study – in the context of an interbacterial interaction was also not known.
To investigate the role of Tsi2 in resisting T6S-dependent Tse2-based intoxication, we monitored changes in donor and recipient CFU during interbacterial competition experiments between P. aeruginosa strains. Both donor and recipient strains were generated in the P. aeruginosa ΔretS background. The deletion of retS relieves tight negative posttransciptional regulation of the H1-T6SS and reveals a robust T6S- and Tse2-dependent competitive fitness advantage between strains. Recipient strains bore an additional deletion of the tse2 tsi2 bicistron, which renders them sensitive to Tse2. Both tse2 and tsi2 must be deleted in this strain, as the deletion of tsi2 alone is lethal in the presence of tse2. Donor strains were distinguished from recipients by chromosomal lacZ expression from the neutral attB site. Interestingly, we found that while total CFU of the donor strain increased exponentially over the course of the competition experiment, CFU of the recipient remained constant (Figure 1A). Consistent with our earlier findings, this inhibition of proliferation required tse2 in the donor and the absence of tsi2 in the recipient (Figure 1B & 1C).
We considered three explanations for our finding that the overall population of recipient cells lacking Tse2 immunity did not change during competition experiments against donor cells actively exporting Tse2 by the T6SS: 1) recipient cells are efficiently targeted (approaching 100%) and Tse2 is always bacteriostatic, 2) recipient cells are inefficiently targeted and Tse2 is bactericidal, and 3) recipient cells are efficiently targeted, but differentially affected by Tse2 (unaffected, growth-inhibited, or killed). In the latter two scenarios, the balance between proliferation and death (2), or between proliferation, non-proliferation, and death (3), could produce the stable overall population of recipient bacteria observed. For either of these possibilities, we would expect to observe elevated cell death that is Tse2-dependent in competition experiments between a donor bacterium and a sensitive recipient. However, we found equivalent fractions of dead cells when a sensitive strain was competed against a donor strain capable of delivering Tse2 or one lacking Tse2 (Figure 1D). From these data, we conclude that recipient cells are efficiently targeted by the T6SS, and that the function of Tsi2 is to protects cells from stasis induced by Tse2.
The substrates of many bacterial secretion pathways require dedicated chaperones for their export. We hypothesized that in addition to its immunity function, Tsi2 might also serve as a dedicated chaperone for Tse2. Although our earlier work has shown that Tse2 and Tsi2 interact in P. aeruginosa, whether the proteins bind directly was not determined [13]. As a first step toward investigating the involvement of Tsi2 in Tse2 export, we probed for direct interaction between the proteins using an E. coli bacterial two-hybrid (B2H) assay [24]. In the system we employed, fusions of candidate interaction partners are made to a zinc-finger DNA-binding protein, Zif, and the ω subunit of RNA polymerase. Association of the fusion proteins promotes transcription of a lacZ reporter gene as described by Dove and colleagues [25]. The broadly toxic nature of Tse2 is a complicating factor for analyzing the protein in heterologous systems such as E. coli. Indeed, upon induction of its synthesis from B2H vectors, we found that cellular physiology was rapidly modified, obscuring interpretation of results (data not shown). To facilitate the study of Tse2 in the B2H, we used an allele of the gene encoding a non-toxic Tse2 variant, Tse2T79A S80A (Tse2NT). A description of this mutant is provided below. E. coli strains expressing Tsi2 C-terminally fused to the vesicular stomatitis virus glycoprotein (VSV-G) epitope followed by the Zif protein (Tsi2–V–Zif) and Tse2NT–ω showed significantly enhanced LacZ activity over control strains, indicating that Tsi2 and Tse2 directly interact (Figure 2A).
The interactions between dedicated secretion chaperones and their effector substrates often enhance stability of the effector in the bacterial cytoplasm [26]. However, if Tsi2 behaves analogously to typical antitoxin components of TA modules, even in the absence of stress it would be expected to have a shorter lifetime than the toxin – leaving it unable to act directly in stabilizing Tse2 [27]. Therefore, prior to determining if Tsi2 influences the stability of Tse2, we queried the relative stabilities of the two proteins in P. aeruginosa. Western blot analyses of cells following treatment with the protein synthesis inhibitor tetracycline showed an almost complete loss of intact Tse2–V after one hour (Figure 2B). However, no significant decrease in Tsi2–V levels was observed in the cells over the same time period. This was not the result of Tsi2 stabilization by its C-terminal VSV-G fusion, as an N-terminal VSV-G-fused Tsi2 (V–Tsi2) displayed similar stability (Figure S1). The finding that Tsi2 is more stable in cells than its cognate toxin motivated us to further investigate its potential chaperone activity.
Next we sought to ascertain the influence of Tsi2 on the stability of Tse2 in P. aeruginosa. Since Tse2 is toxic in cells lacking Tsi2, this experiment required the use of a non-toxic variant of Tse2. Furthermore, it was necessary that this variant bore only minimal perturbations, such that its stability and overall structure accurately reflected that of the native protein. In light of the additional objective of the study to probe the involvement of Tsi2 in Tse2 secretion, we added the requirement that the non-toxic variant was competent for export through the H1-T6S apparatus.
The mechanism of action of Tse2 is not known and our analyses have so far failed to identify sequence motifs that would facilitate the prediction of residues essential for its function. Therefore, we adopted a scanning mutagenesis approach for defining minimal inactivating mutations. Using site-directed mutagenesis, we generated 15 tse2 alleles encoding adjacent double alanine substitutions at ten amino acid intervals along the length of the protein. In the four cases wherein one of these positions already encoded an alanine, only one substitution was made. Toxicity of the mutant alleles was determined by monitoring the effect of their ectopic expression on the growth of PAO1 Δtse2 Δtsi2. Two alleles, tse2T79A S80A –V (tse2NT–V) and tse2V109A K110A–V, displayed a marked decrease in cytotoxicity relative to the wild-type protein (Figure 3A). Western blot analysis showed that the mutant proteins accumulated to levels similar to those of the wild-type protein, suggesting that their lack of toxicity is due to inactivation of the toxin rather than poor expression or decreased stability (Figure S2).
The sequence and structural determinants for effector export by the T6SS remain unresolved; therefore, we proceeded to empirically determine whether the non-toxic Tse2 variants retained H1-T6SS-dependent secretion. Expression plasmids directing the synthesis of Tse2NT–V and Tse2V109A K110A–V, or Tse2–V were introduced into P. aeruginosa ΔretS Δtse2. The genes inactivated in this strain lead to constitutive export of effectors by the H1-T6SS (ΔretS) and avoid potential competition between native Tse2 and the ectopically-produced protein for the secretory apparatus (Δtse2). As an additional control, we also transformed a plasmid directing the synthesis of Tse2–V into ΔretS Δtse2 ΔclpV1. The clpV1 gene encodes a AAA+ family ATPase that is an important determinant of effector export by the H1-T6SS. Consistent with our earlier finding that Tse2 is a substrate of the H1-T6SS, the wild-type protein was detected in concentrated culture supernatants in a manner dependent on clpV1 (Figure 3B). The level of secreted Tse2NT–V was similar to that of the wild-type protein, whereas secretion of Tse2V109A K110A–V was not detected. From these data, we conclude that Tse2NT is a non-toxic substrate of the H1-T6SS.
With a non-toxic and secreted Tse2 mutant in hand, we were able to test the involvement of Tsi2 in Tse2 stability and secretion by the H1-T6SS. The contribution of Tsi2 to Tse2 stability was determined by comparing the lifetime of Tse2NT–V when co-expressed in cells with Tsi2–V versus when expressed in cells devoid of Tsi2. For co-expression, Tse2–V and Tsi2–V were produced in their native bicistronic configuration under the control of an inducible promoter. Our data showed that the presence of Tsi2–V significantly extends the lifetime of Tse2NT–V in P. aeruginosa. In the absence of Tsi2–V, intact Tse2NT–V was not detected beyond 15 minutes following the inhibition of protein synthesis; however, it persisted for 60 minutes in the presence of the immunity protein (Figure 3C). The short cellular lifetime of Tse2–V was not due to either its fusion to the VSV-G epitope tag nor to secretion via the H1-T6SS (Figure S1).
To determine the contribution of Tsi2 to Tse2 export by the H1-T6SS, levels of Tse2NT–V in culture supernatants from P. aeruginosa ΔretS strains with or without tsi2 were compared using quantitative western blotting. This analysis clearly demonstrated that Tsi2 is not required for Tse2 secretion by the H1-T6SS (Figure 3D). Therefore, despite the direct interaction of Tsi2 with Tse2, and the role of this interaction in stabilizing intracellular Tse2, Tsi2 appears to have no role in targeting Tse2 to the secretion apparatus.
To gain additional mechanistic insights into Tsi2 function, we solved its X-ray crystal structure to a resolution of 1.00 Å (Table S1, Figure 4A). Phasing of the structure was obtained using the multiwavelength anomalous diffraction method on a 1.68 Å resolution dataset collected from a crystal containing selenomethionine-substituted, C-terminal hexahistidine-tagged Tsi2 (Tsi2–H6) [28]. A molecular model fit to a 1.68 Å resolution electron density map was used to calculate phases for a 1.00 Å data set collected from an isomorphous crystal of native Tsi2–H6. Two monomers interacting through extensive contacts were modeled into the crystallographic asymmetric unit (Figure 4B). Each Tsi2 monomer consists of two large α-helices (α1, amino acids 4–26 (based on monomer A); α2, 30–62) arranged as an anti-parallel coiled-coil connected by a short turn (T1, 27–29) (Figure 4C). The remaining C-terminal end of the protein is composed of a short helical segment (α3, 67–72) located between two extended loops (L1 and L2).
In the observed Tsi2 dimer, the long axes of the two monomers are arranged approximately perpendicular to each other and the molecules pack via their coiled-coils. This interaction involves a large (725 Å2, 13% total surface area) and predominately hydrophobic (68%) surface area, indicative of a physiologically relevant interface. In agreement with this, the molecular weight of purified Tsi2 measured by gel filtration chromatography was found to closely approximate that of the dimer (calculated, 19.16 kDa; measured, 19.46 kDa), and we observed strong interaction between Tsi2 monomers using the B2H assay (Figure 4D and Figure S3). Superimposition of the Tsi2 monomers showed that overall the two subunits adopt highly similar structures (Cα r.m.s.d, 1.1 Å).
As Tse2 has proven recalcitrant to in vitro reconstitution, a direct biochemical study of the Tse2–Tsi2 interface has not been feasible. As an alternative strategy, we mutagenized 27 solvent-accessible Tsi2 residues to alanine and probed for effects on toxin immunity as a proxy for functional interaction with Tse2. None of these substitutions, nor a truncation of Tsi2 lacking residues C-terminal of α3, showed a measurable impact on Tse2 interaction (Figure S4). We did not attempt to analyze more extensive truncations of Tsi2, as removing residues from α1 or α2 would likely disrupt its overall fold. From these results we conclude that the interaction of Tse2 with Tsi2 does not require the C-terminal loops and helix of Tsi2, and that interaction is resilient to minor perturbations of the Tsi2 surface.
The difficulty we encountered rationally dissecting the Tse2 binding site of Tsi2 led us to pursue a genetic screening strategy. The screen we designed exploited our ability to detect Tsi2 homodimerization and Tse2–Tsi2 association using the B2H (Figure 5A). A random PCR-generated mutant library of tsi2 was inserted into the pACTR::V–zif B2H vector such that clones lacking nonsense mutations would generate N-terminal fusions to V–Zif (Tsi2*–V–Zif). Next, the B2H was used to identify those pACTR::tsi2*–V–zif clones that did not activate lacZ expression when co-transformed into cells expressing Tse2NT–ω. After cultivation of positive clones, pACTR::tsi2*–V–zif plasmids were isolated, pooled and transformed into cells expressing Tsi2–ω. At this stage, clones of pACTR::tsi2*–V–zif that retained homotypic interaction, and therefore expressed high levels lacZ in the presence of Tsi2–ω, were selected for further characterization. While this second stage of our screen was critical for removing major sources of false positives, including tsi2 nonsense mutations and mutations that abrogated tsi2 expression, we are also aware of the caveat that it systematically eliminated the potential to recover tsi2* clones that affect both Tse2 and Tsi2 interaction. This issue is addressed below.
Despite screening approximately 20,000 Tsi2* clones, we were able to identify only one single amino acid substitution, Tsi2E38K, that specifically abrogated Tse2 interaction when reconstructed and retested in the B2H (Figure 5B). Interestingly, modeling of the electrostatic surface potential of Tsi2 showed that residues in the vicinity of Glu38, including Asp23, Asp45 and Asp49, generate a prominent negatively charged surface feature (Figure 5C). Based on these findings, we hypothesized that this acidic patch on the surface of Tsi2 contributes directly and critically to Tse2 binding. Although our structure of Tsi2 indicates that Glu38 does not mediate intramolecular interactions, we sought to rule-out the possibility that its non-conservative substitution with Lys perturbs native Tsi2 structure and indirectly leads to a loss of Tse2 binding. To this end, we purified Tsi2E38K–H6 and Tsi2–H6 from E. coli and compared their secondary structure by circular dichroism spectroscopy (CD). Consistent with our X-ray crystal structure of Tsi2, the CD spectrum of the wild-type protein showed strong helical character (Figure S5). The CD spectrum of Tsi2E38K–H6 displayed close agreement with the wild-type, suggesting that the Lys substitution does not significantly alter Tsi2 structure (Figure S5).
Tsi2 is a strongly acidic protein (calculated isolectric point, 4.1) with several solvent-exposed negatively charged amino acids located outside of the Glu38-containing acidic surface patch (Figure 5C). To further investigate the specific involvement of this region on Tse2 interaction, we compared the effects of substituting Glu and Asp with Lys within, and outside of, its boundary. In total, we constructed nine additional lysine substitution mutants: three within the acidic patch (D23K, D45K, D49K) and six outside (D30K, E36K, E56K, D69K, E73K, E74K). Using the B2H assay, each Tsi2 substitution mutant was probed for its capacity to both dimerize and associate with Tse2. The nine variants expressed to similar levels as the wild-type protein, and, as expected, none of the substitutions affected Tsi2 dimerization (Figures 5B and S6). Interestingly, while substitutions outside of the acidic patch had no effect on Tse2 interaction, Tsi2 bearing a lysine at position 45 (Tsi2D45K), located within the acidic patch, was incapable of binding to Tse2 (Figure 5B). CD spectroscopy confirmed that Tsi2D45K–H6 retained the secondary structure of the wild-type protein (Figure S5). Surprisingly, no effect on Tse2 binding was caused by the Asp23Lys or Asp49Lys substitutions, suggesting that the residues critical for Tse2 binding within the acidic patch of Tsi2 are Glu38 and Asp45. The observed lack of Tse2 binding by Tsi2E38K–V and Tsi2D45N–V was also reflected in the immunity properties of the proteins. We observed that both proteins fail to rescue Tse2-based toxicity when expressed in E. coli (Figure 5D).
The prominent role of Glu38 and Asp45 in Tse2 binding is further supported by the results of additional mutagenesis studies. Conservative substitutions introduced at these positions displayed a synergistic effect on Tse2 binding. Neither Tsi2 substitutions Glu38Gln nor Asp45Asn alone had a measurable impact on Tse2 interaction, however their combination reduced Tse2 binding by approximately 50%, as determined using the B2H assay (Figure 5E). From these data, we conclude that Glu38 and Asp45 of Tsi2 are major determinants of Tse2 interaction. Furthermore, the substantial loss of Tse2 binding observed upon mutation of these residues to glutamine and asparagine, respectively, suggests that the Tse2–Tsi2 interface is stabilized in part by electrostatic interactions.
As mentioned above, one caveat of our screen for Tse2-binding determinants of Tsi2 is that it excluded those mutations that also disrupt Tsi2 dimerization. It is conceivable that disruption of the Tsi2 dimer has a generally negative impact on Tse2 binding. Such a scenario could explain the difficulty we encountered in identifying Tsi2 substitutions that lose Tse2, but not Tsi2 interaction. To address this issue, we probed the requirement for Tsi2 dimerization in the interaction of the protein with Tse2. Inspection of the dimer interface permitted the identification of several candidate non-conservative single amino acid substitutions that we predicted could destabilize the Tsi2 dimer. To minimize the probability that our mutations – if successful in disrupting the Tsi2 dimer – would not disrupt Tse2 binding, we limited our mutagenesis to single substitutions. Our initial attempts focused on Arg18 and Glu21, which form a network of polar intersubunit interactions at the origin of the non-crystallographic two-fold rotation axis relating the two subunits in the dimer (Figure 6A). However, B2H analyses of alanine substitution mutants at these positions showed that disruption of this network does not significantly destabilize the overall dimer interface (Figure S7). Together these residues constitute the most significant polar intersubunit contacts, thus we concluded that interfering with hydrophobic interactions would be necessary to disrupt the Tsi2 dimer.
Hydrophobic interactions between Tsi2 monomers are extensive. To increase the likelihood that single amino acid substitutions in non-polar residues at this interface resulted in significant perturbation, we replaced selected residues with the large polar amino acid glutamine. Three spatially distributed small hydrophobic interface residues (all >90% solvent inaccessible), Val10, Cys14, and Ala47, were selected, substituted with glutamine, and tested for homodimerization (Figure 6A). Of the three mutant proteins, only Tsi2A47Q displayed decreased activity in the B2H (Figure 6B). Levels of this protein were similar to the other mutant proteins and the wild-type, consistent with the diminished activity observed specifically resulting from a failure to efficiently homodimerize. As an independent measure of dimer formation by Tsi2A47Q, we employed an in vitro cysteine accessibility assay. This assay is based on our observations that Tsi2 possesses only one cysteine residue, and that this amino acid is solvent inaccessible at the dimer interface (Figure 6A). Destabilization of the interface is expected to increase the reactivity of the Cys14 sulfhydryl side chain to small soluble maleimide-containing probes. To test for differential reactivity at this site, we purified Tsi2–H6, Tsi2A47Q–H6, and Tsi2C14A–H6 and Tsi2E21A–H6 as controls, and incubated the proteins with biotin-maleimide. Reactions were separated by SDS-PAGE and protein-biotin-maleimide conjugates were visualized by probing with Neutravidin-HRP. Consistent with our B2H data, Tsi2A47Q–H6 reacted more rapidly than the wild-type protein or Tsi2E21A–H6 (Figure 6C). Tsi2C14A–H6 displayed no observable reactivity under these conditions, indicating that the products observed were specifically due to the reactivity at this site. Taken together with our B2H data, these data strongly suggest that the native Tsi2 dimer interface is significantly disrupted in Tsi2A47Q.
With a Tsi2 dimer-defective variant in hand, we next sought to measure the impact of dimer disruption on heterotypic interactions of Tsi2 with Tse2. First, using the B2H assay, we observed no significant difference between Tsi2–V and Tsi2A47Q–V binding with Tse2 (Figure 6D). As a second, functional measure of Tse2–Tsi2 interaction, we also tested the capacity of Tsi2A47Q to provide immunity to Tse2. Our data showed that Tsi2A47Q–V, like Tsi2–V, provides full protection against Tse2-based toxicity in E. coli (Figure 6E). In total, these data show that the Tse2–Tsi2 interaction is insensitive to the oligomeric state of Tsi2. Taken together with our screening data, we conclude that the regions of Tsi2 important for Tsi2 homotypic versus Tse2 heterotypic interactions are topologically distinct; Tse2 binding occurs at the face of Tsi2 opposite the site of homodimerization.
This study has shown that the Tse2–Tsi2 system has a unique set of properties that do not neatly conform to existing paradigms for toxin-antitoxin and effector-chaperone systems. For example, unlike canonical antitoxins, Tsi2 is more stable than its cognate toxin. This may reflect different physiological functions of the two systems. While the roles of TA systems are variable, and in certain instances remain a matter of debate, it is well established that they serve important functions in gene maintenance and response to stress [20]. For both of these functions, the activity of the TA system involved is mediated by modulation of toxin activity through antitoxin degradation. Our finding that Tsi2 stability greatly exceeds that of Tse2 suggests that this system has not evolved to conditionally release the toxin. Therefore, assuming an adequate expression level, P. aeruginosa strains endowed with tsi2 are likely to possess stable, non-dynamic immunity to growth inhibition by Tse2. In this way, Tsi2 is more akin to certain colicin immunity proteins, which bind their cognate toxin with extraordinary affinity and provide complete protection against both endogenous and xenogenous cognate toxin [17], [29].
Despite the functional disparities between the Tse2–Tsi2 and TA systems, they do possess notable parallels. One common property of TA systems is that the components have strongly opposing electrostatic character [19]. In the majority of instances, the antitoxin is more acidic than its cognate toxin. This is also the case for the Tse2–Tsi2 system, wherein Tsi2 is highly acidic and the difference in the calculated isoelectric points of the two proteins is 2.6. This analogy holds, and is even more striking, when one considers the other Tse-Tsi pairs. The differences in toxin and immunity isoelectric points are greater in magnitude for these proteins (Tse/i1, 4.1; Tse/i3, 3.6), and in both cases the immunity proteins are acidic. A second shared physical attribute of TA and the Tse2–Tsi2 systems is that their antitoxin and immunity components, respectively, display modularity in their homotypic and heterotypic interactions. Although type II antitoxins are highly diverse at the sequence level, recent biochemical and structural analyses of these proteins indicate that they often exist as dimers and, despite their small size, homomeric contacts occur at a site physically removed from the site of cognate toxin interaction [30]. Our discovery that amino acids positions of Tsi2 critical for Tse2-binding reside on the face of the protein opposite from those involved in its homodimerization, taken together with our ability to readily generate specific loss-of-function mutations at either of these sites, strongly suggest an analogous configuration of the Tse2–Tsi2 complex.
We found that co-expression of tsi2 with tse2 leads to a significant increase in the stability of the toxin, suggesting that the two proteins closely interact. Despite this, cells lacking Tsi2 have no detectable defect in Tse2 secretion, indicating that Tsi2 does not–in addition to its immunity properties–play a role in targeting Tse2 to the secretion apparatus. The specialization of Tsi2 as an immunity protein is in line with our current understanding of the function of other T6SS effector immunity proteins, Tsi1 and Tsi3. These proteins reside in the periplasm and therefore are unavailable to assist in ushering their cognate toxins to the H1-T6SS [14]. This leaves open the question of how T6S effectors are recognized by the apparatus. One possibility is that effectors are bound by yet unidentified specialized secretion chaperone(s). In this case, such a protein might remove Tse2 from the Tse2–Tsi2 complex prior to export. Since we observed no impact of Tsi2 on Tse2 secretion, we would expect that such a protein would either bind Tse2 with higher affinity than Tsi2, or that it would bind a region of Tse2 not involved in Tsi2 binding and target the protein to the secretion machine, where Tsi2 would be readily removed.
An alternative explanation for our finding that Tsi2 has no impact on Tse2 export is that the Tse proteins are exported co-translationally. In this model, Tsi2, like Tsi1 and Tsi3, might be present largely to protect against cognate Tse proteins arriving in trans via the T6SSs of adjacent bacteria. Co-translational export of the Tse proteins could also help reconcile how periplasmic effectors, in particular Tse1, which possess numerous cysteine residues, avoid misfolding in the reduced cytoplasmic environment. Based on this model, one would predict that the H1-T6SS and its effectors would be tightly co-regulated. In P. aeruginosa, expression of tse and HSI-I genes (encode the H1-T6S apparatus) are coordinately co-regulated by the Gac/Rsm pathway [7], [31]. Interestingly, this pathway stringently controls expression at the posttransciptional level, which would appear logical if a build-up of cytoplasmic effector was undesirable.
Teng and colleagues recently reported the crystal structure of the N-terminal three-helix bundle (Habc) domain of the yeast SNARE (soluble N-ethylmaleimide-sensitive factor activating protein receptor) Vti1p bound to its epsin-like adaptor protein Ent3p [32]. According to analyses using DALI, the Vti1p Habc domain is the most similar structure to Tsi2 in the current protein databank (Z score, 7.8; Cα r.m.s.d, 1.2 Å). Specifically, a close match of the length and curvature of the two large helices of Tsi2 to helices A and B of the Habc domain leaves these regions of the two structures nearly indistinguishable (Figure 7). Not only are these proteins structurally related, they also appear to interact with binding partners in a spatially and chemically similar fashion. Two acidic residues located on helix B of the Habc domain, Glu42 and Asp46, were identified by Teng and colleagues as critical determinants of Ent3p binding. Substitution of either residue with arginine severely disrupted the interaction of Vti1p with Ent1p and led to mislocalization of Vti1p in yeast [32]. Interestingly, in an overlay of the two structures, these residues are found in close proximity to the acidic residues of Tsi2 discovered in our study to mediate Tse2 interaction (Figure 7). Although the simple structure of Tsi2 necessarily reduces confidence in interpreting the significance of structural similarity to the protein, we find the extent of structural and functional similarity between Tsi2 and the Habc domain striking. Our overall limited understanding of T6S makes it difficult to reconcile the relatedness of Tsi2 to the Habc domain, however it is worth noting that Tsi2 is now the second T6S protein shown to possess structural homology to an N-terminal regulatory domain of a SNARE protein. We recently reported that TagF, a negative posttranslational regulator of the T6SS, displays significant similarity to the N-terminal longin domain of Sec22b [33]. This domain has no structural homology to the Habc domain, however both can function analogously in directing subcellular localization by mediating interactions with adaptor proteins [34]. It will be of interest to determine the evolutionary mechanisms underlying the relationship between Tsi2 and TagF and proteins involved in vesicle trafficking.
The structure of Tsi2 marks an important first step toward a complete molecular characterization of the Tse2–Tsi2 T6S toxin-immunity pair. However, many key outstanding questions remain. Foremost among these remains the mechanism of action of Tse2. Our analysis of the effects of Tse2 on P. aeruginosa cells during intraspecies competition suggests that the protein acts efficiently and specifically to cease growth, and avoid killing targeted cells. Such effects have been observed for TA system toxins that act by cleaving mRNA, such as RelE [35]. Strong evidence that Tse2 also functions as a ribonuclease is lacking, however there are noteworthy indications. For example, the Phyre (Protein Homology/AnalogY Recognition Engine) structure prediction algorithm reports similarity between Tse2 and enzymes that bind and hydrolyze nucleic acids [36]. We found that an acidic patch of amino acids located on the surface of Tsi2 mediates interaction with, and immunity against Tse2. If like most antitoxins, Tsi2 inhibits its cognate toxin by active site occlusion, it is conceivable that the negatively charged character of Tsi2 could engage basic residues of Tse2 that would otherwise participate in nucleic acid binding [37].
Pseudomonas aeruginosa strains used in this study were derived from the sequenced strain PAO1 [38]. P. aeruginosa strains were grown on Luria-Bertani (LB) media at 37°C supplemented, when appropriate, with 30 µg/ml gentamycin, 25 µg/ml irgasan, 40 µg/ml X-gal, and stated concentrations of Isopropyl β-D-1-thiogalactopyranoside (IPTG). Escherichia coli strains used in this study included DH5α for plasmid maintenance, SM10 for conjugal transfer of plasmids into P. aeruginosa, and BL21 pLysS for expression of Tse2 and Tsi2. The tse2, tsi2 genes and tse2 tsi2 and tse2–vsv-g tsi2 bicistronic sequences were PCR-amplified from P. aeruginosa genomic DNA and cloned into pPSV35CV [13], pET29b+ and pET21a+ vectors (Novagen). Site-directed mutants of tsi2 and tse2 were generated using either QuickChange (Stratagene) or Kunkel mutagenesis procedures [39]. Chromosomal fusions and in-frame gene deletions were generated as described previously and were verified by DNA sequencing [7], [40]. The ΔHSI-I strain was constructed such that all sequence between nucleotide 2015 of PA0074 (ppkA) and nucleotide 754 of PA0091 (vgrG1) were deleted.
P. aeruginosa cultures were grown overnight at 37°C in LB broth containing 0.01% L-arabinose. In each experiment, the donor strain contained lacZ inserted at the neutral phage attachment site [41]. LacZ-labeled donor and non-labeled recipient strains were mixed at a ratio of 1∶1 and spotted onto a 0.2 µm nitrocellulose membrane (Whatman) on a 3% LB agar plate containing 0.2% L-arabinose. Competitions were incubated at 37°C and harvested at the indicated time points by resuspending bacterial cells in LB and plating onto LB plates containing 40 µg/ml X-gal for CFU enumeration.
Growth competition assays of P. aeruginosa ΔretS and ΔretS Δtse2 against P. aeruginosa ΔretS Δtse2 Δtsi2 were performed on filters as described above. At 4 hrs after initiating the experiment, the filters were removed from agar plates and resuspended in 3 ml LB broth. The cells were collected by centrifugation, washed once with 1× phosphate buffered saline (PBS) and resuspended in 100 µl PBS. The bacterial suspension was stained with the LIVE/DEAD BacLight Bacterial Viability Kit (Molecular Probes) according to the protocol of the manufacturer. Viability was measured using a fluorescence microscope equipped with FITC and mCherry filters. The ratio of live/dead cells was determined by calculating the green/red fluorescent cells for 12 random fields per competition. Three independent experiments were performed.
C-terminal hexahistidine-tagged Tsi2 (Tsi2–H6) proteins were overexpressed in E. coli BL21 pLysS. Overnight cultured cells were back-diluted 1∶1000 into fresh 2× Yeast Tryptone (YT) media or defined SelenoMet medium base and SelenoMet nutrient mix medium (Athena Enzyme Systems). Expression was induced at an OD600 of 0.5 with 0.1 mM IPTG for 16 hrs. at 20°C. Cells were harvested by centrifugation (8000× g; 20 min, 4°C) and resuspended in lysis buffer [50 mM Tris-HCl, pH 7.5, 0.5 M NaCl, 1% (v/v) Triton X-100, 5% (v/v) glycerol, 1 mM DTT, and protease inhibitor cocktail (Roche Diagnostics)]. Tsi2–H6 was purified by affinity chromatography using a HisTrap FF column (GE Healthcare) followed by size-exclusion chromatography on a HiPrep 16/60 Sephacryl S-200 high-resolution column (GE Healthcare) using the AKTA Explorer FPLC System. Purified proteins were stored in a buffer containing 50 mM Tris-HCl pH 7.5, 500 mM NaCl, 1 mM DTT, and 5% (w/v) glycerol and dialyzed into a buffer containing 5 mM Tris-HCl pH 7.5, 5 mM NaCl, and 1 mM DTT for crystallization.
Crystals of Tsi2-H6 were grown at 25°C by hanging drop vapor diffusion. An equal volume of 10 mg/ml protein sample was mixed with the crystallization solution (0.1 M sodium acetate, pH 5.0 and 8% polyethylene glycol (PEG) 4000). Crystals were cryo-protected in reservoir solution containing 25% PEG 4000 and flash frozen in liquid nitrogen. Diffraction data were collected at the Lawrence Berkeley National Laboratory Advanced Light Source (ALS) Beamline 8.2.1 (University of California, Berkeley). Data were reduced using HKL2000 [42]. Phases were obtained experimentally with data obtained from selenomethionine-substituted Tsi2-H6 for structure solution by multi-wavelength anomalous dispersion (MAD) using the SOLVE program [43]. The final model was built by iterative model building and maximum likelihood refinement with Refmac-5 [44]. Finally, 123 well-defined water molecules were added, and refinement was continued until the R-value converged to 0.144 (Rfree = 0.176) for all reflections to 1.00 Å resolution. The CCP4 [45] suite and XtalView [46] were used for crystallographic calculations. Molecular figures were generated with PyMOL [47] and CCP4 Molecular Graphics [48]. The model was validated using MolProbity [49]. All residues in the final model lie within allowed regions of a Ramachandran plot and 99.4% lie within the Ramachandran favored region. The crystal structure and structure factors have been deposited in the Protein Data Bank (PDB entry 3RQ9) [50].
Tse2 and Tsi2 derivatives were cloned into pBRGP–ω and pACTR–V–zif plasmids [25], [51]. The pBRGP::tsi2–ω and pBRGP::tse2NT–ω derivatives direct the synthesis of Tsi2 or Tse2 wild-type and mutant alleles as N-terminal fusions to the ω subunit of E coli RNAP. Plasmid pACTR::tsi2– V–zif directs the synthesis of the Tsi2-VSV-G fusion to the N-terminus of the zinc finger DNA-binding domain of murine Zif268 (Zif). The tsi2 gene was mutagenized randomly by PCR with Taq DNA polymerase. A pool of plasmids encoding the resulting tsi2 mutants were ligated into the pACTR– V–zif plasmid and transformed into DH5α-F′IQ cells. All resulting transformants were pooled for plasmid isolation. Pooled plasmids were co-transformed with pBRGP::tse2NT–ω into KDZif1ΔZ competent cells and plated onto LB plates containing 12.5 µg/ml tetracycline, 150 µg/ml carbenicillin, 50 µg/ml kanamycin, 40 µg/ml X-gal, and 500 µM Phenylethyl-β-D-galactosidase (tPEG). LacZ negative (white) colonies were selected for inoculation into 96 well plates, subcultured three times to cure plasmid pBRGP::tse2NT–ω and pooled for plasmid isolation yielding pACTR::tsi2*–V–zif. Purified plasmids were digested by ScaI and T7 exonuclease for removal of pBRGP::tse2 NT–ω. After purification, the mutated pACT::tsi2*–V–zif plasmids were co-transformed with pBRGP::tsi2–ω into KDZif1ΔZ competent cells and transformants were plated onto LB plates containing 12.5 µg/ml tetracycline, 150 µg/ml carbenicillin, 50 µg/ml kanamycin, 40 µg/ml X-gal and 500 µM tPEG. LacZ positive (blue) colonies were subcultured and subjected to plasmid isolation and subsequent sequencing analysis.
E. coli KDZif1ΔZ cells were grown to an OD600 of 1.0, permeabilized with 10% CHCl3, and β-galactosidase activity was quantitatively assayed using a Galacto-Light Plus kit as previously described [52]. Assays were performed with at least two individual experiments in triplicate. Representative data sets are shown and values consist of averages based on three independent measurements from one experiment.
Cell associated and supernatant protein samples were prepared as previously described [7]. Western blotting was performed as described previously using α-VSV-G, α-Tse2 and α-RNA-polymerase, with the modification that α-VSV-G antibody probing was performed in 5% BSA in Tris-buffered saline containing 0.05% v/v Tween 20 [14]. HisProbe-HRP Kit was used for direct detection of recombinant His-tagged proteins according to the manufacturer's instructions (Thermo Scientific).
For growth curves, E. coli BL21 pLysS cells harboring pET29b+ expressing Tse2 and Tsi2 derivatives were grown overnight in liquid LB broth at 37°C and back-diluted into LB broth (1∶1000) supplemented with 50 µg/ml kanamycin and 12.5 µg/ml chloramphenicol. Cultures were grown to an OD600 of 0.1–0.2 and induced with 0.2 mM IPTG. OD600 measurements were determined for E. coli strains in LB broth using an automated BioScreen C Microbiology plate reader with agitation at 37°C. Three independent measurements were performed in triplicate for each strain. A VSV-G epitope sequence was fused to tsi2 to allow for analyzing Tsi2 expression. For growth on solid medium, E. coli BL21 pLysS cells expressing Tse2 and Tsi2 derivatives were grown on LB agar plates with or without IPTG induction.
CD spectra were recorded on a Jasco J810 Circular Dichroism Spectrometer using a 1 mm path-length quartz cuvette (Starna). Tsi2 proteins were measured in triplicate at 195–260 nm in 1× PBS buffer, pH 7.5 at 25°C. A total of three scans were recorded and averaged for each spectrum.
Purified proteins were exchanged into a 1× PBS buffer containing 50 mM NaCl, pH 7.5. Biotin-maleimide was solubilized in Dimethyl sulfoxide (DMSO) and added to the protein samples at a final concentration of 10 µM. Protein samples (10 µM) were incubated with biotin-maleimide at room temperature and reactions were quenched at indicated time points by the addition of a final concentration of 0.1 mM Tris, pH 8.0. Western blots were used to detect biotin-maleimide labeled Tsi2 with NeutrAvidin and to detect His-tagged Tsi2 with HisProbe-HRP Kit.
Purified Tsi2 was loaded onto a Superdex-200 10/300GL HR10/30 column (GE Healthcare) equilibrated with a buffer containing 50 mM Tris pH 7, 500 mM NaCl, and 5% glycerol. Protein standards included ribonuclease A (13700 Da), carbonic anhydrase (29000 Da), ovalbumin (43000 Da), conalbumin (75000 Da), and aldolase (158000 Da).
P. aeruginosa Δtse2Δtsi2 strains harboring pPSV35::tse2NT–V, pPSV35::tse2–V tsi2–V or pPSV35::tse2NT–V tsi2–V were grown at 37°C with aeration in LB broth containing 30 µg/ml gentamycin. Overnight cultures were back-diluted 1∶500 into LB containing 30 µg/ml gentamycin and 0.5 mM IPTG. After P. aeruginosa cells were grown at 37°C for 5 hrs., protein synthesis was inhibited with the addition of 250 µg/ml tetracycline. Samples were taken at indicated time points and analyzed by Western blot.
|
10.1371/journal.ppat.1002777 | Candida albicans Scavenges Host Zinc via Pra1 during Endothelial Invasion | The ability of pathogenic microorganisms to assimilate essential nutrients from their hosts is critical for pathogenesis. Here we report endothelial zinc sequestration by the major human fungal pathogen, Candida albicans. We hypothesised that, analogous to siderophore-mediated iron acquisition, C. albicans utilises an extracellular zinc scavenger for acquiring this essential metal. We postulated that such a “zincophore” system would consist of a secreted factor with zinc-binding properties, which can specifically reassociate with the fungal cell surface. In silico analysis of the C. albicans secretome for proteins with zinc binding motifs identified the pH-regulated antigen 1 (Pra1). Three-dimensional modelling of Pra1 indicated the presence of at least two zinc coordination sites. Indeed, recombinantly expressed Pra1 exhibited zinc binding properties in vitro. Deletion of PRA1 in C. albicans prevented fungal sequestration and utilisation of host zinc, and specifically blocked host cell damage in the absence of exogenous zinc. Phylogenetic analysis revealed that PRA1 arose in an ancient fungal lineage and developed synteny with ZRT1 (encoding a zinc transporter) before divergence of the Ascomycota and Basidiomycota. Structural modelling indicated physical interaction between Pra1 and Zrt1 and we confirmed this experimentally by demonstrating that Zrt1 was essential for binding of soluble Pra1 to the cell surface of C. albicans. Therefore, we have identified a novel metal acquisition system consisting of a secreted zinc scavenger (“zincophore”), which reassociates with the fungal cell. Furthermore, functional similarities with phylogenetically unrelated prokaryotic systems indicate that syntenic zinc acquisition loci have been independently selected during evolution.
| The capacity of disease-causing microbes to acquire nutrients from their host is one of the most fundamental aspects of infection. Host organisms therefore restrict microbial access to certain key nutrients in a process known as nutritional immunity. Recently, it was found that infected vertebrates sequester zinc from invading microorganisms to control infection. Therefore, the mechanisms of microbial zinc acquisition represent potential virulence attributes. Here we report the molecular mechanism of host-derived zinc acquisition by the major human fungal pathogen, Candida albicans. We show that C. albicans utilises a secreted protein, the pH-regulated antigen 1 (Pra1), to bind zinc from its environment. Pra1 then reassociates with the fungal cell via a syntenically encoded (genetically-linked) membrane transporter (Zrt1) to acquire this essential metal. Deletion of PRA1 prevented utilisation of host zinc and damage of host cells in the absence of exogenous zinc. Finally, we demonstrate that this zinc-scavenging locus arose in an ancient fungal lineage and remains conserved in many contemporary species. Syntenically arranged zinc acquisition systems have evolved independently in the fungal and bacterial kingdoms, suggesting that such an arrangement is evolutionary beneficial for microorganisms.
| Assimilation of essential nutrients by pathogenic microorganisms from their host environment is one of the most fundamental aspects of infection. Host organisms therefore restrict microbial access to certain key nutrients in a process termed nutritional immunity. The mechanisms of iron sequestration, together with the strategies that successful pathogens employ to overcome this restriction have been extensively studied [1]. Zinc is the second most abundant trace metal in vertebrates and an important cofactor for around 9% of eukaryotic proteins [2]. However, unlike iron, the microbial mechanisms of zinc acquisition are not as well understood. Recently, Corbin and coworkers demonstrated that infected mice actively sequester zinc from invading bacteria [3]; therefore, the scope of nutritional immunity has expanded beyond iron [4] and the mechanisms of microbial zinc acquisition represent potential virulence factors.
Candida albicans is one of the few fungal species of the normal human microbiome. Although typically a commensal of the oral cavity, gastrointestinal and urinogenitary tracts, C. albicans is also an extremely frequent cause of superficial infections such as vaginitis. Moreover, common iatrogenic procedures, such as gastrointestinal surgery, implantation of a central venous catheter or antibiotic treatment are major risk factors for disseminated candidiasis. This form of systemic candidiasis is now the third most common cause of nosocomial bloodstream infections and the mortality of severe sepsis caused by Candida species is over 50% [5].
C. albicans virulence relies on a number of factors, including morphological plasticity, the expression of adhesins and invasins, robust stress responses, immune evasion, metabolic flexibility and nutrient acquisition [6], [7], [8]. A number of studies have focused on how C. albicans assimilates iron [9]; however the mechanisms of zinc acquisition by pathogenic fungi are poorly understood.
In the current study we sought to elucidate the mechanism of C. albicans zinc acquisition from host cells. We found that C. albicans secretes a scavenger protein (a “zincophore”), Pra1, which sequesters zinc from host cells and re-associates with the fungus via a co-expressed zinc transporter, Zrt1. Furthermore, we show that syntenic zinc acquisition loci are conserved in many fungal species with functional similarities to bacterial ABC transport systems.
Our first objective was to determine whether C. albicans can acquire zinc from host cells. During colonisation of the oral mucosa, vagina or gastrointestinal tract, C. albicans coexists with other members of the microbial flora and is exposed to a complex milieu of nutrients. However, following infection of otherwise sterile body sites, the only nutrients available to the fungus are from host cells, tissues or extracellular matrix and fluids. We therefore decided to focus on a specific stage of systemic candidiasis: interaction with human endothelial cells.
We first created zinc-depleted cell culture medium by treating Dulbecco's Modified Eagle's Medium (DMEM) with CHELEX-100 beads and then reconstituting all metals, with the exception of zinc, to their original concentrations (DMEM-Zn). To ensure that this zinc restriction did not adversely affect the endothelia, monolayers of HUVECs were incubated with DMEM or DMEM-Zn for 24 h and host cell damage was assayed by measuring release of lactate dehydrogenase (LDH). Zinc depletion did not result in increased LDH release, demonstrating that removal of this metal was not cytotoxic over the investigated time frame (data not shown).
To investigate whether C. albicans can utilise host zinc, zinc-starved yeast cells were incubated either with or without endothelial monolayers in zinc-depleted medium for 3.5 h and hyphal length determined as a measure of growth (Figure 1A). C. albicans formed significantly longer hyphae in the presence of endothelial cells (Figure 1B). This was not due to enhanced filamentation upon contact with endothelial cells per se, as supplementation of the medium with zinc restored growth. (Figure 1B). These observations suggested that C. albicans can use zinc from host cells.
To investigate this further, we treated the cells with zinquin, a specific dye which fluorescently labels zinc. As shown in Figure 1A, the mother cells of hyphae (both in the presence and absence of endothelia) stained positive for zinquin, suggesting that the inoculated yeast cells carried over some zinc from the preculture. Emergent hyphae, on the other hand, exhibited punctate zinquin staining only in the presence of endothelia (Figure 1A). Indeed, the fluorescence intensity of endothelial-associated hyphae was around 5-fold higher than hyphae grown without endothelia (Figure 1C).
As the only other zinc available to C. albicans is from host cells, we reasoned that fungal invasion of the endothelia may facilitate zinc acquisition. We therefore performed differential fluorescent staining [10] to directly visualise and discriminate invading and non-invading fungal elements. As predicted, the invading hyphal elements bound zinc at significantly higher levels than the non-invading sections (Figure 1C).
Together these data demonstrate that invading C. albicans hyphae are able to sequester zinc from endothelial cells.
The mechanisms of microbial iron acquisition are well documented [1]. Arguably one of the most widespread strategies is the utilisation of siderophores – small secreted molecules which chelate iron with high affinity and can return to a microbial cell to deliver their iron load. Given the evolutionary success of siderophores, we hypothesised that C. albicans may employ an analogous system, secreting a metal-binding molecule to capture zinc. We reasoned that proteins constitute promising candidates for such a function, as around 9% of eukaryotic proteins can bind zinc [2].
The amino acid sequences of all 55 confirmed and predicted C. albicans secreted proteins (GO term, extracellular region – Candida Genome Database) were analysed using protein-pattern-find for the presence of zinc binding motifs [11]. Of the C. albicans secretome, only the pH regulated antigen 1, encoded by PRA1, contained multiple zinc-binding motifs (highlighted in red Figure 2A).
Three-dimensional modelling of the Pra1 sequence predicted a best-fit with Deuterolysin of Aspergillus oryzae. The three-dimensional structure is shown in Figure 2B. Pra1 was predicted to host two zinc-binding coordination sites (Figure 2C & D) with additional multiple zinc binding histidine residues in the tail (Figure 2D). Deuterolysin is an M35 metalloprotease [12], characterised by an HEXXH motif at the catalytic site. However, manual analysis of the Pra1 sequence revealed degeneration of this motif with a glutamic acid to arginine substitution in Pra1 (Figure 2A & C). Indeed, when we measured the proteolytic activity of purified Pra1 against the metalloprotease substrate casein, Pra1 did not exhibit protease activity, in contrast to a positive control (thermolysin), (Figure S1). From these analyses, we conclude that Pra1 has lost (or has never had) protease activity, but may possess zinc binding capacity.
To test this hypothesis, we directly measured the zinc binding capacity of recombinant purified His-tagged Pra1 [13]. His-tagged β galactosidase [14] was included as a control to account for the metal binding properties of the His tag. Zinc-loaded protein samples were sequentially washed on 10 kDa microspin columns and the zinc content of each flow-through was measured using a PAR assay. Following 10 washes of β galactosidase, zinc was no longer detectable in the flow-through. In contrast, we continued to measure zinc in the Pra1 flow-through until the 23rd wash (Figure 3A), suggesting that Pra1 can loosely bind relatively large amounts of zinc.
The fully washed protein samples were then digested with proteinase K and assayed for zinc release. As shown in Figure 3B, digested Pra1 released 4.6 moles zinc per mole protein. In contrast, His-tagged β galactosidase released only 1.8 moles zinc per mole protein. Therefore, in agreement with our modelling approach (Figure 2), Pra1 can tightly bind approximately three zinc atoms. We also tested the ability of Pra1 to bind iron, calcium, copper and manganese. We observed no binding of iron, calcium or manganese and only moderate binding of copper (data not shown). We therefore conclude that zinc is the dominant metal-substrate of Pra1.
Having established that C. albicans hyphae can sequester host zinc and that Pra1 is able to effectively bind this metal in vitro, we anticipated that Pra1 is responsible for zinc sequestration during invasive growth. C. albicans zinc-starved wild type, pra1Δ and pra1Δ+PRA1 strains were used to infect endothelial cells in the absence or presence of exogenous zinc. All strains formed hyphae on endothelia (Figure 4A). However, in the absence of exogenous zinc, pra1Δ hyphae were significantly shorter than the wild type (Figure 4B). Normal hyphal growth of the pra1Δ mutant was restored either by genetic complementation with a single copy of PRA1 or by reconstituting the medium with 20 µM zinc (Figure 4B). Differential fluorescence staining revealed that, despite the reduced hyphal length of pra1Δ in the absence of zinc, the actual percentage of invading hyphae were similar amongst strains (data not shown). Therefore, the growth defect of pra1Δ was not due to a general inability to physically access intra-endothelial zinc. To determine whether Pra1 binds zinc within endothelia, we stained with zinquin to visualise localisation of the metal. In contrast to wild type and pra1Δ+PRA1 strains, punctate zinquin staining was not observed on the hyphae of the pra1Δ mutant (Figure 4A) and the fluorescence intensity of pra1Δ hyphae was significantly lower (Figure 4C). Therefore, Pra1 is required for zinc sequestration within host endothelial cells.
To examine the impact of Pra1-mediated zinc scavenging on the pathogenicity of C. albicans, we co-incubated different zinc-starved C. albicans strains with endothelial monolayers in the presence of varying concentrations of exogenous zinc for 24 h, and assayed host cell damage by measuring release of lactate dehydrogenase (LDH). The zinc content of the medium did not influence endothelial damage caused by wild type cells, suggesting that C. albicans can efficiently assimilate this micronutrient directly from the endothelia. In contrast, pra1Δ caused very little damage to host cells in the absence of exogenous zinc. However, reconstitution of the medium with zinc resulted in near wild type levels of endothelial damage by pra1Δ (Figure 5). Therefore Pra1 is specifically required for host cell damage in the absence of free zinc.
PRA1 is expressed at alkaline, but not acidic pH [15]. As PRA1 is specifically required for hyphal extension and endothelial damage under zinc limitation and encodes a zinc scavenger, we predicted that this gene should be responsive to environmental zinc levels. We therefore grew C. albicans harbouring a codon optimised GFP (green fluorescent protein) [16] under the control of the PRA1 promoter (PPRA1) in Lee's medium buffered to either pH 5.5 or pH 7.4. As expected [15], PPRA1 drove robust GFP expression at alkaline, but not at acidic pH. The addition of exogenous zinc (100 µM) to Lee's pH 7.4 fully blocked PPRA1-GFP expression (Figure 6). Zinc depletion at pH 5.5, on the other hand, did not induce PRA1 expression, indicating that PRA1 is strongly repressed at acidic pH. We also tested the effect of other metals (iron, copper and manganese at 100 µM) on PRA1 expression at alkaline pH. In contrast to high levels of zinc, these other transition metals did not repress PRA1 expression (data not shown). Therefore, PRA1 expression is regulated by both environmental pH and zinc.
As noted by Nobile and colleagues [17], PRA1 shares its upstream intergenic region with ZRT1, which encodes a predicted high affinity zinc transporter. Moreover, these two genes are transcriptionally co-regulated, belonging to the same transcription module [18] and exhibit similar expression patterns during invasion of oral epithelial cells [19] and liver tissue [20].
We postulated that PRA1-ZRT1 represents a zinc acquisition locus. We therefore deleted ZRT1 in C. albicans and tested the growth behaviour of the resultant mutant in zinc depleted medium. The zrt1Δ mutant grew poorly in the absence of exogenous zinc (Figure S2). Growth was restored to wild type levels either by genetic complementation with a single copy of ZRT1 or by supplementing the medium with zinc. Next, to investigate whether Zrt1 was also required for assimilation of zinc from host cells, we assayed fungal growth with or without endothelial monolayers in zinc-depleted medium. Under control conditions without endothelial cells, wild type, zrt1Δ or pra1Δ strains grew poorly (Figure 7). In the presence of endothelia, wild type cells formed larger, regularly shaped micro-colonies. In contrast, growth of zrt1Δ and pra1Δ was not enhanced by the presence of endothelia (Figure 7). Supplementation of the medium with zinc resulted in large regularly shaped micro-colonies for all strains, irrespective of the presence of endothelia (Figure S3). This suggests that C. albicans employs both Pra1 and Zrt1 for efficient assimilation of zinc from host cells.
We conclude that, in agreement with phylogenetic evidence and ZRT1 over-expression analysis [17], [21], Zrt1 is indeed a zinc transporter and that PRA1-ZRT1 represents a zinc acquisition locus.
Amich and co-workers [22] have recently reported that A. fumigatus zrfC (encoding a high affinity zinc transporter) shares its upstream intergenic region with aspf2 (encoding a fibrinogen binding allergen). The authors also demonstrated that these genes are co-regulated and required for growth under zinc limitation.
Interestingly, C. albicans Pra1 and A. fumigatus Aspf2 have similar properties [17], [23]. Indeed, upon sequence alignment, we observed 43% identity between the Pra1 and Aspf2 amino acid sequences and 48% identity between Zrt1 and ZrfC sequences. As A. fumigatus aspf2-zrfC and C. albicans PRA1-ZRT1 are syntenic and all four genes required for efficient zinc assimilation, it would appear that the zinc acquisition locus is conserved in these two species. We therefore investigated whether this locus is also conserved in other species.
BLASTp analysis with the C. albicans Pra1 amino acid sequence as query identified orthologues in a diverse, yet limited, number of fungal species (Figure S4).
In agreement with Amich et al. [22], we observed synteny of the A. fumigatus orthologues of PRA1 (aspf2) and ZRT1 (zrfC). Indeed, PRA1 and ZRT1 orthologues were syntenic in all sequenced Aspergillus species. Within the Candida (CUG) clade, ZRT1-PRA1 synteny was conserved in some, but not all species (Figure 8). Interestingly, in the distantly related Basidiomycetes, Ustilago maydis and Sporisorium reilianum, PRA1 and ZRT1 orthologues also share their upstream intergenic region. We did not detect PRA1 orthologues in Zygomycete or Microsporidia species. However, the basal Chytrid Spizellomyces punctatus does have a PRA1 orthologue, although it does not share synteny with a ZRT1 orthologue. Together, these observations suggest that PRA1 arose in an ancient fungal lineage and that its syntenic arrangement with ZRT1 occurred before divergence of the Basidiomycota and Ascomycota – an event which occurred at least 452 million years ago [24]. Subsequently, however, PRA1 has been lost multiple times. Indeed, most contemporary fungal clades encompass species both with and without PRA1 orthologues (Figure 8).
PRA1 orthologues are strictly specific to fungi. ZRT1, on the other hand, belongs to the ZIP (ZRT/IRT-like protein) family of transporters, an ancient family found in both eukaryotes and prokaryotes [25]. Unlike their fungal counterparts, however, bacterial ZIP transporters are encoded at loci unlinked to other known zinc acquisition system components. Rather, bacteria possess ZnuABC systems for high affinity zinc uptake [26]. These ABC transporters generally consist of an ATP-binding protein (ZnuC), a permease (ZnuB) and a plasma membrane or periplasmic substrate-binding protein (ZnuA), encoded at the same locus.
Significantly, ZnuB permeases are not phylogenetically related to the ZIP family of zinc transporters and the ZnuA zinc binding proteins are unrelated to fungal Pra1 orthologues.
In Gram negative bacterial ABC transport systems, the soluble binding protein (ZnuA) associates with its substrate in the periplasm for delivery to the cognate membrane permease (ZnuB) [27]. As soluble Pra1 has been shown to reassociate with the C. albicans cell surface [13], we hypothesised that, analogous to the bacterial ABC systems, Zrt1 and Pra1 may physically interact to facilitate zinc uptake.
Three-dimensional modelling of Zrt1 and molecular docking algorithms predicted an interaction between Zrt1 with Pra1 (Figure 9A). To test this prediction experimentally, we exposed C. albicans to recombinant His-tagged Pra1 and visualised binding with a fluorescently conjugated anti-His antibody. We detected association of rPra1 to wild type, but not zrt1Δ cells. Complementation of the zrt1Δ mutant with a single copy of ZRT1 restored Pra1 binding (Figure 9B). In a parallel approach, we exposed C. albicans to recombinant His-tagged Pra1 and assayed binding via Western blot analysis. As shown in Figure 9C, the presence of Pra1 was clearly detectable in protein extract from wild type, but not zrt1Δ cells. Complementation with ZRT1 restored Pra1 binding albeit to a slightly lesser extent than the wild type. Together these data demonstrate that Zrt1 is essential for reassociation of soluble Pra1 to C. albicans cells.
Whilst iron acquisition is an established virulence determinant of many microbial pathogens, the role of zinc, and other essential transition metals, in pathogenicity, has historically received less attention. Recently, however, it has been demonstrated that Staphylococcus aureus-infected mice can actively sequester zinc and manganese from invasive bacterial cells. Therefore the scope of nutritional immunity has expanded beyond iron [4]. Moreover, a number of bacterial zinc transport systems have now been characterised and shown to be involved in virulence [28], [29]. In contrast, the mechanisms employed by human pathogenic fungi to acquire zinc from their hosts have remained unclear.
Pra1, was first identified as a fibrinogen-binding mannoprotein [23], [30], and then, based on expression studies, designated pH-regulated antigen 1 [15]. Pra1 is notable for its complex relationship with innate immunity [31]. On the one hand, Pra1 expression can increase C. albicans recognition by neutrophils, as it serves as the major ligand for the leukocyte integrin αMβ2 [32]. Indeed, deletion of PRA1 can dampen neutrophil activation [33], protect C. albicans from killing by leukocytes and actually increases the virulence of C. albicans in some animal models [34]. In this context, it is noteworthy that PRA1 expression is 29-fold down-regulated during incubation with human blood [35]. On the other hand, expression of Pra1 benefits C. albicans by recruiting complement inhibitors (factor H, factor H-like protein-1 and C4b-binding protein) to the fungal cell surface and by binding C3, thus preventing its cleavage to C3a and C3b [13], [36], [37].
In the current study we provide evidence that C. albicans Pra1 is an extracellular zinc scavenger: (i) invading C. albicans hyphae sequestered zinc from endothelial cells (Figure 1); (ii) the C. albicans secretome hosts a single protein (Pra1) with multiple zinc binding motifs; (iii) the predicted three-dimensional structure of Pra1 accommodates several zinc binding sites (Figure 2); (iv) recombinant Pra1 exhibits zinc binding capacity (Figure 3); (v) PRA1 deletion precludes sequestration and utilisation of host zinc by C. albicans (Figure 4); (vi) PRA1 is required for host cell damage only in the absence of exogenous zinc (Figure 5); (vii) Zrt1 is essential for cellular reassociation of Pra1 (Figure 9). Based on these data we propose the following model for C. albicans zinc acquisition from the host cell (Figure 10). Following host cell invasion, PRA1 is expressed due to the alkaline pH and low soluble zinc status of the intracellular environment [38], [39], [40], [41]. A fraction of secreted Pra1 is released from the fungal cell surface [32]. This component binds host cellular zinc (either free cytosolic or bound to host protein) and returns to the fungal cell via physical interaction with Zrt1 to deliver its zinc load. Based on this, we suggest the new term “zincophore”, for a secreted zinc binding protein which can sequester this metal from the environment and reassociate with the microbial cell.
In C. albicans, Pra1 interaction with components of the innate immune system may be related to its zinc-scavenging function. We note that the cofactor activities of both C4b-binding protein and factor H are zinc dependent [42] and that factor H binds zinc at multiple locations on its surface [43]. It is tempting to speculate that interactions between Pra1 and these complement regulators are zinc mediated. In line with this concept, it is noteworthy that the pneumococcal histidine triad protein of Streptococcus pneumoniae has both zinc binding and factor H recruitment properties [44], [45].
Although it is clear that Pra1 plays important immunomodulatory roles, we propose that Pra1 has an evolutionary older biological function, which has remained elusive until now. This function is zinc acquisition. This is based on phylogenetic analyses: PRA1 orthologues were identified in a number of fungal species which do not associate with vertebrate hosts in nature (Figure S4). Moreover, we identified PRA1 orthologues in both ascomycetes and basidiomycetes, groups which diverged at least 452 million years ago [24]. Indeed, the presence of a PRA1 orthologue in the basal chytrid, S. punctatus, suggests that the origin of this gene may be even older. However, the accuracy of this assessment is limited by the low numbers of sequenced basal fungi. The relatively high sequence identity of PRA1 orthologues (>30% at the amino acid level) across such distantly related fungal species suggests that the gene is under positive evolutionary selection. Furthermore, conservation of PRA1 and ZRT1 synteny between the Ascomycota and Basidiomycota is indicative of an ancient and highly successful evolutionary adaptation.
It is likely that PRA1-ZRT1 modularity simplifies the regulation of these two genes. Indeed, it would appear that transcriptional circuitry is also conserved. In C. albicans, PRA1/ZRT1 is regulated by environmental pH and zinc availability via Rim101 and Zap1, respectively (this study, [17], [46]). In A. fumigatus, the orthologous gene pair, aspf2/zrfC, is similarly regulated by pH and zinc via PacC (the Rim101 orthologue) and ZafC (the Zap1 orthologue), respectively [22]. It remains to be determined whether the orthologous zinc scavenger/transporter loci in more distantly related fungal species (e.g. basidiomycetes) are also pH regulated. As bio-available zinc is less soluble at alkaline pH, coupling the regulation of zinc acquisition systems to environmental pH-sensing pathways may be evolutionarily conserved.
Despite the conservation of this zinc scavenging system throughout the fungal kingdom, it has been lost multiple times during evolution (Figure 8). As Pra1 serves as the major ligand for leukocyte integrin αMβ2 [32], its loss from the proteomes of human pathogenic fungi (such as Paracoccidioides brasiliensis, Cryptococcus neoformans, Histoplasma capsulatum, C. glabrata, C. lusitaniae and C. parapsilosis) may partially contribute to immune evasion by these species.
However, a clear correlation between retention/loss of PRA1 and ecological niche is not obviously apparent, and the dynamic loss of PRA1 orthologues from the genomes of many modern fungal species suggests that these species rely on alternative uptake systems. Indeed, given the essential nature of zinc, it is unlikely that fungi rely solely on one single acquisition system. It is possible that other (Pra1−) fungi rely solely on zinc transporters for uptake of this metal from their environment. Indeed, fungal zinc transporters fall into two relatively distinct phylogenetic clusters: those related to C. albicans ZRT1, and a second class, related to C. albicans ZRT2 (Figure S5). Alternatively, a different zinc-binding protein may be employed: in principle, another secreted protein with zinc binding properties has the potential to function similarly to Pra1. Finally, some fungi may synthesise secondary metabolites to capture environmental zinc.
Although this is, to our knowledge, the first description of a microbial zinc scavenger (“zincophore”), the Pra1-Zrt1 system shares some functional similarities with bacterial zinc ABC transporters. In both cases, the genes encoding an extracellular zinc binding protein (Pra1 in fungi, ZnuA in bacteria) and zinc transporter (Zrt1 in fungi, ZnuB in bacteria) are encoded at the same locus – bacteria additionally encode an ATP-binding protein (ZnuC) [26]. In prokaryotic ABC import systems, the periplasmic or membrane-associated binding protein (ZnuA) directs its substrate to the permease (ZnuB), thus facilitating import via ATP-hydrolysis (ZnuC) [27]. Similarly, the extracellular zinc-binding protein of C. albicans, Pra1, reassociates with the fungal cell via interaction with the transporter, Zrt1 (Figure 9). Therefore, our data indicate that syntenic zinc acquisition loci, encoding an extracellular zinc binding protein and a membrane transporter, have been selected twice during evolution.
Future studies on the C. albicans zinc acquisition system (or its counterpart in other fungi) will be required to further elucidate details of the Pra1-Zrt1 interaction. For example, our own three-dimensional modelling approach indicates that Pra1-Zrt1 may interact via histidine residues, possibly through the formation of zinc bridges. Therefore, it is possible that zinc-loaded Pra1 binds to Zrt1 with higher affinity than apo-Pra1. Such a simple mechanism may enhance zinc delivery to the fungal cell and may even facilitate release of Pra1 for multiple rounds of zinc scavenging.
C. albicans strains were routinely grown in YPD (1% yeast extract, 2% Bacto peptone, 2% glucose). Transformants were selected on SD medium (0.17% Difco yeast nitrogen base; 0.5% ammonium sulphate, 2% glucose, 2% Oxoid agar) supplemented with appropriate amino acids and/or uridine. For zinc starvation, cells were grown in limited zinc medium (LZM). LZM was prepared as described previously [47] with the following modification: zinc was omitted and FeCl2 was included at 25 µM. LZM was supplemented with ZnSO4 as indicated. For alkaline-induced gene expression, Lee's medium, buffered to pH 5.5 or pH 7.4 was prepared as described previously [48]. For zinc-limited cell culture medium Dulbecco's Modified Eagles Medium (DMEM, PAA, or, where stated Promocell Endothelial growth medium, Promocell) was stripped of metals by incubating with 10% w/v CHELEX-100 beads (Sigma) overnight at 4°C with shaking. All metals with the exception of zinc were then restored to the following concentrations: NaCl 4.4 mg/l, MgSO4 97.7 mg/l, NaHCO3 3.7 mg/l, CaCl2 265 mg/l, KCl 400 mg/l, NaH2PO4 109 mg/l. ZnSO4 was supplemented as indicated.
C. albicans strains used in this study are listed in Table S1. Deletion strains produced in this study were generated in the BWP17 background [49] as described previously [50], [51]. A PPRA1-GFP reporter was generated by amplifying the PRA1 promoter and cloning into pGFP [16] at XhoI and HindIII sites. All strains were verified by colony PCR. Primers used for mutant production and verification are listed in Table S2. Complementation plasmids were generated by amplifying the gene of interest, including the upstream and downstream intergenic regions followed by cloning into CIp10 [52]. Resultant complementation constructs were linearised with StuI and transformed into strains as stated above.
Endothelial cells (human umbilical cord endothelial cells – HUVEC) were maintained as described previously [53]. For short term (3.5 h) infections, 105 endothelial cells were seeded on glass cover-slips in 24 well plates and grown for 2 days. Monolayers were washed 3 times with PBS and the medium replaced with zinc-depleted DMEM cell culture medium supplemented with indicated concentration of ZnSO4. Monolayers were infected with 105 zinc-pre-starved C. albicans cells (via 3 subpassages in LZM at 30°C) for 3.5 h. For long term damage assays, 2×104 cells were seeded into 96 well plates and grown for 2 days. Monolayers were washed 3 times with PBS and medium replaced with zinc depleted DMEM+1% fetal bovine serum (FBS, PAA) supplemented with indicated concentration of ZnSO4. Monolayers were infected with 5×104 zinc-starved C. albicans cells. At 24 h, supernatant was assayed for lactate dehydrogenase (LDH) release as described previously [54]. For micro-colony experiments, endothelial cells were grown in Promocell endothelial growth medium (Promocell), seeded into 24 well plates and grown for 2 days. Monolayers were washed 3 times with PBS, the medium replaced with zinc-depleted cell culture medium supplemented with 20 µM ZnSO4 or 8 µM EGTA and infected with around 50 C. albicans cells. For these experiments, C. albicans strains were pre-grown for 24 h in YPD to ensure the presence of single cells for inoculation. Following 16 h incubation, micro-colonies were photographed using a Leica DMIL inverse microscope.
Endothelial monolayers, infected with C. albicans, were washed 2 times with CHELEX-100-treated PBS and fixed with Roti-Histofix 4% for 30 minutes at room temperature. In subsequent steps, samples were washed 3 times with CHELEX-100-treated PBS between each step and every treatment was performed in CHELEX-100-treated PBS. The extracellular (non-invading) C. albicans cells stained with rat anti-C. albicans Mab CA-1 (Acris antibodies) (dilution 1/2000) at 30°C 130 rpm. C. albicans cells were then counterstained with secondary Mab anti-rat IgM conjugated with Alexa 467 (dilution 1/5000) at 37°C. The endothelial cells were then permeabilized in 0.5% Triton X-100. Entire C. albicans cells (extracellular and intracellular) were then stained for 20 min with Concanavalin-A-488 (Molecular Probes) (17 µg/ml). Labile zinc was then stained with 12 µM zinquin (Sigma) for 3 h at 30°C and 80 rpm. The coverslips were mounted inverted on a microscope slide, with ProLong Gold Antifade Reagent and observed under epifluorescence using filters to detect Alexa-467, Alexa-488 and DAPI (for zinquin).
Invading C. albicans cells were determined as previously described [54]. Hyphal length was determined using Leica Application Suite software. Hyphal-zinc localisation was determined by selecting the relevant section (either the entire hypha, extracellular section only or invading section only) of each filament using Leica Application Suite AF selection tool and measuring fluorescence. The fluorescence values were then normalised per 0.5 µm2. The experiment was performed twice in triplicate and at least 25 cells were measured (both for length and fluorescence) per replicate (<150 cells per sample).
The amino acid sequences of all 55 confirmed and predicted C. albicans secreted proteins (GO term: extracellular region, Candida Genome Database) were analysed using protein-pattern-find (bioinformatics.org/sms2/protein_pattern) for the presence of zinc binding motifs [11]. Three-dimensional models of both Pra1 and Zrt1 were generated using Phyre2 (www.sbg.bio.ic.ac.uk/phyre2). Only sections of the protein which were modelled with a confidence of 30% or more are displayed in the predicted structures. The Pra1 best structural hit (Deuterolysin from A. oryzae) was additionally identified using SWISS-MODEL tool (swissmodel.expasy.org) [55] and metal binding sites predicted with Findsite-metal (cssb.biology.gatech.edu/findsite-metal) [56] and 3D ligand site (www.sbg.bio.ic.ac.uk/3dligandsite). To predict interactions between the two proteins, the structures generated in Phyre2 were analysed with Patchdock (bioinfo3d.cs.tau.ac.il/PatchDock) [57]. To confirm the predicted interaction shown in Figure 9, PepSite (pepsite.russelllab.org/index.html) [58] was used. Briefly, the C-terminal tail of Pra1 was divided into 10 amino acid segments and tested for interaction with the three-dimensional model of Zrt1. Reciprocally, the N-terminal extracellular tail of Zrt1 (defined by TMHMM - www.cbs.dtu.dk/services/TMHMM) was divided into 10 amino acid peptides and tested for interaction with the three-dimensional model of Pra1.
Proteolytic activity of recombinant proteins against casein was tested with the Enzchek Protease Assay Kit (Molecular Probes). Briefly, BODIPY FL casein (10 µg/ml) was incubated with 0.5 µg recombinant Pra1 or thermolysin (Sigma) in 50 mM Tris-HCl pH 8 at 37°C and fluorescence was measured at 485 nm excitation/525 nm emission. Selected samples contained the metalloprotease inhibitor phenanthroline (10 mM).
Equal amounts (0.4 nmol) of either recombinantly-expressed His-tagged Pra1 [13] and His-tagged β-galactosidase [14] were loaded onto 10 kDa microspin columns (Ambion), washed twice with 500 µl HS buffer (50 mM HEPES-KOH pH 7.5, 200 mM NaCl), transferred to reaction tubes and incubated with 0.1 mM ZnSO4 for 1 h at 30°C. The zinc-loaded proteins were then transferred to 10 kDa microspin columns and sequentially washed with 150 µl of HS buffer. Each flow-through was assayed for zinc content by PAR assay as described previously [59]. Briefly, 4-(2-pyridylazo)resorcinol (PAR) was added to each sample to 0.1 mM and optical density measured at 490 nm against a ZnSO4 standard curve. Following 30 washes, undigested and digested (40 µg proteinase K, 60°C for 30 min) samples were again assayed for zinc content by PAR assay.
C. albicans strains carrying empty vector, PACT1-GFP [60] or PPRA1-GFP were grown overnight in non-buffered Lee's medium [48] and diluted to OD600 = 0.05 in Lee's medium buffered to either 5.5 (MES) or 7.4 (HEPES), containing standard (3.7 µM) or high (100 µM) ZnSO4. Cell suspensions were incubated in flat, transparent-bottomed, black-walled 96 well plates (Nunclon) in a Magellan TECAN plate reader with 30 s shaking and fluorescence (484/525 nm) measurements taken every 15 min. For each sample condition and time point, auto-fluorescence determined from C. albicans carrying the empty vector was subtracted from PPRA1-GFP and PACT1-GFP values, yielding normalised fluorescence values. For clarity, measurements taken at 2 h intervals were plotted.
Strains were pre-grown for 2 days at 30°C in unbuffered LZM, without additional zinc. Cells were diluted to OD600 = 0.05 in LZM, buffered to pH 7.4 (HEPES) supplemented with indicated concentrations of ZnSO4. Cultures were incubated at 37°C in a Magellan TECAN plate reader with 30 s shaking and OD600 determined every 15 min. For clarity, measurements taken at 2 h intervals were plotted.
For microscopic evaluation of Pra1 binding to C. albicans, 106 cells from an LZM overnight preculture were seeded onto coverslips in 24 well tissue culture plates in 1 ml RPMI 1640 (PAA)+10% FBS and incubated for 3 h at 37°C, 5% CO2. Coverslips were washed once with PBS and incubated for 1 h in 500 µl PBS with 1% bovine serum albumin (BSA) and 60 µg/ml recombinant His-tagged Pra1 [13]. Coverslips were then washed three times with PBS to remove non-bound rPra1 and fixed with 500 µl Roti-Histofix 4%. Following fixation, coverslips were washed three times with PBS and incubated with rabbit anti-His FITC-conjugated antibody (Abcam) (diluted 1∶200 in PBS with 1% BSA) for 1 h at room temperature. Coverslips were then washed three times with PBS and mounted inverted onto microscope slides using ProLong Gold Antifade Reagent (Invitrogen). At least 50 cells per sample were photographed with a Leica DM 5500B microscope (Leica) using the filter set to detect FITC.
For Western blot evaluation of Pra1 binding to C. albicans, strains were grown overnight in Lee's medium buffered to pH 7.4 at 37°C. Cells were then washed once with PBS, adjusted to 5×108 cells/ml in PBS with 40 µg/ml recombinant His-tagged Pra1 and incubated for 1 h at 37°C. Cells were then washed 3 times with PBS to remove unbound rPra1. Total protein was extracted by resuspending cells in 200 µl PBS containing 3 mM KCl, 2.5 mM MgCl2, 0.1% TritonX-100 and protease inhibitor cocktail (Roche) in screw-cap reaction tubes containing 10% w/v glass beads. Cells were lysed in a bead mill (Precellys) for 2 braking cycles of 30 s. at 2800× g. Samples were then centrifuged for 10 min at 18900× g at 4°C. The lysate was transferred to new reaction tubes and protein content quantified with DC protein assay (Biorad). Forty µg of protein extract of each strain were separated in 12% Tris-Glycine SDS-PAGE and blotted onto nitrocellulose membrane. The membrane was blocked in PBS+0.05% Tween-20 (PBST) containing 2% milk powder and 1% BSA for 60 min at room temperature. The membrane was hybridised with rabbit anti-His-tag antibody (Abcam) (dilution 1∶500) in PBST containing 2% milk powder overnight at 4°C. The membrane was then washed 3 times with PBST for 5 min at room temperature and hybridized with secondary anti-rabbit-immunoglobulin-HRP-conjugated antibody (Dako), (dilution 1∶2000) in PBST containing 2% milk powder for 2 h at room temperature. After 3 washes with PBST for 5 min at room temperature and an additional PBS wash of 1 min, bands were identify by Enhanced Chemiluminescence (ECL) using the SuperSignal West Dura kit (Thermo Scientific). The membrane was then stripped according to a standard protocol and re-hybridised with rat anti-α tubulin antibody (AbD Serotec) (dilution 1∶2000) and anti-rat-immunoglobulin-HRP-conjugated antibody (Dako), (dilution 1∶2000) as described above. The experiment was performed twice with comparable results.
The C. albicans Pra1 sequence (Candida Genome Database) was analysed by BLASTp at the NCBI (blast.ncbi.nlm.nih.gov/Blast.cgi) and BROAD (www.broadinstitute.org) databases. Identified orthologues were analysed with MegAlign and aligned with ClustalW, Gonnet series protein weight matrix. Resulting alignment was displayed as a phylogram. Additionally, C. albicans Zrt1 and Zrt2 orthologues were identified by BLASTp at NCBI and BROAD. Selected best hits were aligned with MegAlign.
Syntenic arrangements of fungal PRA1 and ZRT1 orthologues were investigated at the BROAD institute database, yeast gene order browser [61], Candida gene order browser [62], the Candida Genome database or the Aspergillus Genome database (www.aspgd.org). Bacterial zinc transporter encoding genes were investigated at the PathoSystems Resource Integration Center (www.patricbrc.org).
|
10.1371/journal.ppat.1000664 | Inclusion Biogenesis and Reactivation of Persistent Chlamydia trachomatis Requires Host Cell Sphingolipid Biosynthesis | Chlamydiae are obligate intracellular pathogens that must coordinate the acquisition of host cell-derived biosynthetic constituents essential for bacterial survival. Purified chlamydiae contain several lipids that are typically found in eukaryotes, implying the translocation of host cell lipids to the chlamydial vacuole. Acquisition and incorporation of sphingomyelin occurs subsequent to transport from Golgi-derived exocytic vesicles, with possible intermediate transport through endosomal multivesicular bodies. Eukaryotic host cell-derived sphingomyelin is essential for intracellular growth of Chlamydia trachomatis, but the precise role of this lipid in development has not been delineated. The present study identifies specific phenotypic effects on inclusion membrane biogenesis and stability consequent to conditions of sphingomyelin deficiency. Culturing infected cells in the presence of inhibitors of serine palmitoyltransferase, the first enzyme in the biosynthetic pathway of host cell sphingomyelin, resulted in loss of inclusion membrane integrity with subsequent disruption in normal chlamydial inclusion development. Surprisingly, this was accompanied by premature redifferentiation to and release of infectious elementary bodies. Homotypic fusion of inclusions was also disrupted under conditions of sphingolipid deficiency. In addition, host cell sphingomyelin synthesis was essential for inclusion membrane stability and expansion that is vital to reactivation of persistent chlamydial infection. The present study implicates both the Golgi apparatus and multivesicular bodies as key sources of host-derived lipids, with multivesicular bodies being essential for normal inclusion development and reactivation of persistent C. trachomatis infection.
| The genus Chlamydia is composed of a group of obligate intracellular bacterial pathogens that cause several human diseases of medical significance. C. trachomatis is the most commonly encountered sexually transmitted pathogen, as well as the leading cause of preventable blindness worldwide. The prevalence of chlamydial infections, and the extraordinary morbidity and health care costs associated with chronic persisting disease, justifies the research efforts in this area of microbial pathogenesis. Despite their clinical importance, the mechanisms by which these intracellular bacteria obtain nutrients essential to their growth remain enigmatic. Acquisition of sphingolipids, from the cells that chlamydiae infect, is essential for bacterial propagation. This study identifies a requirement for the lipid sphingomyelin from the infected host cell for bacterial replication during infection, and for long-term subsistence in persistent chlamydial infection. Blockage of sphingomyelin acquisition results in premature release of bacteria, a reduced bacterial number, and failure of the bacteria to cause a persisting infection. In this study, we have identified and subsequently disrupted specific sphingomyelin transport pathways, providing important implications on therapeutic intervention targeting this successful microbial pathogen.
| The genus Chlamydia is composed of obligate intracellular prokaryotic pathogens that cause a range of clinical sequelae in humans encompassing ocular, genital, and respiratory tract infections. Consequences of subsequent chronic disease include blindness, infertility, arthritis, and possible coronary heart disease [1],[2]. Despite their notoriety clinically, the molecular interactions between Chlamydia and its host cell that allow for propagation, persistence, and subsequent pathology, remain elusive. The defining biological characteristic of these successful pathogens is a unique process of intracellular development, with an infectious elementary body (EB) initiating uptake into a target host cell. The chlamydial EB subsequently differentiates to the noninfectious, metabolically active reticulate body (RB) within the confines of a membrane-bound vacuole termed an inclusion. Successive growth and replication, giving rise to a large inclusion body containing a multitude of infectious EBs, is contingent upon the acquisition of biosynthetic constituents from the nutrient-rich host cell cytosol. In response to nutrient or immunological stress [3], Chlamydiae can also enter into a persistent phase of development characterized by morphologically altered RBs that can be maintained intracellularly for extended periods of time. Alternating infectious and persistent phases of chlamydial growth correlate with acute and chronic infections in vivo [4]. The cellular biosynthetic constituents that sustain persistent chlamydiae, and allow for emergence from a persistent state, are poorly understood.
The intricacies of this host-pathogen interaction, which allow for acquisition of biosynthetic precursors from the host cell, remain largely undefined. Vacuole-bound chlamydiae attain nucleotides, amino acids, and lipids from the host cell [5]. Eukaryotic-derived phospholipids, sphingomyelin, and cholesterol are found within purified chlamydiae, suggesting that these host-derived constituents traverse the inclusion membrane with subsequent incorporation into the bacterium [6],[7]. Translocation of lipid droplets to the chlamydial inclusion lumen represents one potential source of neutral lipids [8],[9]. Host cell sphingolipids are required for the intracellular growth of C. trachomatis [10], with sphingomyelin attained via the intersection of the chlamydial inclusion with Golgi-derived exocytic vesicles destined for the plasma membrane [11]–[13]. Multivesicular bodies (MVBs), late endocytic organelles abundant in sphingolipids and central to intracellular lipid segregation, also serve as a source for host-derived lipids and a potential intermediate in Golgi to inclusion transport [14],[15]. To further delineate lipid acquisition pathways pirated by the chlamydial inclusion, specific inhibitors of host cell lipid biosynthesis and/or trafficking were evaluated for their effects on chlamydial growth and inclusion development. The present study focuses on sphingomyelin biosynthesis, a host cell pathway validated as essential for growth and replication of chlamydiae by Engel and colleagues [10]. Our studies indicate that sphingomyelin biosynthesis is requisite to inclusion membrane biogenesis and stability, and demonstrate that MVBs are a major source for this essential lipid.
Specific inhibitors of sphingomyelin biosynthesis and trafficking were evaluated for effects on chlamydial growth and inclusion development. Treatment of infected cells with 25 µM myriocin, a potent inhibitor of serine palmitoyltransferase (SPT), the initial enzyme in the biosynthesis of sphingomyelin (Figure 1) [16], revealed striking morphological alterations in inclusion maturation. Confocal analysis of untreated Chlamydia-infected cells revealed normal inclusion development with the vacuole expanding in size from 24 to 36 hr postinfection (pi) (Figure 2). Infected cells cultured in the presence of myriocin, revealed a marked loss of inclusion membrane integrity with disruption of the inclusion and release of intracellular bacteria, initially evident at 24 hr pi (22% of infected cells with disrupted inclusions) and most notable at 30 hr pi (61%) (Figure 2). At 36 hr pi, myriocin-treated cells contained small multiple inclusions of heterogeneous size, rather than the large single inclusion typical of untreated cells (Figure 2). The concentration of myriocin used in these studies had no effect on host cell viability.
The CHO-K1 mutant cell line, LY-B [17], which contains a mutation in the LCB1 gene and therefore does not express SPT, was used to independently test the role of sphingomyelin. C. trachomatis inclusions in LY-B cells showed a collapse of membrane integrity, similar to myriocin treatment (Figure 2). In addition, at 36 hr pi, LY-B-infected cells contained small multiple inclusions comparable to those observed in myriocin-treated HEp-2 cells. The complemented cell line, LY-B/LCB1, supported normal inclusion development comparable to that observed in both CHO-K1 and HEp-2 cells (data not shown), confirming that maintenance of inclusion membrane integrity was dependent on host cell SPT activity.
To confirm that the loss of inclusion membrane integrity was a consequence of a deficiency in host cell sphingomyelin rather than an indirect effect of depleted SPT activity, cells were cultured in the presence of 5 µM dihydroceramide or 5 µM sphingosine prior to infection. Dihydroceramide and sphingosine are precursors of sphingomyelin, positioned downstream of SPT, allowing for the restoration of sphingomyelin synthesis under conditions of SPT inactivity (Figure 1) [10],[18]. These sphingomyelin precursors reversed the detrimental effects of SPT-deficiency in LY-B cells or myriocin-treated HEp-2 cells, with growth and expansion of intact inclusions morphologically comparable to those present in untreated control cells at 24 to 36 hr pi (Figure 2) (data for sphingosine not shown).
The intracellular developmental cycle of C. trachomatis E requires approximately 72 hr to complete, with redifferentiation of RBs to infectious EBs occurring prior to release of infectious progeny. At 24 to 36 hr pi, the expanding inclusion contains predominantly noninfectious RBs that, if released indiscriminately from the infected cell, are incapable of initiating an infectious cycle. The presence of multiple small inclusions at 36 hr pi, under conditions of disrupted host cell sphingomyelin biosynthesis, suggested premature release of infectious progeny and subsequent reinfection. To analyze possible early emergence of infectious EBs, the expression of OmcB, an EB-specific protein detectable late in the developmental cycle, was analyzed. In untreated cells, low levels of OmcB were evident at 30 to 36 hr pi (Figure 3), with peak levels emerging at 48 to 72 hr as inclusions reached maximal size and approached lysis (not shown). Myriocin treatment resulted in expression of OmcB as early as 24 hr pi with EBs being dispersed upon premature loss of both inclusion and host cell membrane integrity (Figure 3). Infected SPT-deficient LY-B cells also displayed early emergence of OmcB-positive EBs, temporally similar to those observed under conditions of myriocin treatment (not shown). Western blot analysis confirmed the higher levels of OmcB at 27–36 hr pi in infected cells treated with myriocin as compared to control cells (Figure 3). In addition, higher levels of infectious progeny were released from myriocin-treated cells versus control cells at early times post infection (Figure 3). Collectively, these results indicate that the absence of sphingomyelin results in loss of inclusion membrane integrity, early redifferentiation, and premature release of infectious chlamydiae.
A distinguishing trait of prototypic C. trachomatis strains is homotypic fusion of inclusions [19]. Infecting a single cell with multiple EBs of a defined serovar, results in multiple bacterial-containing vacuoles that fuse early in the developmental cycle to form a single inclusion. The presence of multiple inclusions at 36 hr pi in sphingomyelin-depleted cells, suggests reinfection with subsequent disruption of homotypic fusion. To analyze the effect of sphingomyelin deficiency on homotypic fusion, cells were infected with a high MOI of five bacteria per cell and inclusion numbers were determined at 16 hr pi (Figure 4). HEp-2 and CHO-K1 cells generally contained a single inclusion per infected cell as shown in the histogram inserts. HEp-2 cells cultured in the presence of 25 µM myriocin or 5 µg/ml fumonisin B1 (a potent inhibitor of sphingonine and sphinosine N-acetyltransferase, Figure 1), or the SPT-deficient LY-B cells, revealed multiple inclusions per cell. Complementation of the LY-B cells with the LCB1 gene, resulted in the restoration of host cell sphingomyelin biosynthesis, and the recovery of the inclusion fusion phenotype as shown by a single inclusion per infected cell (Figure 4). To confirm that lack of inclusion fusion was a consequence of a deficiency in host cell sphingomyelin rather than an indirect effect of depleted SPT activity, cells were cultured in the presence of dihydroceramide and sphingosine prior to infection. These sphingomyelin precursors restored the fusion capability to infected cells cultured under conditions of SPT-deficiency with a majority of cells containing a single inclusion (Figure 4). Collectively, these findings indicate that host cell sphingomyelin biosynthesis is required for homotypic fusion of chlamydia inclusions within a single infected cell.
Persistence is a hallmark of natural chlamydial diseases, and is characterized by the retention of nonreplicating, aberrant reticulate bodies within the host cell for extended periods of time [3]. Host cell sphingomyelin biosynthesis is essential for maintenance of inclusion integrity during normal chlamydial development, and is likely essential during reactivation of persistent infection, a process concurrent with inclusion membrane expansion. The role of host cell sphingomyelin was tested in a model system of IFN-γ-induced persistence [20]. HEp-2 cells were infected with C. trachomatis B, a strain sensitive to IFN-γ-mediated alterations in intracellular growth [21]. Untreated Chlamydia-infected cells revealed normal inclusion development with large inclusions at 48 hr pi, while IFN-γ-treated cells harbored smaller inclusions containing enlarged RBs as confirmed by fluorescence and electron microscopy (Figure 5). The persistent state was reversible as shown by the expansion of the inclusion and reactivation of infectious EBs following removal of IFN at 48 hr pi and culturing in fresh medium for an additional 48 hr (Figure 5). In contrast, culturing in the presence of myriocin during the recovery phase resulted in disruption in inclusion membrane integrity and failure of persistent forms to completely reactivate to infectious EBs (Figure 5). These results were confirmed in an alternate in vitro model system of penicillin-induced persistence [22],[23]. In C. trachomatis serovar B- and servovar E-infected cells treated with penicillin to induce aberrant, persistent chlamydial development, the presence of myriocin during the recovery phase prevented the recovery of infectious EBs (not shown). These studies implicate host cell-derived sphingomyelin as an essential component for maintenance of inclusion membrane integrity during reactivation of persistent chlamydial infection.
The precursors of sphingomyelin are synthesized in the endoplasmic reticulum with subsequent transfer of ceramide to the Golgi apparatus, the site of the final step in sphingomyelin biosynthesis (Figure 1). Hackstadt and colleagues have demonstrated the transport of sphingomyelin from the Golgi to the chlamydial inclusion, with incorporation of the sphingolipid into the inclusion membrane and the cell wall of chlamydiae [12],[13]. MVBs, late endocytic organelles abundant in sphingomyelin, have been proposed to provide essential lipids to the chlamydial inclusion and may be an intermediate in Golgi to inclusion transport [14],[15]. To decipher the source of Chlamydia-acquired sphingomyelin, the phenotypic effects of inhibitors of Golgi and MVB transport on inclusion maturation were compared to inclusion development under conditions of sphingomyelin deficiency. The inhibitors were used at concentrations that disrupt transport of ceramide-derived sphingomyelin from the Golgi apparatus to the chlamydial inclusion, but have no effect on host cell viability [13],[14]. HEp-2 cells were infected with a high MOI of five bacteria per cell and treated with the indicated inhibitors at 1 hr pi, then analyzed for homotypic fusion at 16 hr pi (Figure 6). Control cells generally contained a single inclusion per infected cell as shown in the histogram inserts. HEp-2 cells were cultured in the presence of golgicide A (GCA), a potent, highly specific inhibitor of GBR1 (Golgi BFA resistence factor 1) that disrupts both anterograde and retrograde transport through the Golgi [24]. GCA-treatment revealed a slight disruption in vacuole fusion with a mean of 2.6 inclusions per infected cell (Figure 6), with a similar result observed upon treatment with 1 µg/ml brefeldin A (BFA) another inhibitor of Golgi function [25] (not shown). HEp-2 cells cultured in the presence of 10 µM U18666A, a pharmacological agent that disrupts trafficking from MVBs [26]–[28], revealed multiple inclusions per infected cell (Figure 6), similar to the conditions of sphingomyelin deficiency (Figure 4). Therefore, interruption of sphingomyelin trafficking from the Golgi delayed inclusion fusion, while a block in MVB trafficking completely impeded fusion, implicating MVBs, an organelle abundant in sphingolipids, as a principle source of chlamydiae-acquired sphingomyelin.
To analyze the effect of inhibitors on inclusion maturation, HEp-2 cells were infected with a low MOI of C. trachomatis E, treated with the indicated inhibitors at 1 hr pi and analyzed at 36 hr pi. Confocal analysis of GCA-treated Chlamydia-infected cells revealed a slight delay in inclusion maturation with smaller inclusions compared to those in untreated control cells (Figure 6). There was no evidence of inclusion membrane instability as observed under conditions of sphingomyelin deficiency (Figure 2), indicating that sphingolipids may be acquired from an alternate source such as MVBs. Infected cells cultured in the presence of the MVB inhibitor U18666A, revealed a dramatic interruption in inclusion development with significantly smaller inclusions (Figure 6). There was no evidence of inclusion membrane instability as observed under conditions of sphingomyelin deficiency (Figure 2). However, the complete interruption in RB division and subsequent inclusion expansion, implicates additional MVB-derived constituents necessary for normal chlamydial inclusion expansion and development.
Host cell sphingomyelin biosynthesis is essential for maintenance of membrane integrity during expansion of the inclusion following reactivation of persistent infection (Figure 5). Because trafficking from MVBs was essential to sphingomyelin-dependent inclusion expansion, the potential significance of these sphingolipid-rich organelles in reactivation of persistent infection was analyzed. Following induction of the persistent state by IFN-γ treatment for 48 hr, removal of IFN and subsequent culturing in the presence of the MVB inhibitor U18666A for an additional 48 hr, resulted in a lack of inclusion expansion, disruption in inclusion membrane integrity, and complete failure of aberrant persistent forms to reactivate to infectious EBs (Figure 5). These studies implicate MVB-derived sphingomyelin, and potentially other MVB constituents, requisite to inclusion membrane integrity during reactivation of persistent chlamydial infection.
The present studies were initiated to identify lipid biosynthetic and transport pathways essential to the intracellular propagation of chlamydiae. These studies revealed novel effects on the intracellular development of chlamydiae under conditions that inhibit sphingomyelin biosynthesis. As demonstrated in classic studies by Hackstadt and colleagues, sphingomyelin synthesized in the Golgi apparatus is transported from the trans-Golgi to the chlamydial inclusion with successive incorporation into the bacterial cell wall [12],[13]. In subsequent studies by Engel and colleagues, host cell-derived sphingomyelin was shown to be essential for intracellular development of C. trachomatis and optimal production of infectious progeny [10]. In the present study, we further explore this requirement and demonstrate that sphingomyelin biosynthesis is necessary for stability and expansion of the inclusion membrane during both normal intracellular development and reactivation of persistent infection. Blockage of this pathway results in premature egress, reduced bacterial output, and failure to emerge from a persistent state. Hence, disruption of lipid trafficking may provide a novel means to thwart intracellular pathogens.
Chlamydiae undergo their entire intracellular developmental cycle within an inclusion that is bound by a membrane, providing a protected intracellular environment for bacterial replication. Treatment of infected cells with myriocin interrupted inclusion membrane functionality, with complete disruption of membrane integrity resulting in premature dispersal of intracellular bacteria from their protected niche (Figure 2). Myriocin is a potent inhibitor of SPT, the initial enzyme in sphingomyelin biosynthesis (Figure 1) [16]. Analysis of inclusion development in SPT-deficient LY-B cells, and under conditions of concurrent pretreatment with precursors of sphingomyelin, revealed that the compromise in inclusion membrane integrity was a direct result of host cell sphingomyelin deficiency (Figure 2). Actin and intermediate filaments have been shown to stabilize the chlamydial inclusion, with disruption of these host cytoskeletal structures resulting in loss of inclusion membrane integrity and release of bacteria into the host cell cytosol [29]. In the present studies, immunofluorescence analyses of actin and intermediate filaments of both uninfected and chlamydiae-infected cells revealed no obvious morphological alterations in the cytoskeletal structure upon inhibition of sphingomyelin biosynthesis (data not shown).
The disruption of inclusion membrane integrity under conditions of sphingomyelin deficiency occurred concomitantly with the early redifferentiation of noninfectious RBs to infectious EBs (Figure 3). This implies that the procurement of host cell sphingomyelin may be required for inclusion membrane expansion and stability, and programmed conversion to infectious forms. The signals that trigger the replicative RBs to convert to infectious EBs remain elusive. However, it is clear that this developmental transformation coincides with a contact-dependent interaction of the type III secretion (TTS) system with the inclusion membrane. RBs amass at the periphery of the inclusion, with projections of the TTS system mediating intimate contact between the bacteria and the inner face of the inclusion membrane [30],[31]. The proposed chlamydial injectisome acts as a molecular syringe, translocating effector proteins directly from the intrainclusion chlamydiae to the host cell cytosol [32]. This association may be requisite to RB replication and potentially inclusion expansion allowing for nutrient acquisition from the host cell cytosol [33]. The physical detachment of RBs from the inclusion membrane, coupled to inactivation of TTS, signals the initation of late redifferentiation [32]. In the present studies, lipid deprivation may signal the loss of TTS intimate contact and RB detachment leading to premature conversion of RBs to infectious EBs. Host cell-derived sphingomyelin associates transiently with the chlamydial inclusion membrane and incorporates into the bacterial cell wall [12]. Failure of this sphingolipid to incorporate into the inclusion membrane may cause the normally contiguously intact membrane to become indiscriminately permeable to environmental changes that potentially signal RB to EB conversion. Alternately, incorporation of sphingomyelin into the chlamydial cell wall may be essential to RB division and proliferation, with lack of available sphingomyelin being a potential cue for premature redifferentiation.
A secondary function of the inclusion membrane of C. trachomatis, distinct from inclusion membrane integrity, is homotypic fusion of multiple inclusions to a single vacuole in multiply-infected cells. The resulting multiple inclusions with greater surface area would require more lipid incorporation into the chlamydial inclusion membrane, indicating that early in infection other host cell lipids are available for incorporation into the expanding inclusion under conditions of sphingomyelin deficiency. Fusion of inclusions is a temperature-dependent process that requires export of the chlamydial incA protein to the inclusion membrane [34],[35]. Characteristic homotypic fusion of inclusions was interrupted when multiply-infected cells were cultured in the presence of myriocin (Figure 4). Analysis of the fusion of multiple inclusions in SPT-deficient LY-B cells, and under conditions of concurrent pretreatment with precursors of sphingomyelin, revealed that the disruption in homotypic fusion was a direct result of host cell sphingomyelin deficiency (Figure 4). These studies did not reveal an alteration in IncA incorporation into the inclusion membrane under conditions of sphingomyelin deficiency, implicating a role for host cell sphingolipids in homotypic fusion independent of incA. Culturing C. trachomatis-infected cells under conditions of sphingomyelin deficiency has two distinct phenotypic effects on chlamydial inclusion biogenesis. Interruption in homotypic fusion is observed early in chlamydial inclusion development, while a compromise in inclusion membrane integrity occurs later. These distinct anomalies may result from the failure of sphingomyelin incorporation into the inclusion membrane, implicating a direct role for host cell lipid in maintaining normal inclusion functionality. However, the effect of sphingomyelin deficiency on other lipid biosynthetic or signaling pathways that indirectly alter inclusion biogenesis cannot be disregarded.
Further studies determined the source of sphingomyelin essential to inclusion biogenesis, which includes membrane stability and the capacity for homotypic fusion. As described previously, inhibition of sphingomyelin transport from the Golgi apparatus using the inhibitor BFA, results in smaller, compact inclusions that retain a burst size comparable to untreated controls [12]. In the present studies, this observation was reproduced using both BFA and GCA. In addition, treatment of infected cells with concentrations of BFA or GCA that prevent the incorporation of newly synthesized Golgi-derived sphingomyelin into the chlamydial inclusion, failed to completely disrupt inclusion fusion or inclusion membrane integrity (Figure 6). This implicates another source of sphingomyelin available to the chlamydial inclusion under conditions of disrupted Golgi transport. These studies identify MVBs, late endocytic organelles abundant in sphingolipids and pivotal for intracellular distribution, as a potential source of sphingomyelin essential to homotypic fusion and maintenance of inclusion membrane integrity. U18666A treatment of infected cells, utilizing concentrations that block MVB transport and prevent the incorporation of newly synthesized Golgi-derived sphingomyelin into the chlamydial inclusion [14],[15], revealed complete inhibition of homotypic fusion of inclusions (Figure 6). These findings were identical to the disruption of inclusion fusion observed under conditions of sphingomyelin deficiency (Figure 4). However, inhibition of MVB transport had much more profound effects on RB division and normal inclusion development than what was observed under conditions of sphingomyelin deficiency. A deficit in host cell sphingomyelin resulted in RB division and the expansion of the chlamydial inclusion to a moderate size with subsequent loss of inclusion membrane integrity at 24 to 36 hr pi (Figure 2). In contrast, interruption in MVB transport impeded early RB division and inclusion membrane expansion at a stage in development prior to imposing stress on inclusion membrane integrity. Collectively these studies implicate sphingomyelin, and potentially additional constituents derived from MVBs, essential for inclusion expansion during normal development and the reactivation of persistent C. trachomatis infection. However, a pleiotropic effect of inhibitors of MVB transport, on cellular function or potential acquisition of sphingomyelin from alternate sources, cannot be disregarded.
Within the confines of a protected intracellular environment, chlamydiae coordinate the expansion of the inclusion and acquisition of biosynthetic constituents from the host cell cytosol. In the presence of eukaryotic protein synthesis inhibitors, intracellular development proceeds normally, indicating that inclusion expansion may be linked to host cell lipid biosynthesis. These studies identify host cell sphingomyelin biosynthesis as a requisite to C. trachomatis inclusion membrane biogenesis and functionality. This encompasses inclusion membrane expansion, homotypic fusion, and stability during normal inclusion development and reactivation of a persistent chlamydial infection. In addition, identification of potential sphingomyelin transport pathways may have important implications when deciphering this unique host-pathogen interaction.
Rabbit anti-incG was kindly provided by Dr. Ted Hackstadt (Rocky Mountain Laboratories, NIH, NIAID, Hamilton, MT). Rabbit anti-outer membrane complex protein B (OmcB) was generously provided by Dr. Thomas Hatch (University of Tennessee Health Science Center, Memphis, TN). Monoclonal antibody (mAb) L2I-10 to the major outer membrane protein (MOMP) of C. trachomatis, was kindly provided by Dr. Harlan Caldwell (Rocky Mountain Laboratories, NIH, NIAID, Hamilton, MT). MAb A57B9 against the chlamydial heat shock protein-60 (hsp60), was generously provided by Dr. Richard Morrison (University of Arkansas for Medical Sciences, Little Rock, AK). Antibodies to chlamydial LPS and eukaryotic actin (clone C4) were obtained from Chemicon International (Billerica, MA). TOPRO-3 (monomeric cyanine nucleic acid stain), and secondary antibodies conjugated to Alexa Fluor 488 and Alexa Fluor 568 were obtained from Invitrogen (Eugene, OR). Myriocin, fumonisin B1, dihydroceramide, sphingosine, 3-β-(2-diethylaminoethoxy)-androstenone HCl (U18666A), and brefeldin A were obtained from BioMol International (Plymouth Meeting, PA). Recombinant human IFN-γ was purchased from BD Biosciences (San Jose, CA). Golgicide A was kindly provided by Dr. David Haslam (Washington University School of Medicine, St. Louis, MO).
C. trachomatis serovar E (provided by Dr. Harlan Caldwell) and C. trachomatis serovar B (provided by Dr. Ted Hackstadt) were propagated in HEp-2 cells (ATCC, Manassas, VA) and elementary bodies (EBs) were purified by Renografin gradient centrifugation as previously described [36]. HEp-2 cells were maintained in Iscove's DME medium supplemented with 12.5 mM HEPES, 10% (vol/vol) FBS, and 10 µg/ml gentamicin, and grown at 37°C with 5.5% CO2. CHO-K1, LY-B, and LY-B/LCB1 cells, obtained from Dr. Kentaro Hanada (National Institute of Infectious Disease, Tokyo, Japan), were maintained in Ham's F12 medium supplemented with 10% (vol/vol) FBS, and 10 µg/ml gentamicin at 37°C with 5.5% CO2. Cells were infected by incubating monolayers with Chlamydia EBs at a multiplicity of infection (MOI) of 0.2 or 5 for 1 hr at 37°C, washed and incubated in fresh culture medium for the times indicated.
For immunofluorescence analyses, infected cells were fixed and permeabilized for 1 min with cold methanol. Cells were then incubated with the indicated primary and fluorophore-conjugated secondary antibodies, labeled with the nucleic acid stain TOPRO-3, and mounted in ProLong Anti-Fade (Invitrogen), as previously described [14]. Images were acquired using a Zeiss LSM510 Meta laser scanning confocal microscope (Carl Zeiss Inc., Thornwood, NY) equipped with a 63X, 1.4 numerical aperature Zeiss Plan Apochromat oil objective. Confocal Z slices of 0.5 µm were obtained using the Zeiss LSM510 software.
One hour post infection (pi), infected HEp-2 cells were incubated with medium containing inhibitors and the effects on inclusion development were determined by immunofluorescence, Western blot analysis, and infectivity assays, when indicated. To quantify the disruption of inclusions, one hundred infected cells were scored by fluorescence microscopy as indicated. Data are presented as the mean percent of disrupted inclusions. To quantify the number of inclusions per cell, one hundred infected cells were scored by fluorescence microscopy at 16 hr pi and presented as the mean number of inclusions per infected cell.
Infected monolayers cultured in the presence of myriocin or IFN-γ were scraped from culture dishes, and supernatant and cells were analyzed to determine the number of infectious forming units (IFU) per ml (per 7.5×105 infected cells). Data are presented as the mean+/−standard error of mean (s.e.m.) from one of three representative experiments.
At the times indicated, infected monolayers were dissolved in Laemmli buffer and equivalent protein concentrations were analyzed by 10% SDS-PAGE. Western blots were probed with antibody to chlamydial OmcB, and antibody to host cell actin, which served as a loading control.
HEp-2 cells were pretreated with 1 ng/ml IFN-γ for 48 hr prior to infecting with C. trachomatis B. Infected cells were then cultured in the presence of 1 ng/ml IFN for 48 hr, IFN was subsequently removed, and cells were incubated for an additional 48 hr with fresh culture medium with or without 25 µg/ml myriocin or 10 µM U18666A. At the indicated time points, inclusion development and infectivity were analyzed by immunofluorescence analysis and infectivity assays, respectively.
For ultrastructural analysis, infected HEp-2 cells were fixed in 2% paraformaldehyde/2.5% glutaraldehyde (Polysciences Inc., Warrington, PA) in 100 mM phosphate buffer, and processed as described previously [14].
|
10.1371/journal.ppat.1007709 | Nlrp3 inflammasome activation and Gasdermin D-driven pyroptosis are immunopathogenic upon gastrointestinal norovirus infection | Norovirus infection is the leading cause of food-borne gastroenteritis worldwide, being responsible for over 200,000 deaths annually. Studies with murine norovirus (MNV) showed that protective STAT1 signaling controls viral replication and pathogenesis, but the immune mechanisms that noroviruses exploit to induce pathology are elusive. Here, we show that gastrointestinal MNV infection leads to widespread IL-1β maturation in MNV-susceptible STAT1-deficient mice. MNV activates the canonical Nlrp3 inflammasome in macrophages, leading to maturation of IL-1β and to Gasdermin D (GSDMD)-dependent pyroptosis. STAT1-deficient macrophages displayed increased MAVS-mediated expression of pro-IL-1β, facilitating elevated Nlrp3-dependent release of mature IL-1β upon MNV infection. Accordingly, MNV-infected Stat1-/- mice showed Nlrp3-dependent maturation of IL-1β as well as Nlrp3-dependent pyroptosis as assessed by in vivo cleavage of GSDMD to its active N-terminal fragment. While MNV-induced diarrheic responses were not affected, Stat1-/- mice additionally lacking either Nlrp3 or GSDMD displayed lower levels of the fecal inflammatory marker Lipocalin-2 as well as delayed lethality after gastrointestinal MNV infection. Together, these results uncover new insights into the mechanisms of norovirus-induced inflammation and cell death, thereby revealing Nlrp3 inflammasome activation and ensuing GSDMD-driven pyroptosis as contributors to MNV-induced immunopathology in susceptible STAT1-deficient mice.
| Gastrointestinal norovirus infections form a serious socio-economic worldwide problem, with about 684 million cases annually worldwide and mortality occurring both in underdeveloped countries and in immunocompromised patients. Despite the urgent global need, no specific treatments are available for norovirus induced gastroenteritis, partly because innate immune responses upon gastrointestinal norovirus infection are largely unresolved. Using the murine norovirus (MNV) model we showed that MNV infection in macrophages leads to a lytic form of cell death termed pyroptosis as well as to the maturation and release of the pro-inflammatory cytokine IL-1β. Maturation of IL-1β was observed also in vivo, after gastrointestinal infection of MNV-susceptible Stat1 knockout mice. We found that these innate immune responses upon MNV infection crucially depended on activation of the Nlrp3 inflammasome leading to Gasdermin D-driven pyroptosis, and inactivating this signaling pathway delayed lethality of MNV-susceptible STAT1 knockout mice after gastrointestinal MNV infection. We thus identified Nlrp3 inflammasome activation and ensuing pyroptosis as a signaling pathway contributing to norovirus-induced immunopathology. Taken together, this study resulted in a more detailed understanding of MNV-induced inflammatory and cell death pathways and provided insights into how gastrointestinal viruses induce immunopathology.
| Human norovirus infection is the most common cause of food-borne gastroenteritis worldwide, responsible for an estimated 684 million cases per year [1]. While norovirus gastroenteritis is self-limiting in immunocompetent individuals, it can evolve to a seriously debilitating and even life-threatening infection in conditions of genetic or acquired immunosuppression [2]. Moreover, norovirus infections also account for more than 200,000 deaths annually, mostly affecting children below five years of age in developing countries [1]. Despite this significant global socio-economic burden, the cellular processes that are induced upon gastrointestinal norovirus infection are poorly understood.
Norovirus challenge studies in human volunteers showed that viral shedding lasts long after the acute vomiting and diarrhea symptoms have subsided, and also revealed that in some individuals the viral burden peaks after clinical symptoms had been resolved [3]. This poor correlation between viral titers and gastrointestinal manifestations suggests that in addition to norovirus replication, some of the host innate responses to infection may contribute to provoking pathology. Studies using the murine norovirus (MNV) model showed that Interferon (IFN)-induced STAT1-dependent responses are required for controlling viral replication and associated pathogenesis [4–6], but the deleterious innate immune mediators that contribute to norovirus-induced intestinal and systemic inflammation remain to be elucidated.
Several studies identified inflammasome activation as a crucial innate immune mechanism that protects the host from a wide variety of viral infections [7, 8]. Inflammasomes are cytosolic multi-protein complexes in which caspase-1 and -11 proteolytic activities exert important functions in innate immunity by mediating maturation of the pro-inflammatory cytokines Interleukin (IL)-1β and IL-18, as well as by initiating a lytic form of cell death termed pyroptosis through cleaving Gasdermin D (GSDMD) [9]. For instance, inflammasome activation was shown to preserve respiratory function and to prevent lethality upon Influenza infection [10–12]. Similar host protective roles for inflammasomes were reported in encephalitis caused by West Nile virus as well as in gastroenteritis induced by rotavirus [13–15]. However, the role of inflammasomes in norovirus-induced immunopathology and norovirus pathogenesis remains to be elucidated.
Here, we show that inflammasome activation has a detrimental rather than a beneficial function during gastrointestinal norovirus infection. We show that MNV activates the canonical Nlrp3 inflammasome leading to IL-1β secretion as well as GSDMD-dependent pyroptosis in macrophages. In addition, we show that these Nlrp3- and GSDMD-mediated responses do not contribute to the diarrheic manifestations of gastroenteritis but do promote MNV-induced intestinal inflammation and lethality in STAT1-deficient mice. These data show that Nlrp3 inflammasome signaling and GSDMD-driven pyroptosis contribute to immunopathology in the setting of gastrointestinal norovirus infection.
As an in vivo model for norovirus-induced gastroenteritis with concomitant systemic inflammatory responses, we infected Stat1-/- mice via oral gavage with the non-persistent MNV-1 CW1 strain [16] (hereafter referred to as MNV), which was lethal for infected Stat1-/- mice but not for their Stat1+/- littermates (S1A Fig). In order to identify putative innate immune mediators contributing to MNV-induced intestinal inflammation and lethality in Stat1-/- mice, we first characterized in vivo MNV replication kinetics. To this end, cohorts of Stat1-/- and Stat1+/- littermate mice were orally infected with MNV and tissues were harvested at one, two or three days post-infection. Quantification of viral RNA levels showed that Stat1-/- mice displayed overall significantly higher viral burdens when compared with Stat1+/- littermates over the entire gastrointestinal tract, in Peyer’s patches (PP) and mesenteric lymph nodes (MLN), as well as in the spleen and liver only at 3 days post-infection (S1B–S1J Fig). Analysis of IFNβ expression levels at a time point when systemic viral spread had occurred, indicated that MNV infection of Stat1-/- mice caused robust anti-viral type I IFN responses in all of these tissues examined, with the exception of the distal colon (S2 Fig). MNV genomes detected in the distal colon may therefore derive from non-replicating viral particles, or alternatively MNV-induced IFN responses in the distal colon may occur with different kinetics or may be less potent compared to other parts of the gastrointestinal tract. Nevertheless, the above analyses showed that Stat1-/- mice displayed intestinal, as well as systemic dissemination of MNV at three days post-infection, which was associated with anti-viral responses within the corresponding tissues.
We next investigated how these viral replication kinetics correlated with MNV-induced diarrhea and intestinal inflammation. Consistent with the peak of viral titers in these mice, we observed that gastrointestinal MNV infection caused diarrhea only in Stat1-/- mice at three days post-infection (S3A Fig). Gastrointestinal MNV infection is known to provoke only subtle inflammation in the intestine, even in highly susceptible Stat1-/- mice. For instance, a study performing detailed histological examinations could not reveal increased numbers of inflammatory cells in the intestinal lamina propria of MNV-infected Stat1-/- mice [17]. Therefore, to monitor intestinal inflammation in MNV-infected mice non-invasively over time in a sensitive manner we measured fecal Lipocalin-2 (Lcn-2) levels that were shown to correlate closely with varying degrees of inflammation restricted to the intestine [18]. While fecal Lcn-2 levels did not increase in MNV-infected Stat1+/- mice, Stat1-/- littermates displayed a significant elevation of fecal Lcn-2 levels at two and three days after oral MNV infection (S3B Fig). This observation validated fecal Lcn-2 as a sensitive marker of MNV-induced intestinal inflammation, and indicated that the kinetics of intestinal pathology correlated with previously observed viral replication and anti-viral responses. Therefore, three days post-infection was selected as an appropriate time point for screening innate immune responses potentially associated with MNV-induced intestinal and systemic pathology.
For this purpose, we quantified the expression levels of a panel of inflammatory chemokines and cytokines at selected MNV replication sites in Stat1-/- and Stat1+/- littermates that had been infected for three days with either UV-inactivated or live MNV. Statistical log-linear regression analysis identified several chemokines and cytokines that were specifically associated with MNV-induced pathology, as defined by statistically significant higher expression levels in Stat1-/- mice infected with live MNV relative to Stat1-/- mice challenged with UV-inactivated MNV, as well as relative to live MNV-infected Stat1+/- littermate controls (Fig 1A, S1 Table). The various inflammatory cytokines expressed at higher levels in live MNV-infected Stat1-/- mice indicated a complex immune response raised to MNV in these mice. For instance, live MNV-infected Stat1-/- mice displayed increased levels of IFNγ, IL-4 and IL-17A in MLN, suggestive of ongoing ILC1/Th1, ILC2/Th2 as well as ILC3/Th17 innate and/or adaptive lymphocyte responses (Fig 1A, S1 Table). Amongst the unambiguously innate immune mediators associated with severity of MNV infection in Stat1-/- mice, the cell death-associated alarmin IL-1α and the inflammasome-dependent cytokine IL-1β were prominent in both MLN and the spleen (Fig 1A and 1B). Because IL-1β is produced as a cytosolic precursor protein that needs maturation by inflammasomes for exerting biological activity, we performed additional Western blotting analyses. Consistent with inflammasome activation occurring in MNV-infected mice, these analyses showed in vivo maturation of pro-IL-1β to its active IL-1β form in live MNV-infected Stat1-/- MLN and spleen (Fig 1C and 1D). In addition, although overall IL-1β levels in liver were not significantly increased (Fig 1A and 1B), liver tissue of a subset of live MNV-infected Stat1-/- mice contained detectable levels of mature IL-1β (Fig 1E). Together, these results showed that Stat1-/- mice produced mature IL-1β three days after gastrointestinal MNV infection. In line with our observations showing that MNV-infected Stat1-/- mice displayed detectable intestinal inflammation as well as systemic viral spread only at three days post-infection, we could not detect increased IL-1β levels in MNV-infected Stat1-/- mice at earlier time points post-infection in any tissue examined (S4 Fig). These correlating cytokine, intestinal inflammation and systemic viral dissemination kinetics suggested that IL-1β could contribute to local as well as systemic MNV-induced pathologies, prompting us to further investigate the mechanisms by which MNV induces IL-1β maturation and release.
Inflammasomes are cytosolic multiprotein complexes that engage caspase-1 and/or -11 to exert both inflammatory and cytotoxic functions in response to infectious agents [7]. Since myeloid cells are major targets during MNV infection [16, 19, 20], we next turned to primary bone marrow-derived macrophages (BMDMs) as a controlled ex vivo setting to dissect the mechanisms underlying inflammasome responses to MNV infection. MNV-infected macrophages that were primed with the TLR2 agonist Pam3CSK4 secreted IL-1β into the culture supernatant starting 12–16 hours post-infection (Fig 2A). This IL-1β secretion was associated with proteolytic maturation of pro-IL-1β and caspase-1 (Fig 2B), confirming MNV-induced inflammasome activation in these cells. In contrast, unprimed MNV-infected macrophages failed to secrete IL-1β, consistent with the need for prior transcriptional upregulation of pro-IL-1β (Fig 2A). Nevertheless, these unprimed WT macrophages displayed caspase-1 processing starting 12 hours after MNV infection (Fig 2C), demonstrating that MNV itself has the capacity to trigger inflammasome activation even in the absence of sufficient amounts of pro-IL-1β that would contribute to inflammatory responses.
We next set out to dissect the composition of the MNV-activated inflammasome. Experiments in caspase-11-deficient primary macrophages showed that caspase-11 did not contribute to MNV-induced caspase-1 activation and that the latter was sufficient for mediating IL-1β maturation and secretion upon MNV infection (Fig 2D–2F). Next, MNV infections in primary macrophages lacking both caspase-1 and -11, or lacking the adaptor protein ASC demonstrated that MNV induced maturation and secretion of IL-1β in an ASC- and caspase-1-dependent manner (Fig 2G–2I). While MNV infections in Aim2-, Nlrc4- and Nlrp6-deficient primary macrophages showed that these Pattern Recognition Receptors were dispensable for MNV-induced inflammasome activation (S5A–S5D Fig), MNV did not provoke caspase-1 processing or downstream IL-1β maturation and secretion upon infecting primary or immortalized Nlrp3-/- macrophages (Fig 2J–2L, S5A–S5E Fig). Importantly, despite abrogating MNV-induced inflammasome responses, the absence of Nlrp3 did not impair MNV replication rates or MNV-induced type I IFN responses in macrophages (S6A and S6B Fig), showing that MNV-induced inflammasome activation does not interfere with the protective IFN responses that are triggered upon MNV dsRNA recognition, at least in vitro. Together, the above observations established activation of the canonical Nlrp3 inflammasome as an innate immune pathway responsible for IL-1β secretion from MNV-infected macrophages.
We next investigated the cellular mechanisms by which MNV-induced inflammasome activation mediated secretion of IL-1β. Several MNV infection studies using RAW264.7 macrophages showed that MNV provokes apoptotic cell death [21–24]. However, analyzing the kinetics of apoptotic signaling and cellular membrane permeability by simultaneous real-time monitoring of cleavage of a fluorogenic caspase-3/7 substrate and of cell-impermeable Sytox Green (SG) uptake, respectively, showed that MNV-infected primary macrophages displayed SG uptake due to membrane permeability starting from 12 hours post-infection, while caspase-3/7 activity was detected only at later stages of MNV infection (Fig 3A). Consistent with this observation arguing against an exclusive role for apoptosis in MNV-induced cell death, confocal live cell imaging of MNV-infected primary macrophages revealed cells undergoing an increase in membrane permeability followed by cellular swelling, characteristic of a necrotic cell death mode (Fig 3B, S1 Movie). Importantly, the above real-time and imaging analyses detected MNV-induced membrane permeability and cellular swelling from 12 hours post-infection, similar as observed for caspase-1 maturation and IL-1β secretion (see Fig 2A–2C). These concordant kinetics suggested a functional link between inflammasome activation and loss of membrane integrity. Inflammasomes are known to provoke a lytic cell death mode termed pyroptosis. The biochemical hallmark of inflammasome-induced pyroptosis is the proteolytic processing of Gasdermin D (GSDMD), resulting in its N-terminal p30 fragment capable of generating pores in the cellular membrane that eventually lyse the cell due to osmotic pressure build-up [25–30]. We therefore performed Western blotting analyses showing that MNV-infected BMDMs displayed GSDMD processing to its p30 fragment starting at 12 hours post-infection (Fig 3C), in line with previously observed kinetics of inflammasome activation. In contrast, proteolytic activation of caspase-3 indicative of apoptotic cell death was detected only at 24 hours after MNV infection (Fig 3C). Together, the above real-time imaging and biochemical observations suggested that apoptosis is a late-stage event after infecting primary macrophages with MNV, whereas a population of MNV-infected macrophages undergo an earlier lytic cell death associated with GSDMD processing, indicative of pyroptosis.
After observing MNV-induced GSDMD processing concordant with the appearance of lytic cell death, we next sought to validate the functional role of GSDMD in MNV-induced cell death and IL-1β release. Real-time SG uptake analyses showed that both unprimed and TLR2-primed primary Gsdmd-/- macrophages displayed a delayed membrane permeability increase upon MNV infection (Fig 3D–3E). Furthermore, in accordance with our above observations identifying Nlrp3 and caspase-1 but not caspase-11 as inflammasome components engaged by MNV (see Fig 2D–2L), Nlrp3-/- and Casp1-/- macrophages showed a delay in MNV-induced membrane permeability similar to Gsdmd-/- cells, while Casp11-/- macrophages responded like WT cells (Fig 3D–3E). Further biochemical evidence for inflammasome-dependent pyroptosis in MNV-infected macrophages was obtained by Western blotting analyses showing that MNV-induced generation of the GSDMD p30 fragment did not require caspase-11 (Fig 3F) but was dependent on the presence of Nlrp3 (Fig 3G), consistent with the opposing roles of these proteins in MNV-induced inflammasome activation.
We next evaluated the role of GSDMD-mediated pyroptosis in MNV-induced release of IL-1β from macrophages. While GSDMD deficiency did not affect maturation of caspase-1 or pro-IL-1β (Fig 3H)–consistent with its role downstream of inflammasome activation–also MNV-induced secretion of IL-1β into the culture supernatant was not diminished in Gsdmd-/- macrophages (Fig 3I). This persisting IL-1β secretion in MNV-infected Gsdmd-/- macrophages aligns with the incomplete blockade of SG uptake in these cells, and supports the existence of additional mechanisms contributing to cellular permeability during MNV infection. As IL-1β secretion triggered by canonical inflammasomes indeed was reported to involve also GSDMD-independent mechanisms [25], we investigated whether necroptosis contributed to MNV-induced IL-1β release. Necroptosis is executed by the pseudokinase MLKL, which upon phosphorylation creates nanopores in the plasma membrane leading to cellular swelling and lysis [31, 32]. This MLKL-induced membrane damage was suggested as a trigger for Nlrp3 inflammasome activation [33, 34], raising the possibility that upon MNV infection, MLKL could act upstream to activate the Nlrp3 inflammasome as well as downstream to help release mature IL-1β through necroptosis. However, real-time SG uptake measurements did not reveal delayed membrane permeability in MNV-infected Mlkl-/- macrophages (S7A and S7B Fig). In accordance, the kinetics of MNV-induced inflammasome activation and release of mature IL-1β were unchanged in Mlkl-/- macrophages (S7C–S7E Fig). Together, these results showed that MLKL was not required for MNV-induced inflammasome responses. In addition, these results dismissed a role for MLKL-mediated necroptosis as a GSDMD-independent lytic cell death contributing to IL-1β release from MNV-infected cells, leaving apoptotic cells as the remaining potential source of IL-1β secretion. Interestingly, recent studies showed that GSDMD-deficient macrophages could relay inflammasome signaling into the execution of apoptotic cell death [35, 36]. In support of this idea, real time monitoring of a fluorogenic caspase-3/7 substrate in MNV-infected BMDMs showed increased apoptotic activity in the absence of GSDMD (S8A Fig). Moreover, Gsdmd-/- BMDMs also showed earlier and stronger caspase-3 cleavage upon MNV infection compared to WT BMDMs (S8B Fig). Together, these observations showing increased apoptotic signaling in Gsdmd-/- macrophages suggest that the residual IL-1β release that is detected in the absence of GSDMD-driven pyroptosis might derive from apoptotic cell bodies that lyse through secondary necrosis.
As our above observations pointed to a dual cytokine maturation and cell death inducing role of the Nlrp3 inflammasome upon MNV infection, we next evaluated the role of Nlrp3 in MNV-susceptible Stat1-/- conditions. Interestingly, we noted pronounced IL-1β secretion from Stat1-/- macrophages without the need for prior TLR2 priming (Fig 4A). Indeed, while caspase-1 activation and downstream GSDMD p30 generation were evident in both unprimed Stat1+/- and unprimed Stat1-/- macrophages, STAT1 deficiency was additionally associated with pro-IL-1β processing leading to secretion of its mature form (Fig 4A and 4B). The observed effect of STAT1 deficiency allowing IL-1β maturation in unprimed macrophages was specific for MNV-induced Nlrp3 inflammasome activation, as unprimed Stat1-/- macrophages did not display IL-1β maturation upon triggering the Nlrp3 inflammasome with ATP, Nigericin or MSU (S9 Fig). Closer Western blotting examination of unprimed MNV-infected Stat1-/- macrophages revealed that STAT1 deficiency was associated with enhanced expression of pro-IL-1β but not of the Nlrp3 or ASC inflammasome components (Fig 4B). Since MNV-infected unprimed Stat1-/- macrophages displayed faster MNV replication rates compared to Stat1+/- cells (Fig 4C), more pronounced pro-IL-1β expression could be caused by enhanced dsRNA-induced NF-κB signaling triggered by increased viral RNA. Therefore, because MNV-induced NF-κB responses rely on dsRNA triggering of the MDA5-MAVS pathway [37], we examined the role of MAVS in MNV-induced inflammasome responses. Deleting MAVS in otherwise WT BMDMs abrogated MNV-induced IFNβ secretion as expected (S10A Fig), but did not alter MNV-induced inflammasome responses as assessed by IL-1β secretion as well as by pro-IL-1β, caspase-1 and GSDMD processing (S10B–S10D Fig), showing that MNV-induced activation of the Nlrp3 inflammasome does not rely on MAVS signaling. Interestingly however, unlike unprimed STAT1-deficient macrophages, unprimed MAVS-deficient macrophages did not display increased pro-IL-1β expression upon MNV infection and did not alter MNV replication rates in unprimed conditions (S10D and S10E Fig). These pronounced differences between the effects of deleting either STAT1 or MAVS during MNV infection of unprimed BMDMs supported the idea that enhanced MNV-induced signaling associated with increased viral replication could explain the propensity of unprimed Stat1-/- macrophages to boost pro-IL-1β expression. Indeed, abrogating MAVS-mediated signaling in Stat1-/- macrophages did not only abolish MNV-induced IFNβ secretion as expected (S11 Fig), but the resulting Stat1-/-Mavs-/- BMDMs also lost their ability to secrete IL-1β upon MNV infection in unprimed conditions (Fig 4D). Additional MAVS deletion in Stat1-/- macrophages did not alter inflammasome responses in terms of caspase-1 cleavage (Fig 4E), consistent with its dispensable role in MNV-induced inflammasome activation as observed on a WT background, but it abrogated the increase in pro-IL-1β levels observed in unprimed Stat1-/- BMDMs, resulting in the lack of IL-1β maturation in unprimed MNV-infected Stat1-/-Mavs-/- BMDMs (Fig 4E). Together, our observations in MNV-infected Stat1-/-, Mavs-/- and Stat1-/-Mavs-/- macrophages showed that STAT1 controls MNV-induced inflammasome-mediated IL-1β responses by restricting MAVS-mediated signaling towards upregulation of pro-IL-1β levels. However, while being responsible for this MNV-induced priming of the inflammasome, MAVS signaling does not participate in the Nlrp3 activation step of MNV-induced inflammasome responses.
The above ex vivo experiments indicated that STAT1 signaling prevents the pro-inflammatory effects of MNV-induced Nlrp3 inflammasome activation by suppressing expression of its pro-IL-1β substrate. As this could be one of the mechanisms by which STAT1 signaling controls MNV pathogenesis, we generated Stat1-/-Nlrp3-/- mice to investigate whether Nlrp3 inflammasome-mediated IL-1β secretion contributes to the MNV-susceptible phenotype conferred by STAT1 deficiency. Whereas both MNV-infected Stat1-/-Nlrp3-/- and Stat1-/-Nlrp3+/- macrophages produced pro-IL-1β in the absence of TLR priming, MNV only triggered caspase-1 activation, GSDMD cleavage, pro-IL-1β maturation as well as IL-1β secretion in Stat1-/-Nlrp3+/- cells (Fig 5A–5B). This demonstrated that Nlrp3 deficiency prevents the enhanced secretion of bio-active IL-1β from Stat1-/- macrophages by blocking its maturation (Fig 5B). To evaluate whether Nlrp3 was also responsible for inflammasome responses upon MNV infection in vivo, Stat1-/-Nlrp3-/- and Stat1-/-Nlrp3+/- littermates were infected with MNV through the oral route. Western blotting analysis showed that both spleen and MLN from MNV-infected Stat1-/-Nlrp3+/- mice displayed maturation of IL-1β, which was abrogated in MNV-infected Stat1-/-Nlrp3-/- organs (Fig 5C–5D). Moreover, these analyses revealed that MLN from MNV-infected Stat1-/-Nlrp3+/- mice also displayed Nlrp3-dependent GSDMD cleavage after MNV infection, indicative of pyroptosis (Fig 5D). Together, these observations demonstrated that gastrointestinal MNV infection in Stat1-/- mice elicits Nlrp3-dependent IL-1β responses in MLN and spleen, as well as Nlrp3-dependent pyroptosis in MLN based on the appearance of the N-terminal p30 fragment of GSDMD as its biochemical hallmark.
We next evaluated the pathophysiological outcome of Nlrp3-mediated inflammasome responses by assessing MNV-induced lethality in Stat1-/-Nlrp3-/- mice. Gastrointestinal MNV infections revealed that Stat1-/-Nlrp3-/- mice survived statistically significantly longer than their Stat1-/-Nlrp3+/- littermates (Fig 5E), clearly showing that abrogating Nlrp3-mediated IL-1β maturation has a beneficial effect in overall MNV-induced immunopathology in STAT1-deficient mice. Next, to examine whether GSDMD-driven pyroptosis takes part in the Nlrp3-mediated effects during MNV-induced immunopathology, we also generated Stat1-/-Gsdmd-/- mice. Interestingly, Stat1-/-Gsdmd-/- mice phenocopied the statistically significant survival advantage of Stat1-/-Nlrp3-/- mice upon gastrointestinal MNV infection when compared to Stat1-/- mice (Fig 5F). This observation indicated that also GSDMD-driven pyroptosis is crucially involved in the deleterious systemic effects of MNV contributing to lethality in Stat1-/- mice. Finally, in order to evaluate whether Nlrp3 inflammasome activation and GSDMD-driven pyroptosis also exerted deleterious effects locally in the intestinal tract, we evaluated diarrhea responses and we measured fecal Lcn-2 levels in MNV-infected Stat1-/-Nlrp3-/- and Stat1-/-Gsdmd-/- mice. These analyses showed that gastrointestinal MNV infection provoked similar degrees of diarrhea in Stat1-/-Nlrp3-/- and Stat1-/-Gsdmd-/- mice when compared with Stat1-/- mice (Fig 5G), indicating that inflammasome activation does not take part in MNV-induced diarrheic responses. In contrast, the induction of fecal Lcn-2 observed in MNV-infected Stat1-/- mice at three days post-infection was significantly diminished in both Stat1-/-Nlrp3-/- and Stat1-/-Gsdmd-/- mice (Fig 5H). Taken together, while not overtly involved in the diarrhea responses that are most relevant to norovirus gastroenteritis in immunocompetent humans, the fecal Lcn-2 results show that Nlrp3 inflammasome and GSDMD pyroptosis responses do not only take part in systemic inflammatory responses that most likely underlie MNV-induced lethality, but also contribute to local MNV-induced inflammation in the intestine. Thus, our in vivo observations identify activation of the Nlrp3 inflammasome and ensuing GSDMD-driven pyroptosis as a critical innate immune response to MNV that–in contrast to the previously described MNV-induced signaling pathways leading to protective IFN responses [37, 38]–exerts an immunopathologic role upon gastrointestinal MNV infection in Stat1-/- mice.
In this study, we reveal activation of the canonical Nlrp3 inflammasome leading to IL-1β secretion and GSDMD-dependent pyroptotic cell death as a detrimental innate immune response triggered by MNV. Given that several studies demonstrated protective roles for inflammasomes in various viral infections [10–15], our observation that the Nlrp3 inflammasome contributes to MNV-induced lethality in Stat1-/- mice is remarkable. Interestingly however, Influenza infection studies have shown that the level of host IFN responsiveness determines the physiological outcome of inflammasome activation in viral infections. While inflammasome deficiency was detrimental during Influenza infection when compared with wild-type mice [10–12], abrogating inflammasome signaling prevented Influenza-induced lethality in Tlr7-/-Mavs-/- mice that are impaired in mounting type I IFN responses to Influenza [39]. Our study showing that the Nlrp3 inflammasome contributes to lethality upon gastrointestinal MNV infection in IFN-unresponsive Stat1-/- mice is in line with this observation in Influenza-infected mice. Moreover, our observation that STAT1 signaling prevents MAVS-mediated upregulation of pro-IL-1β suggests that enhanced expression of this inflammasome substrate may be one of the reasons why inflammasome activation is detrimental rather than protective in the absence of IFN signaling.
Although enhanced pro-IL-1β expression enables Stat1-/- cells to produce more mature IL-1β upon inflammasome activation, releasing this mature IL-1β across the plasma membrane is still needed to provoke the overwhelming inflammatory responses that may contribute to lethality in Stat1-/- mice. While necroptosis was suggested to perform such a role upon infection with Influenza [40, 41], MNV did not provoke MLKL-mediated membrane permeability but rather induced GSDMD-mediated pyroptosis. Although this contrasts with previous studies showing that MNV infection caused apoptotic cell death in RAW264.7 macrophages [21–24], the latter cell line is known to lack ASC expression [42]. This implies that inflammasome activation and downstream pyroptosis induction were defective in these cells. As such, apoptotic cell death in MNV-infected RAW264.7 cells is in line with our observation that primary macrophages display Casp3/7 activity at late stages upon MNV infection, and this notion supports a balance between pyroptosis and apoptosis during MNV infection in macrophages. Accordingly, we observed that MNV infection provoked elevated Casp3/7 activity and more pronounced caspase-3 cleavage in the absence of GSDMD, suggesting that Gsdmd-/- cells can switch to apoptotic cell death upon MNV infection. This could explain our observation that MNV-infected GSDMD-deficient macrophages still released IL-1β, as secondary necrosis of apoptotic cell bodies may also contribute to IL-1β release in macrophage cultures. Despite this in vitro observation predicting that GSDMD deficiency would not be able to prevent MNV-induced inflammasome-mediated immunopathology, our in vivo experiments demonstrated that Stat1-/-Gsdmd-/- mice reproduced both the survival advantage and the diminished fecal Lcn-2 levels of Stat1-/-Nlrp3-/- mice upon gastrointestinal MNV infection. These concordant observations in Stat1-/-Gsdmd-/- and Stat1-/-Nlrp3-/- mice indicate that Nlrp3 inflammasome-induced pyroptosis represents the dominant cell death mode promoting MNV-induced immunopathology. While difficult to obtain experimental evidence explaining the discordant in vitro and in vivo observations, it is plausible that the IL-1β release deriving from post-apoptotic necrosis of Gsdmd-/- cells in vitro is efficiently prevented by timely efferocytosis of apoptotic cells in vivo. Further supporting our observations pointing to Nlrp3-mediated pyroptosis as the physiologically relevant cell death mode during gastrointestinal MNV infection, Van Winkle et al. recently showed that MNV induced a lytic cell death mode in bone-marrow derived dendritic cells as well as in the BV2 microglia cell line [20]. Moreover, the ability of different MNV strains to induce this lytic cell death correlated with their persistence in the host, as lytic cell death was accompanied with myeloid cell recruitment to the MLN and the PP that provided MNV with novel target cells [20]. Although future experiments will be required to validate this hypothesis, our observations suggest that this cell death mediated process supporting systemic persistence of acute MNV infection could involve Nlrp3-induced and GSDMD-dependent pyroptosis.
In conclusion, this work contributed to a more detailed understanding of norovirus-induced inflammatory and cell death pathways in vitro and in vivo. We identify Nlrp3 inflammasome activation and downstream GSDMD-dependent pyroptosis as an MNV-induced innate immune and cell death response. We demonstrate that Nlrp3 inflammasome activation by MNV provokes both GSDMD-dependent pyroptosis as well as secretion of mature IL-1β in macrophages. We further show that these inflammasome responses contribute to lethality in MNV-susceptible Stat1-/- mice. In addition, Stat1-/-Gsdmd-/- and Stat1-/-Nlrp3-/- mice showed diminished fecal Lcn-2 levels upon gastrointestinal MNV infection. Although the latter observations show that inflammasome responses also act locally to promote MNV-induced intestinal inflammation, diarrheic reactions upon gastrointestinal MNV infection were not decreased in Stat1-/-Gsdmd-/- or Stat1-/-Nlrp3-/- mice. This indicates that the mechanisms regulating intestinal inflammation and diarrhea in MNV-infected Stat1-/- mice are uncoupled, and suggests that the immunopathologic role of inflammasome activation in Stat1-/- mice may not be extrapolated to human norovirus gastroenteritis as occurring in immunocompetent individuals. Instead, our observations showing inflammasome responses upon systemic MNV dissemination such as occurring in Stat1-/- mice suggest that Nlrp3 inflammasome activation and GSDMD-driven pyroptosis may represent an immunopathologic response that could be more relevant in life-threatening cases of human norovirus infections such as in immunocompromised individuals [2].
All animal experiments were performed according to institutionally approved protocols according to national (Belgian Laws 14/08/1986 and 22/12/2003, Belgian Royal Decree 06/04/2010) and European (EU Directives 2010/63/EU, 86/609/EEG) animal regulations. Animal protocols were reviewed and approved by the Ethical Committee Animal Experimentation VIB site Ghent—Ghent University—Faculty of Sciences (permit number LA1400091) with approval ID 2016–030. All necessary efforts were made to minimize suffering of the animals.
Stat1−/−, ASC−/−, Casp1/11−/−, Casp1−/−, Casp11−/−, Nlrp3−/−, Nlrp6-/-, Nlrc4−/−, Aim2−/−, Gsdmd−/−, Mlkl−/− and Mavs-/- mice, either generated on C57BL/6 background or backcrossed at least ten generations to C57BL/6J background, were described previously [25, 43–52]. In all in vivo experiments as well as ex vivo primary cell experiments, controls were either in-house bred C57BL/6J mice (WT) or the +/+ or +/- littermates of the respective mice as indicated.
All animal experiments were performed according to approved protocols according to institutional, national and European guidelines. All mice used in this study were bred and housed in individually ventilated cages (IVC) in the Specific Pathogen Free facility at Ghent University, were sex- and age-matched and were fed autoclaved standard rodent feed (Ssniff, Soest, Germany) at libitum with free access to drinking water. In all experiments, up to 5 mice were housed per cage in a 12-h light-12-h dark cycle. Mice were assigned to experimental groups according to genotype and treatment.
Bone-marrow-derived macrophages (BMDMs) were differentiated in Iscove’s modified Dulbecco’s medium (IMDM) supplemented with 30% L929 cell conditioned medium, 10% heat-inactivated FBS, 1% Gibco non-essential amino acids and 1% penicillin/streptomycin at 37°C and 5% CO2. After 6 days, the differentiation medium was aspirated and the cells were scraped in cold infection medium (IMDM supplemented with 10% heat-inactivated FBS and 1% Gibco non-essential amino acids). Then, 8.5 x 105 cells per well were seeded in a 12-well plate and incubated overnight at 37°C and 5% CO2. On day 7 the MNV infection experiments were initiated, all with the MNV-1 CW1 strain [53]. Macrophages derived from immortalized myeloid progenitors (generated by Eicke Latz, Institute of Innate Immunity, Bonn, Germany; and kindly provided by Dr. Ashley Mansell, Centre for Innate Immunity and Infectious Diseases, Clayton, Australia) were grown in DMEM (Gibco) supplemented with 10% heat-inactivated FBS and 2 mM glutamine.
For MNV infection experiments, unprimed macrophages were used, or macrophages were first primed with TLR agonists. TLR2 priming in BMDMs was performed by aspirating the medium and adding 500 μl/well of culture medium (DMEM + 10% FBS) containing 500 ng/ml Pam3CSK4 for 5h. TLR4 priming of immortalized macrophage was performed by aspirating the medium and adding culture medium (DMEM + 10% FBS) containing 100 ng/ml LPS for 3h. MNV infections were performed using viral dilutions made from a stock solution stored at -80°C in cell culture medium. First, 250 μl of medium was aspirated and then either 250 μl of medium was added (mock-infected controls), 250 μl of UV-inactivated MNV was added, or 250 μl of live MNV was added, the latter two both at a multiplicity of infection (MOI) 5. Control UV-inactivated MNV was prepared by placing the virus under UV light for at least 1h. After the infection the cells were incubated for 1h at 37°C with regular shaking, after which the medium was removed, the cells were washed with PBS and 1 ml culture medium was added. The cells were then incubated for indicated time periods at 37°C and 5% CO2 before sample collection. For canonical Nlrp3 inflammasome triggering, BMDMs were stimulated with 1 μg/ml Ultrapure LPS from Salmonella Minnesota (Invivogen) for 3 hours, after which ATP (5 mM, Roche), Nigericin (20 μM, Sigma Aldrich) or Silica (0.5 mg/ml, Min—U—Sil5 US Silica) were applied for 45 min, 45 min and 6 hours, respectively, in 37°C and 5% CO2 cell culture conditions.
Cellular membrane integrity and Caspase-3/7 enzymatic activity were determined with an IncuCyte FLR imaging system (Essen Bioscience), using the non-cell-permeable SYTOX Green (SG) DNA staining agent (30 nM) (Invitrogen) and the Casp-3/7 fluorogenic substrate (1:46 v/v) (Life technologies), respectively, according to the manufacturer’s protocol. The percentage of positive cells was calculated with the IncuCyte software package. These percentages were normalized to cell confluency as well as to a 100% cell count control achieved by SG-labelling of Triton X-100 treated wells, and were corrected for the number of positive cells observed at time point 0.
Age- and sex-matched mice were infected by oral gavage with the MNV-1 CW1 strain. The infectious dose used was 106 or 107 PFU per mouse, as indicated, administered in 100 μl PBS. Survival was monitored daily. Peyer`s patches, mesenteric lymph nodes, spleen, liver, ileum and stool were collected at indicated days post-infection for homogenization and further analysis.
Tissue samples were weighed and were homogenized in 500 μl PBS, after which lysis was completed by addition of lysis buffer (20 mM Tris HCl (pH 7.4), 200 mM NaCl, 1% Nonidet P-40) and incubation for 10 minutes on ice. Full speed centrifugation for 30 minutes cleared the homogenate and supernatant was used for further analysis. Mouse cytokines in cell culture supernatants and tissue homogenates were determined by magnetic bead-based multiplex assay using Luminex technology (Bio-Rad) and type I IFN levels were analyzed using eBioscience IFNα/IFNβ Procartaplex, all according to the manufacturer’s protocols. In S5E Fig, IL-1β was detected by the BD OptEIA ELISA kit (BD biosciences) according to the manufacturer’s protocol. Cytokines from tissue homogenates were normalized to weight of tissue, while cytokines from cell culture supernatants were expressed as concentration per ml of cell culture medium. For measuring Lipocalin-2 levels, fecal pellets were weighed and were homogenized in 500μl PBS using sterile soil grinding SK38 2ml tubes (Bertin Technologies). Stool homogenates were collected in 1.5 ml eppendorf tubes and cleared upon full speed centrifugation for 30 minutes. Lipocalin-2 levels in stool supernatants were then analyzed using the mouse Lipocalin-2/NGAL duoset ELISA (R&D systems) according to the manufacturer’s instructions, and were normalized per mg of stool.
Tissue samples collected on indicated days were homogenized in 500 μl TRIsure (Bioline). For RNA isolation from primary macrophages grown in monolayer, cells were directly lysed in the culture plate by adding 700 μl of TRIsure. Cell lysate was pipetted up and down to ensure sufficient cell disruption. Total RNA isolation was performed according the manufacturer’s protocol. cDNA was synthesized using the iScript gDNA clear cDNA Synthesis kit (Biorad) and quantitative PCR was performed using Taqman gene expression assays (Life technologies). Ifnβ mRNA levels were normalized to the levels of the Tbp reference gene. To determine MNV genome copy numbers, cell culture supernatants were aspirated, after which adherent BMDMs were lysed in TRIsure. RNA was isolated and qRT-PCR was performed using the specific 5’-CCGCAGGAACGCTCAGCAG-3’ and 5’-GGCTGAATGGGGACGGCCTG-3’ primers together with the custom made Taqman probe 5’-ATGAGTGATGGCGCA-3’ [16]. Viral genome copies were quantified according to a serially diluted MNV DNA standard.
To visualize cell morphology over time, 105 BMDMs per well were seeded onto a poly-L-lysine (Sigma)-coated eight-well slide chamber (ibidi). The next day, the cells were infected with UV-inactivated or live MNV (MOI 5), followed by incubation at 37°C and 5% CO2. After 1 hour of infection, the medium was removed, cells were washed with PBS and 1 ml medium with fluorescent dye was added. To visualize membrane integrity, the non-permeable DNA stain Propidium iodide (BD Bioscience) was added to the medium (50 ng/ml). Cells were imaged every 30 minutes over a time-period of 24 hours on a Zeiss Spinning Disk microscope using a 25x objective.
Cells and culture supernatants, and tissue homogenates were incubated with cell lysis buffer (20 mM Tris HCl (pH 7.4), 200 mM NaCl, 1% Nonidet P-40) and denatured in Laemlli buffer by boiling for 10 min. Proteins were separated by SDS-PAGE electrophoresis (Thermo Scientific) after which proteins were transferred to nitrocellulose membranes (Thermo Scientific) using semi-dry (20 min) or turbo (7 min) blotting. Blocking and antibody incubation were performed in PBS or TBS supplemented with 0.05% Tween20 (vol/vol) and 3% or 5% (wt/vol) non-fat dry milk. The membranes were incubated overnight at 4°C with primary antibodies against Caspase-1 (1:1000; Adipogen), IL-1β (1:2000; GeneTex), Nlrp3 (1:1000; Adipogen), Caspase-11 (1:1000; Novus biologicals), GSDMD (1:1000)[29], ASC (1:1000; ECM Biosciences), MLKL (1:1000; Millipore), or Caspase-3 (1:1000; Cell signaling). After washing, membranes were incubated with HRP-conjugated anti-mouse, anti-rabbit or anti-rat secondary antibodies (1:5000; Jackson ImmunoResearch Laboratories, 111-035-144, 112-035-143, 112-035-143) or were incubated with the directly labeled primary antibody β-actin-HRP (1:10000; Santa Cruz) for up to 3 h. Proteins of interest were detected by the enhanced SuperSignal West Pico Chemiluminescent Substrate (Thermo Scientific).
For log-linear regression analysis of the cytokine and chemokine expressions shown in Fig 1A, a Generalized Linear Mixed Model (GLMM) (fixed model: poisson distribution, log link; random model: gamma distribution, log link) as implemented in Genstat v18 (VSN International, Hemel Hempstead, UK; www.genstat.co.uk) was fit to the data. The linear predictor vector of the values was written as follows: log(μ) = η = Xβ + Zν, where the matrix X is the design matrix for the fixed terms (i.e. genotype, treatment and tissue) and their interactions, β is their vector of regression coefficients, Z is the design matrix for the random term (i.e. subject), and ν is the corresponding vector of random effect having a gamma distribution. The significance of the regression coefficients was assessed by a t-test. Estimated mean values and their standard errors were obtained as predictions from the GLMM, formed on the scale of the response variable. Other statistics used include two-sided Student’s t-test with unequal variance and nonparametric log-rank test, both analyzed with Prism 7.0 (GraphPad Software, San Diego, CA).
|
10.1371/journal.pcbi.1005945 | Identifying human diamine sensors for death related putrescine and cadaverine molecules | Pungent chemical compounds originating from decaying tissue are strong drivers of animal behavior. Two of the best-characterized death smell components are putrescine (PUT) and cadaverine (CAD), foul-smelling molecules produced by decarboxylation of amino acids during decomposition. These volatile polyamines act as ‘necromones’, triggering avoidance or attractive responses, which are fundamental for the survival of a wide range of species. The few studies that have attempted to identify the cognate receptors for these molecules have suggested the involvement of the seven-helix trace amine-associated receptors (TAARs), localized in the olfactory epithelium. However, very little is known about the precise chemosensory receptors that sense these compounds in the majority of organisms and the molecular basis of their interactions. In this work, we have used computational strategies to characterize the binding between PUT and CAD with the TAAR6 and TAAR8 human receptors. Sequence analysis, homology modeling, docking and molecular dynamics studies suggest a tandem of negatively charged aspartates in the binding pocket of these receptors which are likely to be involved in the recognition of these small biogenic diamines.
| The distinctive dead smell comes largely from molecules like cadaverine and putrescine that are produced during decomposition of organic tissues. These volatile compounds act as powerful chemical signals important for the survival of a wide range of species. Previous studies have identified the trace amine-associated receptor 13c (or TAAR13c) in zebrafish as the cognate receptor of cadaverine in bony fishes. In this work, we employed computational strategies to disclose the human TAAR6 and TAAR8 receptors as sensors of the putrescine and cadaverine molecules. Our results indicate that several negatively charged residues in the ligand binding pocket of these receptors constitute the molecular basis for recognition of these necromones in humans.
| Olfaction is the major neurosensory function by which many species explore the chemical composition of their natural environments to locate food, avoid potentially harmful situations, recognize territory, identify members of their own group or predators, and choose a mate. Notable among the many olfactory signals is the characteristic pungent odor of a decaying cadaver. The smell of death consists of a complex mixture of volatile organic compounds [1]. Two of the most significant components of the ‘rotting flesh’ odor are putrescine (PUT) and cadaverine (CAD), early described in 1885 by the German physician Ludwig Brieger [2]. PUT and CAD are diamine products of decarboxylation of the amino acids lysine and arginine during decomposition of animal tissue. Both have short hydrocarbon chains with a primary amine group at each end. PUT has four carbon atoms (C4) in the chain between the two amines, whereas there are five carbon atoms (C5) in CAD. These molecules, characterized by a foul-smelling odor that repels most animals, could also act as an attractant for scavengers, parasites and others [3–5].
Recent studies in mouse and fish indicate that CAD activates chemosensory receptors in the olfactory epithelium, called trace amine-associated receptors (TAARs) [6–8]. TAAR genes are found in all vertebrate taxa, varying in number between species, and constitute a sensory subsystem to detect volatile molecules complementary to the canonical olfactory receptors (ORs) [9] and pheromone vomeronasal receptors (VRs) [10]. These membrane proteins generally recognize volatile amines linked to stress, social cues and predator-derived chemicals [11–13]. TAARs belong to family A of G-protein-coupled receptors (GPCR), which are characterized by the transduction of sensory signals of external origin through second messenger cascades controlled by different heterotrimeric guanine nucleotide binding proteins (G-proteins) coupled at their intracellular regions [14]. The predominant signaling pathway described for these receptors involved the Gαolf activation, increasing cAMP levels upon stimulation by trace amines [9, 15]. Thus, TAAR responses are likely mediated by coupling to the canonical odorant transduction cascade, acting on cyclic nucleotide-gated ion channels which allow Na+ and Ca2+ ions to enter into the cell, depolarizing olfactory sensory neurons (OSNs) and beginning an action potential which carries the information to the brain [16].
TAARs share a strong evolutionary relationship with biogenic amine GPCRs such as the serotoninergic (5-HTR), β-adrenergic (ADRB) dopaminergic (DRD) and histaminergic (HRH) receptors [17]. These receptors are characterized by a highly conserved molecular architecture of seven α-helical transmembrane (7-TM) segments connected to each other by three extracellular loops (3-ECL) and three intracellular loops (3-ICL) [18]. X-ray 3D structures of several aminergic GPCRs have revealed topological conserved positions in the TM helix bundle that are critical for ligand-receptor interactions [19]. Particularly, a conserved aspartic acid at position 3.32 in TM3 [number correspond to Ballesteros-Weinstein nomenclature [20]] forms a salt bridge with the positively charged nitrogen of aminergic compounds, and polar residues at positions 5.42, 5.43 and/or 5.46 in TM5 form hydrogen bond interactions with weakly acidic hydroxyl moieties of several ligands. An interesting example in this respect is the presence of two aspartates (Asp3.32 and Asp5.42) essential for the binding of histamine and other dicationic at low pH ligands to the non-chemosensory histamine receptor type-2 (HRH2) [21, 22].
Most mammalian TAARs, and some from teleosts retain the negatively charged Asp3.32, which supports its role for volatile amine recognition [12]. Among these, a small group of TAARs contain a second aspartate at position 5.42 or 5.43 (zebrafish: zTAAR13c, zTAAR13d; human: hTAAR6, hTAAR8; mice: mTAAR6, mTAAR8b; and others). One of the few studies that explored the impact of these two negative charges in the binding of ligands it was shown that CAD binds zTAAR13c via two ionic interactions between the protonated amine and Asp3.32 and Asp5.42 [23]. However, despite the theoretical and empirical importance of this finding, very little is reported in the literature for how PUT or CAD exert their effects, and the TAAR family remain largely understudied compared to other GPCR subfamilies. Following the working hypothesis of the involvement of TAARs in death-odor detection, we have investigated the molecular interactions of PUT and CAD with the hTAAR6 and hTAAR8. The results of molecular modeling and docking experiments, in addition to unrestrained microsecond-scale (μs) molecular dynamics (MD) simulations indicate that PUT and CAD fit into the binding pocket of the human TAAR6 and TAAR8, making stable interactions with Asp3.32 and Asp5.43. This finding supports the importance of the conserved tandem of negatively charged residues in the orthosteric cavity of these receptors, offering a robust modelling hypothesis for the recognition of C4 and C5 diaminated compounds. A structure-informed multiple sequence alignment of several TAARs from well-known classes of vertebrates reveals the conservation of both aspartates in at least one of either TAAR6 or TAAR8 homolog of most mammals, while being absent in amphibians, reptiles and birds.
Numerous structural studies of GPCRs have revealed a strong conservation of the 7-TM helical architecture, as well as in a number of topologically equivalent residues involved in the binding of ligands [24]. This information has been integrated in Multiple Sequence Alignments (MSAs) in order to identify functional amino acids, localize amino acid insertions and deletions or improve classification [25–27]. Fig 1 shows a structure-based MSA of representative biogenic amine receptors, including the structurally determined 5-HT1BR (PDB ID: 4IAR), ADRB2 (2RH1, 3P0G), D3R (3PBL), H1R (3RZE) and selected TAAR6, TAAR8, TAAR13c and TAAR13d sequences from different organisms (see S1 Fig for an extensive list). The sequence similarity between members of the distinct subfamilies (e.g. TAARs vs. 5-HTRs vs. ADRBs vs. DRDs vs. HRHs) is ∼30%, which is archetypal of class A GPCRs despite their high structural resemblance [28]. Nonetheless, all sequences display well-known consensus signatures GN1.50, LAxxD2.50, DR3.50Y, W4.50, P5.50, Y5.58, CWxP6.50, NP7.50xxY [18], including the ECL1 WxFG motif and the highly conserved cysteines in TM3 and ECL2 involved in a disulfide bridge for the majority of class A GPCRs [29].
The key Asp3.32, directly involved in the interaction with aminergic ligands, aligns in all sequences. In addition, a second aspartate (Asp5.42 or Asp5.43, according to the receptor type) is present on TAAR13c, TAAR13d, TAAR6 and TAAR8 sequences (Fig 1 and S1 Fig). Both positions are an integral part of the orthosteric-binding site in most aminergic receptors and are frequently involved in interactions with polar groups of substrates [19]. In the MSA of Fig 1, the Asp5.42 of the teleost fish TAAR13c and TAAR13d sequences is aligned with Asp5.43 of mammalian TAAR6 and TAAR8 by the introduction of a single gap in the MSA. The occurrence of such a gap has been described before in order to amend non-matching amino acids due to local distortions in the α-helical scaffold [25]. In this particular case, we considered that the negatively charged aspartate in TM5 might be similarly positioned to recognize chemicals of comparable size and with two positively charged groups.
Currently, there is no experimental structural data of any TAAR in complex with their cognate substrate. However, the recent breakthroughs in GPCR structure determination [30] allow us to study the molecular basis of their interactions using modeling with high quality, structurally close, templates. Here, we used a structure-based MSA (Fig 1), together with the experimentally determined three-dimensional (3D) atomic coordinates of the ADRB2 in active and inactive conformational states [31, 32], to construct molecular 3D-models of human TAAR6 and TAAR8. From a total of 400 generated models, four representative structures of the agonist bound active- (hTAAR6active-like/hTAAR8active-like) and inactive- (hTAAR6inactive-like/hTAAR8inactive-like) conformations were selected based on their stereochemical quality and subsequently refined by molecular dynamics simulations (S1 Table). In addition, for comparison purposes, computational models of zebrafish TAAR13c were developed using the same methodology (see Methods).
To a great extent, active- and inactive-like human TAARs models displayed a high similarity in the extracellular ligand-binding region (average root mean square deviation RMSD < 2.0 Å), whereas major differences were located at the cytoplasmic G protein-coupling domain. In this region, outward displacements of the TM5 (∼5.0 Å) and TM6 (∼10.0 Å) necessary for coupling the G-protein-mimetic nanobody differentiate the TAAR6active-like/TAAR8active-like from the TAAR6inactive-like/TAAR8inactive-like structures (S2 Fig). Analysis of the biogenic amine GPCRs topologically equivalent ligand-binding pocket (region comprising TMs 3–7) in the hTAAR6, hTAAR8 and zTAAR13c molecular models clearly shows a strong electronegative character (Fig 2 and S3 Fig). An exceptional cluster of six conserved Asp/Glu residues on the TMs contributed to the overall negative electrostatic potential of the binding cavity (Asp3.32, Asp5.43, Asp6.54, Asp6.58 and Glu7.36, identified in Fig 2 and S1 Fig). It has been shown that the presence of charged residues at the orthosteric binding site entrance of GPCRs serve as a floodgate to remove the water solvent shell around ligands during the process of transferring from the extracellular aqueous environment to the binding site crevice in the TM domain [33–35]. This is of particular relevance for dicationic ligands as PUT and CAD. Thus, we hypothesized that the amino acids at the extracellular entrance playing this role are Asp6.54 (hTAAR6 D277; hTAAR8 D276), Asp6.58 (hTAAR6 D281; hTAAR8 D280) or/and Glu7.36 (hTAAR6 E293; hTAAR8 E294). On the other side, we assumed that Asp3.32 (hTAAR6 D112; hTAAR8 D111) and Asp5.43 (hTAAR6 D202; hTAAR8 D201) located at the same height at the bottom of the TM helix cavity, serve as the final anchor points of PUT and CAD (see below).
PUT and CAD are chemically very similar: they are symmetrical molecules with short hydrocarbon chains (C4 & C5 carbon atoms, respectively) and two primary amine groups at each end (average length between nitrogen atoms is 6.3 and 7.4 Å, respectively) (Fig 3). These compounds are smaller than classical aminergic ligands. Thus, owing to the fact that zebrafish TAAR13c has been identified as a high-affinity receptor for the odd-chained diamines CAD (C5) and diaminoheptane (C7) [23], it is reasonable to assume that the shorter PUT and CAD could also fit in the binding pocket of human TAAR6 and TAAR8. To test this hypothesis, we conducted molecular docking experiments of PUT and CAD to the hTAAR6 and hTAAR8 (Fig 3 and S4 Fig). As depicted in Fig 3, the chosen orientations of both molecules in the TAAR6 and TAAR8 was similar to that observed in the adrenaline-activated structure of ADRB2 [36]. The main interactions involved are a double salt-bridge between PUT/CAD protonated amines and carboxylic groups of Asp3.32/Asp5.43, and hydrophobic contacts with V3.33 (hTAAR6 V113; hTAAR8 V112) and Y6.51 (hTAAR6 Y274; hTAAR8 Y273) in close proximity to the central alkyl chains of the ligands. Likewise, similar molecular poses and score energies were obtained for the zTAAR13c bound to CAD (S2 Table and S5 Fig) that, as mentioned earlier, has been experimentally demonstrated.
Unbiased 1μs MD simulations of the ligand-receptor systems were conducted in an explicit lipid bilayer environment to assess the stability of the proposed binding: hTAAR6active-like/PUT; hTAAR6active-like/CAD; hTAAR6inactive-like/PUT; hTAAR6inactive-like/CAD; hTAAR8active-like/PUT; hTAAR8active-like/CAD; hTAAR8inactive-like/PUT, hTAAR8inactive-like/CAD and compared with the zTAAR13cactive-like/CAD and zTAAR13cinactive-like/CAD binding complexes (S3 Table). For the active-like conformations, the MD systems included a receptor-specific nanobody Nb80 with G-protein-like properties [32], coupled to the intracellular part of the receptors (S2 and S6 Figs). This procedure is necessary as agonists are incapable of stabilizing the fully active conformation of the receptor in the absence of the G protein or a G-protein-mimetic nanobody [37, 38]. All MD simulations gave rise to stable trajectories and membrane-protein systems remained steady after relaxation and during the data collection steps. The root mean square deviation (RMSDbackbone < 4.0 Å) in all simulated systems demonstrates the overall structural stability of the modeled receptors. Likewise, the accuracy of the docking poses was confirmed by the small fluctuations of ligands coordinates, in particular for the active-like structures (S7 and S8 Figs). These results support the hypothesis that both natural diamines are likely to interact in a stable manner with human TAAR6 and TAAR8 in the same way as CAD to the zebrafish TAAR13c.
Fig 4 shows the computed distances between the nitrogen atom of the protonated amines of PUT/CAD and the carboxylate groups of Asp3.32/Asp5.43 in the human TAAR6 and TAAR8 along the MD trajectories. Clearly, in the inactive-like models these distances fluctuate through the simulations, revealing that PUT/CAD could spin around inside the binding pocket (Fig 4E–4H). These flip- transitions occur very rapidly (~10ns on average) and are quickly stabilized by salt-bridges with the opposite pairs of the interacting partners. Notably, this effect is not observed in the active-like models (Fig 4A–4D), probably due to the small contraction of the orthosteric cavity observed in the activated state of the receptors [39] that impedes the transition. This is reflected in the initial homology models, depicted in Fig 2, in which the distances between the carboxyl moieties of Asp3.32/Asp5.43 were ~1.0 Å smaller in the active-like conformations (average dist. 10.2 Å) with respect to the inactive ones (average dist. 11.6 Å). A similar trend was observed in the zTAAR13c/CAD complexes (S3 and S8 Figs). In all cases, the TM3-TM5 distance was further reduced during the MD trajectories, dropped below 10 Å in the active-like ligand-receptor simulated complexes (S3 Table).
Furthermore, we analyzed in the MD simulations of active- and inactive-like structures the ‘transmission switch’, comprising amino acids at positions 3.40, 5.50, and 6.44 (Fig 5 and S9 Fig). These residues located below the ligand binding cavity adopt different conformations upon binding of agonists, inverse agonists or allosteric modulators, and thus constitute a good model to study the effect of the ligands on the conformational states of the receptors [24, 38, 40, 41]. Similarly to the agonist-bound ADRB2 in complex with Gαs (Fig 5A in green), the TAAR6/TAAR8 active-like complexes (green in Fig 5B and 5C) were characterized by the inward displacement of TM5 at the highly conserved Pro5.50 (hTAAR6 P209; hTAAR8 P208), steric competition with bulky hydrophobic residues (hTAAR6 L120; hTAAR8 V119) at position 3.40 and small counterclockwise rotation of TM3 which leads to a steric exclusion with the side chain of F6.44 (hTAAR6 F267; hTAAR8 F266) and outward displacement of TM6. Conformational sampling analysis of these residues revealed higher fluctuations in the inactive-like complexes, in particular P5.50 and F6.44 (standard deviations (SD) of Cβ atoms position ≥ 1Å, Fig 5B and 5C in red/light red) with regard to the active-like complexes (SD of Cβ < 1Å, Fig 5B and 5C in green/light green). We believe this is a consequence of the disrupted interactions between PUT and CAD with Asp3.32 and Asp5.43 (Fig 4E–4H). This is in contrast to the strong binding in the active-like receptors (Fig 4A–4D), which suggest that both ligands contribute to the constriction of the binding cavity through stable ionic interactions with the Asp3.32/Asp5.43 pair, stabilizing active conformations same as agonists compounds [39] and consistent with previous observations in the zTAAR13c [7].
In addition to TAARs, the chemosensory function in vertebrates it is carried out by ORs, VRs and taste receptors (TRs) GPCR subfamilies. The number of genes and pseudogenes of these chemosensory receptors, as well as their associated sensory organs, vary enormously among species according their different living environments [42, 43]. Likewise, the TAAR gene repertoire is highly variable among vertebrate taxa [44]. Copy number of TAARs ranges over a hundred in teleosts (zebrafish), to less than ten in amphibians (clawed frog), and only a few (1 to 4) in sauropsids (zebra finch, anole lizard and chicken). The number of TAARs in synapsids is generally larger than in other four-limbed vertebrates, but also varies significantly across species, even within the same taxonomic group (see Fig 6). We searched for the tandem of aspartates in 220 identified vertebrate TAARs [44], and except for the teleosts TAAR13a, TAAR13c, TAAR13d, TAAR13e, TAAR14d and therian TAAR6 and TAAR8 sequences, no other receptor with two conserved negatively charged residues in the TM3 and TM5 helices was found in the monotreme, sauropsid or amphibian lineages.
It has been reported that the identified zTAARs could detect chemicals with two cations. In particular, CAD binds to the zTAAR13c with μM affinity [7], whereas PUT and CAD bind with different affinities to the zTAAR13d [23]. Similarly, mutation of either Asp3.32 or Asp5.42 in these receptors reduced or abolished responses to dicationic ligands. On the other hand, TAAR6 and TAAR8 homologous genes with conserved Asp3.32/Asp5.43 were found in most of placental mammals including terrestrial ungulates (hoofed animals), supraprimates (human, mouse, rat), carnivores (with a notable exception in dogs), and were absent in cetaceans (see Figs 6 and S1). Frequently, these two genes are contiguously located in chromosomal regions (16.6kb distance between hTAAR6 and hTAAR8 on human chromosome 6), which suggests they are products of genome duplication events and, consequently, could share similar ligand binding preferences. This could be consistent with our MD simulation experiments that show stable interactions of the two related diamines in both receptors. Moreover, taking into account that besides the Asp3.32/Asp5.43 pair, all other negatively charged binding pocket residues are also conserved in the TAAR6 and TAAR8 sequences (Fig 2 and S1 Fig). It is reasonable to assume that a common molecular mechanism for PUT and CAD recognition is shared by the mammalian orthologs here identified.
Death’s distinctive smell, characterized among other chemicals by the volatiles diamines PUT and CAD, constitutes an important signal related to risk avoidance, social cues and feeding behaviors which are pivotal for surviving. PUT and CAD belong to the biogenic amine group of naturally occurring compounds found in the whole animal world from bacteria to mammals, including key intracellular signaling molecules with powerful physiological effects such as histamine, serotonin, dopamine and adrenaline [45]. But unlike these well-studied neurotransmitters, the molecular basis and physiological actions of these ‘necromones’ is still largely unknown. Fortunately, there is indication that zebrafish TAAR13c constitutes a diamine sensor that manifests selectivity for odd chain diamines, including CAD. With this knowledge, we explored the sequence-structure relation of TAARs from different organisms and propose the human TAAR6 and TAAR8, and possibly their mammalian orthologs, as the cognate receptors for these compounds. This finding is supported by the analysis of structure-informed sequence alignments of close related aminergic GPCRs, revealing a conserved tandem of negatively charged aspartates in the ligand binding cavity of teleost TAAR13c and mammalian TAAR6 and TAAR8, which are likely to be involved in diamine recognition. Structural models of these receptors based on 3D structures of the ADRB2 in different conformational states, together with molecular docking and MD simulations, sustain this hypothesis, showing feasible interactions between the negatively charged aspartates Asp3.32 (zTAAR13cD112; hTAAR6 D112; hTAAR8 D111) and Asp5.42/5.43 (zTAAR13cD202; hTAAR6 D202; hTAAR8 D201) with diamine moieties of PUT and CAD. The observation that both TAAR6 and TAAR8 could bind these similar molecules is not surprising, in view of the well-known ligand promiscuity among closely related GPCRs (e.g. both adrenaline and noradrenaline display high affinity for alpha-adrenergic ADRA1 and ADRA2 receptors). Unfortunately, our theoretical approach does not allow to predict the binding affinities for these similar binders (C4 vs. C5 alkyl chain lengths), in either TAAR6 or TAAR8. However, since the interactions between Asp3.32/5.43 (-COO-) and PUT/CAD (-NH3+) were more stable in the active-like complexes, following a similar trend as that observed for the CAD binding to the zTAAR13c, we hypothesize that both ligands show a preference for the activated state of the receptors and, consequently, could behave as agonists.
Taking into account that the odor mortis constitutes a primordial class of chemical signal linked to survival, the two-aspartate signature was searched amongst TAARs of other jawed vertebrates. Teleosts (bony fishes) are characterized by a great expansion of TAAR genes (including TAAR13c and TAAR13d) related to the important roles of solubilized polyamines for chemical communication in water environments [3]. Conversely, no identifiable TM3 and TM5 negatively charged signature was found in sauropsids (birds and reptiles) or amphibian lineages, characterized by small number of TAARs, but with large numbers of vomeronasal and taste receptor repertoires [42]. This great amount of variation in chemosensory receptors within organisms, has been linked to a model of birth-and-death evolution, related to living environments [43, 46]. Thus, specific ecological conditions [47], lineage-specific specialization [48] and morphological or physiological adaptations [49] among other factors, could lead to different sensory abilities to detect the PUT and CAD polyamines in these species.
In mammals, the tertiary amine-detecting TAARs display higher rates of gene duplications, which suggest they may have played important roles in terrestrial adaptations. Likewise, the high conservation of the negatively charged Asp3.32/Asp5.43 tandem in TAAR6 and TAAR8 therian sequences seems to provide chemosensory sensitivity to diamines like PUT and CAD in most of terrestrial mammals. Nonetheless, this signature is missing in the non-terrestrial aquatic dolphins and whales, characterized in general by having small number functional chemosensory receptors [50] and in some carnivores like dog [51]. In the latter case, the notable loss of functional TAARs seems to be compensated by a strong evolution of ORs genes (> 800) which almost double the human repertoire [52]. It is known that OR-expressing neurons may also function as detectors of trace amines in the olfactory epithelium [53]. Thus, from this perspective, the rapid evolutionary diversification according to environmental adaptations makes it possible that recognition of PUT and CAD in vertebrates lacking TAAR6 and TAAR8 functional genes, could be undertaken by other chemosensory receptors which may have developed a dication binding site. In any event, these primordial class of chemical signals linked to the survival of many organisms deserve further studies. We hope this work helps provide insight into two scarcely studied human receptors with unknown pharmacology and contribute to the understanding of the mechanism of action of PUT and CAD which may be useful in pharmacological applications and other industrial purposes.
The human TAAR6 (NP_778237.1) and TAAR8 (NP_444508.1) were used as queries to search for homologues using protein-protein blast (blastp) sequence similarity searches (http://www.ncbi.nlm.nih.gov/blast). Twenty-six TAAR6 and TAAR8 mammalian orthologs (including humans) were aligned with ClustalW, using the GPCRtm substitution matrix [54] (see S1 Fig). An additional MSA was constructed with a selection of TAAR6, TAAR8, TAAR13c and TAAR13d sequences and related aminergic receptors with known 3D-structures. This MSA was manually curated in order to satisfy the structural correspondence between conserved sequence motifs in class A GPCRs, including the disulfide bridge between TM3 and ECL2 [29] and a single residue gap in TM5 [25] (see Fig 1). Approximate divergence times between species were estimated with TimeTree [55].
MODELLER v9.12 [56] was used for the construction of hTAAR6, hTAAR8 and zTAAR13c three-dimensional (3D) models using the crystal structures of the closed related ADRB2 as templates (reference MSA on Fig 1). Only non-conserved N-terminal (amino acids 1–20), C-terminal (amino acids 329–345) and ICL3 (amino acids 226–251) regions were excluded for the modeling protocol. One hundred models were generated for each receptor in the active-like (template PDB ID: 3P0G, [32]) and inactive-like conformations (template PDB ID: 2RH1, [31]) (see S2 Fig). The resulting models were evaluated stereochemically with ProSA and PROCHECK (S1 Table). The best evaluated structures were selected for further refinement of loop regions through a MD simulated annealing (SA) protocol. For this purpose, the backbone residues of the TM helices were constrained and the conformation of ECLs and ICLs were optimized in 20 simulated annealing cycles of heating up to 700 K and slowly cooling down to 300 K in successive 10 K, 100 ps steps, followed by an energy minimization with the AMBER ff99SB force field [57].
PUT and CAD were docked into the hTAAR6 and hTAAR8 models using the Molecular Operating Environment (MOE) [58]. The Site Finder application in MOE was employed to localize the binding cavities from the 3D atomic coordinates of the molecular models and 100 conformations per ligand were generated by the stochastic conformation search method. One hundred flexible docking solutions were produced by the triangle matcher algorithm into the active site of the receptor structures (additional details on S2 Table). Top-ranking solutions were visually inspected and the high score conformations in which the protonated amines form ionic interactions with Asp3.32 and Asp5.43 were energy minimized (S4 Fig). A similar protocol was employed for docking CAD to its cognate receptor zTAAR13c (S2 Table and S5 Fig). The selected binding complexes were further studied in explicit membrane MD simulations with the GROMACS MD simulation package.
MD simulations were performed using GROMACS v5.0.7. Ten molecular systems: hTAAR6active-like/PUT; hTAAR6active-like/CAD; hTAAR6inactive-like/PUT; hTAAR6inactive-like/CAD; hTAAR8active-like/PUT; hTAAR8active-like/CAD; hTAAR8inactive-like/PUT; hTAAR8inactive-like/CAD; zTAAR13cactive-like/CAD and zTAAR13cinactive-like/CAD were embedded in pre-equilibrated lipid bilayers containing 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphatidylcholine (POPC), water molecules (TIP3P) and monoatomic Na+ and Cl- ions (0.2 M), with its long axis perpendicular to the membrane interface (additional information on S3 Table). Taking into account that agonists alone are not able to preserve a fully active conformation of the receptor in the absence of the G protein [37], in our simulations, the active-like models were further stabilized by the inclusion of the G protein mimic nanobody particle towards the cytoplasmic region [32] (shown in S2 and S6 Figs). MD systems were subject to a 1000 steps of energy minimization, followed by 20.0 ns of gradual relaxation of positional restraints in protein backbone coordinates before the production phase in order to hydrate the receptor cavities and allow lipids to pack around the protein. After equilibration, 1 μs unrestrained MD trajectories were generated at a constant temperature of 300 K using separate v-rescale thermostats for the receptor, ligand, lipids and solvent molecules. A time step of 2.0 fs was used for the integration of equations of motions. All bonds and angles were kept frozen using the LINCS algorithms. Lennard-Jones interactions were computed using a cutoff of 10 Å, and the electrostatic interactions were treated using PME with the same real-space cutoff under periodic boundary conditions (PBC). The AMBER ff99SB force field was selected for the protein and the parameters described by Berger and co-workers was used for the lipids [59]. PUT and CAD parameters were obtained from the general Amber force field (GAFF) and HF/6-31G*-derived RESP atomic charges.
|
10.1371/journal.pcbi.1002079 | Generative Embedding for Model-Based Classification of fMRI Data | Decoding models, such as those underlying multivariate classification algorithms, have been increasingly used to infer cognitive or clinical brain states from measures of brain activity obtained by functional magnetic resonance imaging (fMRI). The practicality of current classifiers, however, is restricted by two major challenges. First, due to the high data dimensionality and low sample size, algorithms struggle to separate informative from uninformative features, resulting in poor generalization performance. Second, popular discriminative methods such as support vector machines (SVMs) rarely afford mechanistic interpretability. In this paper, we address these issues by proposing a novel generative-embedding approach that incorporates neurobiologically interpretable generative models into discriminative classifiers. Our approach extends previous work on trial-by-trial classification for electrophysiological recordings to subject-by-subject classification for fMRI and offers two key advantages over conventional methods: it may provide more accurate predictions by exploiting discriminative information encoded in ‘hidden’ physiological quantities such as synaptic connection strengths; and it affords mechanistic interpretability of clinical classifications. Here, we introduce generative embedding for fMRI using a combination of dynamic causal models (DCMs) and SVMs. We propose a general procedure of DCM-based generative embedding for subject-wise classification, provide a concrete implementation, and suggest good-practice guidelines for unbiased application of generative embedding in the context of fMRI. We illustrate the utility of our approach by a clinical example in which we classify moderately aphasic patients and healthy controls using a DCM of thalamo-temporal regions during speech processing. Generative embedding achieves a near-perfect balanced classification accuracy of 98% and significantly outperforms conventional activation-based and correlation-based methods. This example demonstrates how disease states can be detected with very high accuracy and, at the same time, be interpreted mechanistically in terms of abnormalities in connectivity. We envisage that future applications of generative embedding may provide crucial advances in dissecting spectrum disorders into physiologically more well-defined subgroups.
| Neurological and psychiatric spectrum disorders are typically defined in terms of particular symptom sets, despite increasing evidence that the same symptom may be caused by very different pathologies. Pathophysiological classification and effective treatment of such disorders will increasingly require a mechanistic understanding of inter-individual differences and clinical tools for making accurate diagnostic inference in individual patients. Previous classification studies have shown that functional magnetic resonance imaging (fMRI) can be used to differentiate between healthy controls and neurological or psychiatric patients. However, these studies are typically based on descriptive patterns and indirect measures of neural activity, and they rarely afford mechanistic insights into the underlying condition. In this paper, we address this challenge by proposing a classification approach that rests on a model of brain function and exploits the rich discriminative information encoded in directed interregional connection strengths. Based on an fMRI dataset acquired from moderately aphasic patients and healthy controls, we illustrate that our approach enables more accurate classification and deeper mechanistic insights about disease processes than conventional classification methods.
| Recent years have seen a substantial increase in the use of functional neuroimaging data for investigating healthy brain function and detecting abnormalities. The most popular type of analysis is statistical parametric mapping (SPM), a mass-univariate encoding model of fMRI data in which the statistical relationship between experimental (or clinical) variables and haemodynamic measurements of neural activity is examined independently for every voxel in the brain [1]. While this approach has led to many insights about functional abnormalities in psychiatric and neurological disorders, it suffers from two limitations. First, since univariate models are insensitive to spatially distributed patterns of neural activity, they may fail to detect subtle, distributed differences between patients and healthy controls that are not expressed as local peaks or clusters of activity [2]. Second, while encoding models such as SPM are excellent for describing regional differences in brain activity across clinical groups, they are less well suited for clinical decision making, where the challenge is to predict the disease state of an individual subject from measured brain activity [3]–[5].
An alternative approach is provided by multivariate decoding methods, in particular classification algorithms. Unlike mass-univariate encoding models, these methods predict an experimental variable (e.g., a trial-specific condition or subject-specific disease state) from the activity pattern across voxels (see [6]–[10] for reviews). Using multivariate decoding models instead of mass-univariate encoding models has interesting potential for clinical practice, particularly for diseases that are difficult to diagnose. Consequently, much work is currently being invested in constructing classifiers that can predict the diagnosis of individual subjects from structural or functional brain data [11], [3], [12], [13], [4], [14]–[16]. Historically, these efforts date back to positron emission tomography (PET) studies in the early 1990s [8]. Today, attempts of using multivariate classifiers for subject-by-subject diagnosis largely focus on MRI and fMRI data [11], [3], [12], [17].
Despite their increasing popularity, two challenges critically limit the practical applicability of current classification methods for functional neuroimaging data. First, classifying subjects directly in voxel space is often a prohibitively difficult task. This is because functional neuroimaging datasets (i) typically exhibit a low signal-to-noise ratio, (ii) are obtained in an extremely high-dimensional measurement space (a conventional fMRI scan contains more than 100,000 voxels), and (iii) are characterized by a striking mismatch between the large number of voxels and the small number of available subjects. As a result, even the most carefully designed algorithms have great difficulties in reliably finding jointly informative voxels while ignoring uninformative sources of noise. Popular strategies include: preselecting voxels based on an anatomical mask [18], or a separate functional localizer [20], [21]; spatial subsampling [22]; finding informative voxels using univariate models [3], [11], [12] or locally multivariate searchlight methods [23], [24]; and unsupervised dimensionality reduction [4], [25]. Other recently proposed strategies attempt to account for the inherent spatial structure of the feature space [23], [26], [27] or use voxel-wise models to infer a particular stimulus identity [28]–[30]. Finally, those submissions that performed best in the Pittsburgh Brain Activity Interpretation Competition (PBAIC 2007) highlighted the utility of kernel ridge regression [31] and relevance vector regression [31], [32]. The common assumption underlying all of these approaches is that interesting variations of the data with regard to the class variable are confined to a manifold that populates a latent space of much lower dimensionality than the measurement space.
The second challenge for classification methods concerns the interpretation of their results. Most classification studies to date draw conclusions from overall prediction accuracies [33], [11], the spatial deployment of informative voxels [19], [34], [18], [35]–[39], the temporal evolution of discriminative information [40], [37], [41], [42], [26], or patterns of undirected regional correlations [43]. These approaches may support discriminative decisions, but they are blind to the neuronal mechanisms (such as effective connectivity or synaptic plasticity) that underlie discriminability of brain or disease states. In other words: while some conventional classification studies have achieved impressive diagnostic accuracy [14], their results have not improved our mechanistic understanding of disease processes.
Generative embedding for model-based classification may provide a solution to the challenges outlined above. It is based on the idea that both the performance and interpretability of conventional approaches could be improved by taking into account available prior knowledge about the process generating the observed data (see [44] for an overview). (The term generative embedding is sometimes used to denote a particular model-induced feature space, or so-called generative score space, in which case the associated line of research is said to be concerned with generative embeddings. Here, we will use the term in singular form to denote the process of using a generative model to project the data into a generative score space, rather than using the term to denote the space itself.) Generative embedding rests on two components: a generative model for principled selection of mechanistically interpretable features and a discriminative method for classification (see Figure 1).
Generative models have proven powerful in explaining how observed data are caused by the underlying (neuronal) system. Unlike their discriminative counterparts, generative models capture the joint probability of the observed data and the class labels, governed by a set of parameters of a postulated generative process. One example in neuroimaging is dynamic causal modelling (DCM) [45]. DCM enables statistical inference on physiological quantities that are not directly observable with current methods, such as directed interregional coupling strengths and their modulation, e.g., by synaptic gating [46]. (We use the term DCM to refer both to a specific dynamic causal model and to dynamic causal modelling as a method.) From a pathophysiological perspective, disturbances of synaptic plasticity and neuromodulation are at the heart of psychiatric spectrum diseases such as schizophrenia [47] or depression [48]. It is therefore likely that classification of disease states could benefit from exploiting estimates of these quantities. While DCM is a natural (and presently the only) candidate for obtaining model-based estimates of synaptic plasticity (cf. [46], [49]), the most widely used approach to classification relies on discriminative methods, such as support vector machines (SVMs) [50], [51]. Together, DCM and SVM methods thus represent natural building blocks for classification of disease states.
Generative embedding represents a special case of using generative kernels for classification, such as the P-kernel [52] or the Fisher kernel [53]. Generative kernels have been fruitfully exploited in a range of applications [54]–[66] and define an active area of research [67]–[70]. In the special case of generative embedding, a generative kernel is used to construct a generative score space. This is a model-based feature space in which the original observations have been replaced by statistical representations that potentially yield better class separability when fed into a discriminative classifier. Thus, an unsupervised embedding step is followed by a supervised classification step. In previous work, we suggested a concrete implementation of this approach for the trial-by-trial classification of electrophysiological recordings [61]. In this paper, we propose a DCM-based generative-embedding approach for subject-by-subject classification of fMRI data, demonstrate its performance using a clinical data set, and highlight potential methodological pitfalls (and how to avoid them).
DCM [45] views the brain as a nonlinear dynamical system of interconnected neuronal populations whose directed connection strengths are modulated by external perturbations (i.e., experimental conditions) or endogenous activity. Here, we will use DCM to replace high-dimensional fMRI time series by a low-dimensional vector of parameter estimates. The discriminative part of our approach will be based on an SVM with a linear kernel. This algorithm learns to discriminate between two groups of subjects by estimating a separating hyperplane in their feature space. Since this paper brings together techniques from different statistical domains that tend to be used by different communities, we have tried to adopt a tutorial-like style and introduce basic concepts of either approach in the Methods section.
Generative embedding for fMRI may offer three substantial advantages over conventional classification methods. First, because the approach aims to fuse the strengths of generative models with those of discriminative methods, it may outperform conventional voxel-based schemes, especially in those cases where crucial discriminative information is encoded in ‘hidden’ quantities such as directed (synaptic) connection strengths. Second, the construction of the feature space is governed and constrained by a biologically motivated systems model. As a result, feature weights can be interpreted mechanistically in the context of this model. Incidentally, the curse of dimensionality faced by many conventional feature-extraction methods may turn into a blessing when using generative embedding: the higher the temporal and spatial resolution of the fMRI data, the more precise the estimation of the parameters of the generative model, leading to better discriminability. Third, our approach can be used to compare alternative generative model architectures in situations where evidence-based approaches, such as Bayesian model selection, are not applicable. We will deal with these three points in more detail in the Discussion.
The remainder of this paper is structured as follows. First, we summarize the general ideas of generative embedding and the specific generative and discriminative components used here, i.e., DCM and SVM. We then inspect different procedures of how generative embedding could be implemented practically while distinguishing between approaches with and without bias. Third, we illustrate the utility of our approach, using empirical data obtained during speech processing in healthy volunteers and patients with moderate aphasia. These data have been explored in a previous study, in which DCM and Bayesian model selection (BMS) were applied to investigate the effective connectivity among cortical areas activated by intelligible speech [71]. In a subsequent study, we extended this analysis to patients with aphasia (Schofield et al., in preparation). In the present paper, we ask whether subject-specific directed connection strengths among cortical regions involved in speech processing contain sufficiently rich discriminative information to enable accurate predictions of the diagnostic category (healthy or aphasic) of a previously unseen individual. In brief, we found that (i) generative embedding yielded a near-perfect classification accuracy, (ii) significantly outperformed conventional ‘gold standard’ activation-based and correlation-based classification schemes, and (iii) afforded a novel mechanistic interpretation of the differences between aphasic patients and healthy controls during processing of speech and speech-like sounds.
The study was approved by the local research ethics committee at UCL, and all participants gave informed consent.
Most methods for classification attempt to find a linear function that separates examples as accurately as possible in a space of features (e.g., voxel-wise measurements). Such discriminative classification methods differ from generative methods in two ways. First, rather than trying to estimate the joint density of observations and class labels, which is not needed for classification, or trying to estimate class-conditional probability densities, which can be difficult, discriminative classifiers directly model the class an example belongs to. Second, many discriminative methods do not operate on examples themselves but are based on the similarity between any two examples, expressed as the inner product between their feature vectors. This provides an elegant way of transforming a linear classifier into a more powerful nonlinear one. (Note that the term discriminative methods is used here to collectively describe the class of learning algorithms that find a discriminant function for mapping an example onto a class label , typically without invoking probability theory. This is in contrast to discriminative models, which model the conditional probability , and generative models, which first model the full joint probability and then derive .)
The most popular classification algorithm of the above kind is the -norm soft-margin support vector machine (SVM) [50], [51], [72], [73]. The only way in which examples enter an SVM is in terms of an inner product . This product can be replaced by the evaluation of a kernel function , which implicitly computes the inner product between the examples in a new feature space, .
The -norm SVM is a natural choice when the goal is maximal prediction accuracy. However, it usually leads to a dense solution (as opposed to a sparse solution) in which almost all features are used for classification. This is suboptimal when one wishes to understand which model parameters contribute most to distinguishing groups, which will be the focus in the Section ‘Interpretation of the feature space.’ In this case, an SVM that enforces feature sparsity may be more useful. One simple way of inducing sparsity is to penalize the number of non-zero coefficients by using an -regularizer. Unlike other regularizers, the -norm (also known as the counting norm) reduces the feature-selection bias inherent in unbounded regularizers such as the - or -norm. The computational cost of optimizing an -SVM objective function is prohibitive, because the number of subsets of items which are of size is exponential in . We therefore replace the -norm by a capped -regularizer which has very similar properties [74]. One way of solving the resulting optimization problem is to use a bilinear programming approach [75]. Here, we use a more efficient difference-of-convex-functions algorithm (Ong & Thi, under review).
In summary, we will use two types of SVM. For the purpose of classification (Section ‘Classification’), we aim to maximize the potential for highly accurate predictions by using an -norm SVM. For the purpose of feature selection and interpretation (Section ‘Interpretation of the feature space’), we will focus on feature sparsity by using an approximation to an -norm SVM, which will highlight those DCM parameters jointly deemed most informative in distinguishing between groups.
Most current applications of classification algorithms in neuroimaging begin by embedding the measured recordings of each subject in a -dimensional Euclidean space . In fMRI, for example, a subject can be represented by a vector of features, each of which corresponds to the signal measured in a particular voxel at a particular point in time. This approach makes it possible to use any learning algorithm that expects vectorial input, such as an SVM; but it ignores the spatio-temporal structure of the data as well as the process that generated them. This limitation has motivated the search for kernel methods that provide a more natural way of measuring the similarity between the functional datasets of two subjects, for example by incorporating prior knowledge about how the data were generated, which has led to the idea of generative kernels, as described below.
Generative kernels are functions that define a similarity metric for observed examples using a generative model. In the case of a dynamic causal model (DCM), for example, the observed time series are modelled by a system of parameterized differential equations with Gaussian observation noise. Generative embedding defines a generative kernel by transferring the models into a vectorial feature space in which an appropriate similarity metric is defined (see Figure 1). This feature space, which we will refer to as a generative score space, embodies a model-guided dimensionality reduction of the observed data. The kernel defined in this space could be a simple inner product of feature vectors, or it could be based on any other higher-order function, as long as it is positive definite [76]. In conclusion, model-based classification via generative embedding is a hybrid generative-discriminative approach: it merges the explanatory abilities of generative models with the classification power of discriminative methods.
The specific implementation for fMRI data proposed in this paper consists of four conceptual steps which are summarized in Figure 1 and described in the following subsections. First, a mapping is designed that projects an example from data space onto a multivariate probability distribution in a parametric family . In our case, we use the fMRI data from each subject to estimate the posterior density of the parameters of a DCM (Sections ‘DCM for fMRI’ and ‘Model inversion’). Second, a probability kernel is constructed that represents a similarity measure between two inverted DCMs. Here, we use a simple linear kernel on the maximum a posteriori (MAP) estimates of the model parameters (Sections ‘Strategies for unbiased model specification and inversion’ and ‘Kernel construction’). Third, this kernel is used for training and testing a discriminative classifier (Section ‘Classification’). Here, we employ a linear SVM to distinguish between patients and healthy controls. Fourth, the constructed feature space can be investigated to find out which model parameters jointly contributed most to distinguishing the two groups (Section ‘Interpretation of the feature space’). We will conclude with an example in which we distinguish between patients with moderate aphasia and healthy controls (Sections ‘Experimental design, data acquisition, and preprocessing,’ ‘Implementation of generative embedding,’ and ‘Comparative analyses’).
DCM regards the brain as a nonlinear dynamic system of interconnected nodes, and an experiment as a designed perturbation of the system's dynamics [45]. Its goal is to provide a mechanistic model for explaining experimental measures of brain activity. While the mathematical formulation of DCMs varies across measurement types, common mechanisms modelled by all DCMs include synaptic connection strengths and experimentally induced modulation thereof [46], [77]–[80]. Generally, DCMs strive for neurobiological interpretability of their parameters; this is one core feature distinguishing them from alternative approaches, such as multivariate autoregressive models [81] which characterize inter-regional connectivity in a phenomenological fashion.
DCMs consist of two hierarchical layers [82]. The first layer is a neuronal model of the dynamics of interacting neuronal populations in the context of experimental perturbations. Critically, its parameters are neurobiologically interpretable, representing, for example, synaptic weights and their context-specific modulation; electrophysiological DCMs describe even more fine-grained processes such as spike-frequency adaptation or conduction delays. Experimental manipulations enter the model in two different ways: they can elicit responses through direct influences on specific regions (e.g., sensory inputs), or they can modulate the strength of coupling among regions (e.g., task demands or learning). The second layer of a DCM is a biophysically motivated forward model that describes how a given neuronal state translates into a measurement. Depending on the measurement modality, this can be a set of nonlinear differential equations (as for fMRI [83]) or a simple linear equation (as for EEG [84]). While the forward model plays a critical role in model inversion, it is the parameters of the neuronal model that are typically of primary scientific interest.
In this paper, we will use the classical bilinear DCM for fMRI [45] as implemented in the software package SPM8/DCM10,(1)(2)where represents the neuronal state vector at time , is a matrix of endogenous connection strengths, represents the additive change of these connection strengths induced by modulatory input , and denotes the strengths of direct (driving) inputs. These neuronal parameters are rate constants with units .
The haemodynamic forward model is given by the function , a nonlinear operator that links a neuronal state to a predicted blood oxygen level dependent (BOLD) signal via changes in vasodilation, blood flow, blood volume, and deoxyhaemoglobin content (see [83] for details). This forward model has haemodynamic parameters and Gaussian measurement error . The haemodynamic parameters primarily serve to account for variations in neurovascular coupling across regions and subjects and are typically not of primary scientific interest. In addition, the haemodynamic parameters exhibit strong inter-dependencies and thus high posterior covariances and low precision [83], which makes it difficult to establish the distinct contribution afforded by each parameter. For these reasons, the model-induced feature spaces in this paper will be based exclusively on the neuronal parameters .
In summary, DCM provides a mechanistic model for explaining measured time series of brain activity as the outcome of hidden dynamics in an interconnected network of neuronal populations and its experimentally induced perturbations. Inverting such a model (see next section) means to infer the posterior distribution of the parameters of both the neuronal and the forward model from observed responses of a specific subject. Its mechanistic interpretability and applicability to single-subject data makes DCM an attractive candidate for generative embedding of fMRI data.
Bayesian inversion of a given dynamic causal model defines a map that projects a given example (i.e., data from a single subject) onto a multivariate probability distribution in a parametric family . The model architecture specifies the neuronal populations (regions) of interest, experimentally controlled inputs , synaptic connections, and a prior distribution over the parameters . Given the model and subject-specific data , model inversion proceeds in an unsupervised and subject-by-subject fashion, i.e., in ignorance of the subject label that will later be used in the context of classification. (The literature on DCM has adopted the convention of denoting the hidden states by and the data by . Here, in order to keep the notation consistent with the literature on classification, we use for the data and for the labels. A distinct symbol for the hidden states is not required here.) DCM uses a fully Bayesian approach to parameter estimation, with empirical priors for the haemodynamic parameters and conservative shrinkage priors for the coupling parameters [85], [45]. Combining the prior density over the parameters with the likelihood function yields the posterior density . This inversion can be carried out efficiently by maximizing a variational free-energy bound to the log model evidence, , under Gaussian assumptions about the posterior (the Laplace assumption; see [86] for details). Given parameters, model inversion thus yields a subject-specific probability density that can be fully described in terms of a vector of posterior means and a covariance matrix .
Model specification and selection is an important theme in DCM [87]. In this paper we are not concerned with the question of which of several alternative DCMs may be optimal for explaining the data or for classifying subjects; these issues can be addressed using Bayesian evidence methods [88], [89] or by applying cross-validation to the classifications suggested by each of the models, respectively (see [61] for an example). However, an important issue is that model specification cannot be treated in isolation from its subsequent use for classification. Specifically, some procedures for selecting time series can lead to biased estimation of classification accuracy. In the next section, we therefore provide a detailed assessment of different strategies for time series selection in DCM-based generative embedding and highlight those procedures which safeguard against obtaining optimistic estimates of classification performance.
For conventional fMRI classification procedures, good-practice guidelines have been suggested for avoiding an optimistic bias in assessing classification performance [8], [10]. Generally, to obtain an unbiased estimate of generalization accuracy, a classifier must be applied to test data that have not been used during training. In generative embedding, this principle implies that the specification of the generative model cannot be treated in isolation from its use for classification. In this section, we structure different strategies in terms of a decision tree and evaluate the degree of bias they invoke (see Figure 2).
The first distinction is based on whether the regions of interest (ROIs) underlying the DCM are defined anatomically or functionally. When ROIs are defined exclusively on the basis of anatomical masks (Figure 2a), the selection of voxels is independent of the functional data. Using time series from these regions, the model is inverted separately for each subject. Thus, given subjects, a single initial model-specification step is followed by subject-wise model inversions. The resulting parameter estimates can be safely submitted to a cross-validation procedure to obtain an unbiased estimate of classification performance.
Whenever functional contrasts have played a role in defining ROIs, subsequent classification may no longer be unbiased. This is because a functional contrast introduces statistics of the data into voxel selection, which usually generates a bias. In this case, we ask whether contrasts are defined in an across-subjects or a between-groups fashion. In the case of an across-subjects contrast (which does not take into account group membership), one might be tempted to follow the same logic as in the case of anatomical ROI definitions: a single across-subjects contrast, computed for all subjects, guides the selection of voxels, and the resulting DCM is inverted separately for each subject (Figure 2b). Unfortunately, this procedure is problematic. When using the resulting parameter estimates in a leave-one-out cross-validation scheme, in every repetition the features would be based on a model with regions determined by a group contrast that was based on the data from all subjects, including the left-out test subject. This means that training the classifier would no longer be independent of the test data, which violates the independence assumption underlying cross-validation, a situation referred to as peeking [10]. In consequence, the resulting generalization estimate may exhibit an optimistic bias. To avoid this bias, model specification must be integrated into cross-validation (Figure 2c). Specifically, in each fold, we leave out one subject as a test subject and compute an across-subjects group contrast from the remaining subjects. The resulting choice of voxels is then used for specifying time series in each subject and the resulting model is inverted separately for each subject, including the left-out test subject. This procedure is repeated times, each time leaving out a different subject. In total, the model will be inverted times. In this way, within each cross-validation fold, the selection of voxels is exclusively based on the training data, and no peeking is involved. This is the strategy adopted for the dataset analysed in this paper, as detailed in the Section ‘Implementation of generative embedding’.
When functional contrasts are not defined across all subjects but between groups, the effect of peeking may become particularly severe. Using a between-groups contrast to define regions of interest on the basis of all available data, and using these regions to invert the model for each subject (Figure 2d) would introduce information about group membership into the process of voxel selection. Thus, feature selection for both training and test data would be influenced by both the data and the label of the left-out test subject. One way of decreasing the resulting bias is to integrate model specification into cross-validation (Figure 2e). In this procedure, the between-groups contrast is computed separately for each training set (i.e., based on subjects), and the resulting regions are used to invert the model for the test subject. This means that the class label of the test subject is no longer involved in selecting features for the test subject. However, the test label continues to influence the features of the training set, since these are based on contrasts defined for a group that included the test subject. This bias can only be removed by adopting the same laborious procedure as with across-subjects contrasts: by using a between-groups contrast involving subjects, inverting the resulting model separately for each subject, and repeating this procedure times (Figure 2f). This procedure guarantees that neither the training procedure nor the features selected for the test subject were influenced by the data or the label of the test subject.
In summary, the above analysis shows that there are three practical strategies for the implementation of generative embedding that yield an unbiased cross-validated accuracy estimate. If regions are defined anatomically, the model is inverted separately for each subject, and the resulting parameter estimates can be safely used in cross-validation (Figure 2a). Otherwise, if regions are defined by a functional contrast, both the definition of ROIs and model inversion for all subjects need to be carried out separately for each cross-validation fold (Figure 2c,f).
Given a set of inverted subject-specific generative models, the kernel defines the similarity metric under which these models are assessed within a discriminative classifier. In generative embedding, the choice of an appropriate kernel depends on the definition of the generative score space. A straightforward way to create a Euclidean vector space from an inverted DCM is to consider the posterior means or maximum a posteriori (MAP) estimates of model parameters of interest (e.g., parameters encoding synaptic connection strengths). More formally, we can define a mapping that extracts a subset of MAP estimates from the posterior distribution . This simple -dimensional vector space expresses discriminative information encoded in the connection strengths between regions, as opposed to activity levels within these regions. Alternatively, one could also incorporate elements of the posterior covariance matrix into the vector space. This would be beneficial if class differences were revealed by the precision with which connection strengths can be estimated from the data.
Once a generative score space has been created, any conventional kernel can be used to compare two inverted models. The simplest one is the linear kernel , representing the inner product between two vectors and . Nonlinear kernels, such as quadratic, polynomial or radial basis function kernels, transform the generative score space, which makes it possible to consider quadratic (or higher-order) class boundaries and therefore account for possible interactions between features. Nonlinear kernels, however, have several disadvantages for generative embedding. As the complexity of the kernel increases, so does the risk of overfitting. Furthermore, feature weights are easiest to interpret in relation to the underlying model when they do not undergo further transformation; then, the contribution of a particular feature (i.e., model parameter) to the success of the classifier can be understood as the degree to which the neuronal mechanism represented by that parameter aids classification. A simple linear kernel will therefore be our preferred choice.
In summary, in this paper, we define a mapping from a subject-specific posterior distribution of model parameters to a feature vector . We then use a linear kernel for this model-based feature space. Together, these two steps define a probability kernel that represents a similarity metric between two inverted models and allows for mechanistic interpretations of how group membership of different subjects is encoded by spatiotemporal fMRI data.
While a kernel describes how two subjects can be compared using a generative model of their fMRI data, it does not specify how such a comparison could be used for making predictions. This gap is filled by discriminative classification methods. As described in the Section ‘Combining generative models and discriminative methods’, a natural choice is the -norm soft-margin support vector machine (SVM), which currently represents the most widely used kernel method for classification [72].
An estimate of classification performance with minimal variance can be obtained by leave-one-out cross-validation. In each fold, the classifier is trained on subjects and tested on the left-out one. Using the training set only, the SVM can be fine-tuned by carrying out a simple line search over the regularization hyperparameter (Eqn. 1), a procedure known as nested cross-validation [90], [91].
There are many ways of assessing the generalization performance of a classifier. Here, we are primarily interested in the balanced accuracy, that is, the mean accuracy obtained on either class,(3)where , , , and represent the number of true positives, false positives, true negatives, and false negatives, respectively [92]. The balanced accuracy represents the arithmetic mean between sensitivity and specificity. If the classifier performs equally well on either class, it reduces to the ordinary accuracy (i.e., the ratio of correct predictions to all predictions). If, however, the classifier has taken advantage of an imbalanced dataset, then the ordinary accuracy will be inflated, whereas the balanced accuracy will drop to chance (50%), as desired. The balanced accuracy thus removes the bias from estimates of generalizability that may arise in the presence of imbalanced datasets. A probability interval can be computed by considering the convolution of two Beta-distributed random variables that correspond to the true accuracies on positive and negative examples, respectively. A p-value can then be obtained by computing the posterior probability of the accuracy being below chance [92].
Most classification algorithms can not only be used for making predictions and obtaining an estimate of their generalization error; they can also be used to quantify how much each feature has contributed to classification performance. Such feature weights can sometimes be of greater interest than the classification accuracy itself. In the case of a generative score space, as defined above, each feature is associated with a neurobiologically interpretable model parameter. Provided there are no complex transformations of feature weights (see above), they can be interpreted in the context of the underlying model.
As described in the Section ‘Combining generative models and discriminative methods’, the -norm soft-margin SVM is a natural choice when the goal is maximal prediction accuracy. However, its solution usually implies that almost all features are used for classification. This is suboptimal when one wishes to understand which model parameters, and thus mechanisms, contribute most to distinguishing groups. Therefore, for the purposes of interpreting the model-induced feature space, we use an -regularizer. This approach allows us to characterize the feature space by counting how often a particular feature has been selected in leave-one-out cross-validation.
In order to illustrate the utility of generative embedding for fMRI, we used data from two groups of participants (patients with moderate aphasia vs. healthy controls) engaged in a simple speech-processing task. The conventional SPM and DCM analyses of these data are published elsewhere; we refer to [71] and Schofield et al. (in preparation) for detailed descriptions of all experimental procedures.
The two groups of subjects consisted of 26 right-handed healthy participants with normal hearing, English as their first language, and no history of neurological disease (12 female; mean age 54.1 years; range 26–72 years); and 11 patients diagnosed with moderate aphasia due to stroke (1 female; mean age 66.1; range 45–90 years). The patients' aphasia profile was characterized using the Comprehensive Aphasia Test [93]. As a group, they had scores in the aphasic range for: spoken and written word comprehension (single word and sentence level), single word repetition and object naming. It is important to emphasize that the lesions did not affect any of the temporal regions which we included in our model described below (see Schofield et al., in preparation, for detailed information on lesion localization).
Subjects were presented with two types of auditory stimulus: (i) normal speech; and (ii) time-reversed speech, which is unintelligible but retains both speaker identity and the spectral complexity of normal speech. Subjects were given an incidental task, to make a gender judgment on each auditory stimulus, which they indicated with a button press.
Functional T2*-weighted echo-planar images (EPI) with BOLD contrast were acquired using a Siemens Sonata 1.5 T scanner (in-plane resolution 3 mm×3 mm; slice thickness 2 mm; inter-slice gap 1 mm; TR 3.15 s). In total, 122 volumes were recorded in each of 4 consecutive sessions. In addition, a T1-weighted anatomical image was acquired. Following realignment and unwarping of the functional images, the mean functional image of each subject was coregistered to its high-resolution structural image. This image was spatially normalized to standard Montreal Neurological Institute (MNI152) space, and the resulting deformation field was applied to the functional data. These data were then spatially smoothed using an isotropic Gaussian kernel (FWHM 8 mm). In previous work, these data have been analysed using a conventional general linear model (GLM) and DCM; the results are described in Schofield et al. (in preparation). Here, we re-examined the dataset using the procedure shown in Figure 2c, as described in detail in the next subsection.
We compared the performance of generative embedding to a range of alternative approaches. To begin with, we examined several conventional activation-based classification schemes. The first method was based on a feature space composed of all voxels within the predefined anatomical masks used for guiding the specification of the DCMs. As above, we used a linear SVM, and all training sets were balanced by oversampling. We will refer to this approach as anatomical feature selection.
The second method, in contrast to the first one, was not only based on the same classifier as in generative embedding but also used exactly the same voxels. Specifically, voxels were selected on the basis of the same ‘all auditory events’ contrast as above, which is a common approach to defining a voxel-based feature space in subject-by-subject classification [11], [12], [10]. In every cross-validation fold, only those voxels entered the classifier that survived a t-test (, uncorrected) in the current set of subjects. Training sets were balanced by oversampling. We will refer to this method as contrast feature selection.
The third activation-based method employed a locally multivariate ‘searchlight’ strategy for feature selection. Specifically, in each cross-validation fold, a searchlight sphere (radius 4 mm) was passed across all voxels contained in the anatomical masks described above [23]. Using the training set only, a nested leave-one-out cross-validation scheme was used to estimate the generalization performance of each sphere using a linear SVM with a fixed regularization hyperparameter (). Next, all spheres with an accuracy greater than 75% were used to form the feature space for the current outer cross-validation fold, which corresponds to selecting all voxels whose local neighbourhoods allowed for a significant discrimination between patients and healthy controls at . Both outer and inner training sets were balanced by oversampling. We will refer to this method as searchlight feature selection. To illustrate the location of the most informative voxels, we carried out an additional searchlight analysis, based on the entire dataset as opposed to a subset of size , and used the results to generate a discriminative map (see Figure S1 in the Supplementary Material).
The fourth conventional method was based on a principal component analysis (PCA) to reduce the dimensionality of the feature space constructed from all voxels in the anatomical masks described above. Unlike generative embedding, PCA-based dimensionality reduction finds a linear manifold in the data without a mechanistic view of how those data might have been generated. We sorted all principal components in decreasing order of explained variance. By retaining the 22 top components, the resulting dimensionality matched the dimensionality of the feature space used in generative embedding.
In addition to the above activation-based methods, we compared generative embedding to several approaches based on undirected regional correlations. We began by averaging the activity within each region of interest to obtain region-specific representative time series. We then computed pairwise correlation coefficients to obtain a 15-dimensional feature space of functional connectivity. Next, instead of computing spatial averages, we summarized the activity within each region in terms of the first eigenvariate. Thus, in this approach, the exact same data was used to estimate functional connectivity as was used by DCM to infer effective connectivity. Finally, as suggested in [43], we created yet another feature space by transforming the correlation coefficients on eigenvariates into z-scores using the Fisher transformation [96].
In addition to conventional activation- and correlation-based approaches, we also investigated the dependence of generative embedding on the structure of the underlying model. Specifically, we repeated our original analysis on the basis of three alternative models. For the first model, we constructed a feedforward system by depriving the original model of all feedback and interhemispheric connections (Figure 5a); while this model could still, in principle, explain neuronal dynamics throughout the system of interest, it was neurobiologically less plausible. For the second and third model, we kept all connections from the original model but modelled either only the left hemisphere (Figure 5b) or only the right hemisphere (Figure 5c).
In summary, we compared the primary approach proposed in this paper to 4 conventional activation-based methods, 3 conventional correlation-based methods, and 3 generative-embedding analyses using reduced and biologically less plausible models.
The classification performance of generative embedding was evaluated using the procedure described in Figure 2c. This procedure was compared to several conventional activation-based and correlation-based approaches. As an additional control, generative embedding was carried out on the basis of three biologically ill-informed models. In all cases, a leave-one-subject-out cross-validation scheme was used to obtain the posterior distribution of the balanced accuracy [92] as well as smooth estimates of the underlying receiver-operating characteristic (ROC) and precision-recall (PC) curves [97]. Results are presented in Table 2 and Figure 6.
The strongest classification performance was obtained when using generative embedding with the full model shown in Figure 3. The approach correctly associated 36 out of 37 subjects with their true disease state, corresponding to a balanced accuracy of 98%. Regarding conventional activation-based methods, classification based on anatomical feature selection did not perform significantly above chance (balanced accuracy 62%, p≈0.089). Contrast feature selection (75%, p≈0.003), searchlight feature selection (73%, p≈0.006), and PCA-based dimensionality reduction (80%, p<0.001) did perform significantly above chance; however, all methods were outperformed significantly by generative embedding (p≈0.003, p≈0.001, and p≈0.045, paired-sample Wald test). Regarding conventional correlation-based methods, all three approaches performed above chance, whether based on correlations amongst the means (70%, p≈0.011), correlations amongst eigenvariates (83%, p<0.001), or z-transformed correlations amongst eigenvariates (74%, p≈0.002). Critically, however, all were significantly outperformed by generative embedding (p<0.001, p≈0.045, p≈0.006). Regarding generative embedding itself, when replacing the original model shown in Figure 3 by a biologically less plausible feedforward model (Figure 5a) or by a model that captured the left hemisphere only (Figure 5b), we observed a significant decrease in performance, from 98% down to 77% (p≈0.002) and 81% (p≈0.008), respectively, although both accuracies remained significantly above chance (p≈0.001 and p<0.001). By contrast, when modelling the right hemisphere only (Figure 5c), performance dropped to a level indistinguishable from chance (59.3%, p≈0.134).
In order to provide a better intuition as to how the generative model shown in Figure 3 created a score space in which examples were much better separated than in the original voxel-based feature space, we produced two scatter plots of the data (see Figure 7). The first plot is based on the peak voxels of the three most discriminative clusters among all regions of interest, evaluated by a searchlight classification analysis. The second plot, by analogy, is based on the three most discriminative model parameters, as measured by two-sample t-tests in the (normalized) generative score space. This illustration shows how the voxel-based projection (left) leads to classes that still overlap considerably, whereas the model-based projection (right) provides an almost perfectly linear separation of patients and controls.
The low dimensionality of the model-based feature space makes it possible to visualize each example in a radial coordinate system, where each axis corresponds to a particular model parameter (see Figure 8). When using parameters that represent directed connection strengths, this form of visualization is reminiscent of the notion of ‘connectional fingerprints’ for characterizing individual cortical regions [98]. In our case, there is no immediately obvious visual difference in fingerprints between aphasic patients and healthy controls. On the contrary, the plot gives an impression of the large variability across subjects and suggests that differences might be subtle and possibly jointly encoded in multiple parameters.
One way of characterizing the discriminative information encoded in individual model parameters more directly is to estimate class-conditional univariate feature densities (see Figure 9). Here, densities were estimated in a nonparametric way using a Gaussian kernel with an automatically selected bandwidth, making no assumptions about the distributions other than smoothness [99]. While most densities are heavily overlapping, a two-sample t-test revealed significant group differences in four model parameters (denoted by stars in Figure 9): the self-connection of L.HG (parameter 4); the influence that L.HG exerts over L.PT (parameter 5); the influence R.MGB on R.PT (parameter 13); and the influence of R.HG on L.HG (parameter 14). All of these were significant at the 0.001 level while no other parameter survived p = 0.05. An extended plot of all bivariate feature distributions, illustrating how well any two features jointly discriminated between patients and healthy controls, can be found in the Supplementary Material (Figure S2).
In order to better understand which DCM parameters jointly enabled the distinction between patients and controls, we examined the frequency with which features were selected in leave-one-out cross-validation when using an SVM with a sparsity-inducing regularizer [75], [74] (see Figure 10). We found that the classifier favoured a highly consistent and sparse set of 9 (out of 22) model parameters; the corresponding synaptic connections are highlighted in red in Figure 3. Notably, this 9-dimensional feature space, when used with the original -norm SVM, yielded the same balanced classification accuracy (98%) as the full 22-dimensional feature space, despite discarding more than two thirds of its dimensions.
The above representation disclosed interesting potential mechanisms. For example, discriminative parameters were restricted to cortico-cortical and thalamo-cortical connection strengths, whereas parameters representing auditory inputs to thalamic nuclei did not contribute to the distinction between patients and healthy controls. This finding implies that, as one would expect, low-level processing of auditory stimuli, from brain stem to thalamus, is unimpaired in aphasia and that processing deficiencies are restricted to thalamo-cortical and cortico-cortical networks. In particular, the discriminative connections included the top-down connections from planum temporale to Heschl's gyrus bilaterally; the importance of these connections had also been highlighted by the previous univariate analyses of group-wise DCM parameters in the study by Schofield et al. (in preparation). Furthermore, all of the connections from the right to the left hemisphere were informative for group membership, but none of the connections in the reverse direction. This pattern is interesting given the known specialization of the left hemisphere in language and speech processing and previous findings that language-relevant information is transferred from the right hemisphere to the left, but not vice versa [100]. It implies that aphasia leads to specific changes in connectivity, even in non-lesioned parts of the language network, with a particular effect on inter-hemispheric transfer of speech information. This specificity is seen even more clearly when considering only those three parameters which were selected 100% of the time (i.e., in all cross-validation folds) and are thus particularly meaningful for classification (bold red arrows in Figure 3). The associated connections mediate information transfer from the right to the left hemisphere and converge on the left planum temporale which is a critical structure for processing of language and speech [101], [102].
In summary, all selected features represented connectivity parameters (as opposed to stimulus input), their selection was both sparse and highly consistent across resampling repetitions, and their combination was sufficient to afford the same classification accuracy as the full feature set.
Generative embedding for subject-by-subject classification provides three potential advantages over conventional voxel-based methods. The first advantage is that it combines the explanatory strengths of generative models with the classification power of discriminative methods. Thus, in contrast to purely discriminative or purely generative methods, generative embedding is a hybrid approach. It fuses a feature space that captures both the data and their generative process with a classifier that finds the maximum-margin boundary for class separation. Intuitively, this exploits the idea that differences in the generative process between two examples (observations) might provide optimal discriminative information required to enable accurate predictions. In the case of DCM for fMRI, this rationale should pay off whenever the directed connection strengths between brain regions contain more information about a disease state than regional activations or undirected correlations. Indeed, this is what we found in our analyses (cf. Figure 6). Using a DCM-informed data representation might prove particularly relevant in psychiatric disorders, such as schizophrenia or depression, where aberrant effective connectivity and synaptic plasticity are central to the disease process [48], [47].
The second advantage of generative embedding for fMRI is that it enables an intuitive and mechanistic interpretation of features and their weights, an important property not afforded by most conventional classification methods [103], [104]. By using parameter estimates from a mechanistically interpretable model for constructing a feature space, the subsequent classification no longer yields ‘black box’ results but allows one to assess the relative importance of different mechanisms for distinguishing groups (e.g., whether or not synaptic plasticity alters the strengths of certain connections in a particular context). Put differently, generative embedding embodies a shift in perspective: rather than representing sequential data in terms of high-dimensional and potentially highly complex trajectories, we are viewing the data in terms of the coefficients of a well-behaved model of system dynamics. Again, this may be of particular importance for clinical applications, as discussed in more detail below. It is also interesting to note that models like DCM, when used in the context of generative embedding, turn the curse of dimensionality faced by conventional classification methods into a blessing: the higher the spatial and temporal resolution of the underlying fMRI data, the more precise the resulting DCM parameter estimates; this in turn should lead to more accurate predictions.
The third advantage provided by generative embedding is related to model comparison. For any given dataset, there is an infinite number of possible dynamic causal models, differing in the number and location of nodes, in connectivity structure, and in their parameterization (e.g., priors). Competing models can be compared using Bayesian model selection (BMS) [89], [83], [86], [88], where the best model is the one with the highest model evidence, that is, the highest probability of the data given the model [105]. BMS is a generic approach to distinguish between different models that is grounded in Bayesian probability theory and, when group-specific mechanisms can be mapped onto distinct models, represents a powerful technique for model-based classification in itself. However, there are two scenarios in which BMS is problematic and where classification based on generative embedding may represent a useful alternative [61]. First, BMS requires the data to be identical for all competing models. Thus, in the case of current implementations of DCM for fMRI, BMS enables dynamic model selection (concerning the parameterization and mathematical form of the model equations) but not structural model selection (concerning which regions or nodes should be included in the model). Second, BMS is limited when different groups cannot be mapped onto different model structures, for example when the differences in neuronal mechanisms operate at a finer conceptual scale than can be represented within the chosen modelling framework. In this case, discriminability of subjects may be afforded by differences in (combinations of) parameter estimates under the same model structure (see [106] for a recent example).
In both these scenarios, the approach proposed in this paper may provide a solution, in that the unsupervised creation of a generative score space can be viewed as a method for biologically informed feature extraction, and the performance of the classifier reflects how much class information is encoded in the model parameters. This view enables a form of model comparison in which the best model is the one that enables the highest classification accuracy. This procedure can be applied even when (i) the underlying data (e.g., the chosen regional fMRI time series) are different, or when (ii) the difference between two models lies exclusively in the pattern of parameter estimates. In this paper, we have illustrated both ideas: structural model selection to decide between a full model and two reduced models that disregard one hemisphere; and dynamic model selection to distinguish between different groups of subjects under the same model structure.
In summary, BMS evaluates the goodness of a model with regard to its generalizability for explaining the data, whereas generative embedding evaluates a model in relation to an external criterion, i.e., how well it allows for inference on group membership of any given subject. This difference is important as it highlights that the concept of a ‘good’ model can be based on fundamentally different aspects, and one could imagine scenarios where BMS and generative embedding arrive at opposing results. If, for example, discriminability of groups relies on a small subspace of the data, then one model (which provides a good accuracy-complexity trade-off for most of the data except that subspace) may have higher evidence, but another model that describes this subspace particularly well but is generally worse for the rest of the data may result in better classification performance (cf. our discussion in [106]). We will examine the relation and complementary nature of BMS and generative-embedding approaches in future work.
As discussed in this paper, there are three valid strategies for the implementation of generative embedding in fMRI that allow for an unbiased estimate of classification accuracy (Figure 2). If regions (and thus time series) are defined anatomically, the model is inverted separately for each subject, and the resulting parameter estimates can be safely used in cross-validation. If regions are defined by a functional contrast, both time series selection and model inversion for all subjects need to be carried out separately for each cross-validation fold. These procedures clearly have higher computational demands than conventional classification techniques, but the subject-wise nature of model inversion means that generative embedding for fMRI can exploit methods for distributed computing and can thus be implemented even for larger numbers of subjects.
In order to demonstrate the utility of generative embedding for fMRI, we acquired and analysed a dataset consisting of 11 aphasic patients and 26 healthy controls. During the experiment, participants were listening to a series of speech and speech-like stimuli. In an initial analysis (Schofield et al., in preparation), we designed a dynamic causal model to explain observed activations in 6 auditory regions of interest. Here, we extended this analysis by examining whether patients and healthy controls could be distinguished on the basis of differences in subject-specific generative models. Specifically, we trained and tested a linear support vector machine on subject-wise estimates of connection strengths. This approach delivered two sets of results.
First, we found strong evidence in favour of the hypothesis that aphasic patients and healthy controls may be distinguished on the basis of differences in the parameters of a generative model alone. Generative embedding did not only yield a near-perfect balanced classification accuracy (98%). It also significantly outperformed conventional activation-based methods, whether they were based on anatomical (62%), contrast (75%), searchlight feature selection (73%), or on a PCA-based dimensionality reduction (80%). Similarly, our approach outperformed conventional correlation-based methods, whether they were based on regional means (70%) or regional eigenvariates (83% and 74%). Furthermore, it is interesting to observe that group separability was reduced considerably when using a less plausible feedforward model (77%). Finally, performance decreased significantly when modelling only the left hemisphere (81%), and it dropped to chance when considering the right hemisphere by itself (60%), which is precisely what one would expect under the view that the left hemisphere is predominantly, but not exclusively, implicated in language processing. Taken together, our findings provide strong support for the central idea of this paper: that critical differences between groups of subjects may be expressed in a highly nonlinear manifold which remains inaccessible by methods relying on activations or undirected correlations, but which can be unlocked by the nonlinear transformation embodied by an appropriate generative model.
Second, since features correspond to model parameters, our approach allowed us to characterize a subset of features (Figure 10) that can be interpreted in the context of the underlying model (Figure 3). This subset showed four remarkable properties. (i) Discriminative parameters were restricted to cortico-cortical and thalamo-cortical connection strengths. On the contrary, parameters representing auditory inputs to thalamic nuclei did not contribute to the distinction between patients and healthy controls. (ii) We observed a high degree of stability across resampling folds. That is, the same 9 (out of 22) features were selected on almost every repetition. (iii) The set of discriminative parameters was found to be sparse, not just within repetitions (which is enforced by the underlying regularizer) but also across repetitions (which is not enforced by the regularizer; see Figure S3 in the Supplementary Material). At the same time, the set was considerably larger than what would be expected from univariate feature-wise t-tests (Figure 9). (iv) The sparse set of discriminative parameters proved sufficient to yield the same balanced classification accuracy (98%) as the full set. These results are consistent with the notion that a distinct mechanism, and thus few parameters, are sufficient to explain differences in processing of speech and speech-like sounds between aphasic patients and healthy controls. In particular, all of the connections from the right to the left hemisphere were informative with regard to group membership, but none of the connections in the reverse direction. This asymmetry resonates with previous findings that language-relevant information is transferred from the right hemisphere to the left, but not vice versa [100], and suggests that in aphasia connectivity changes in non-lesioned parts of the language network have particularly pronounced effects on inter-hemispheric transfer of speech information from the (non-dominant) right hemisphere to the (dominant) left hemisphere.
It is worthwhile briefly commenting on how the present findings relate to those of the original DCM study by Schofield et al. (in preparation). Two crucial differences are that the previous study (i) applied Bayesian model averaging to a set of 512 models and (ii) statistically examined each of the resulting average connection strengths in a univariate fashion. They found group differences for most connections, highlighting in particular the top-down connections from planum temporale to primary auditory cortex bilaterally. In our multivariate analysis, these two connections were also amongst the most informative ones for distinguishing patients from controls (Figure 3). Schofield et al. also found group differences for interhemispheric connection strengths between left and right Heschl's gyrus, but their univariate approach did not demonstrate any asymmetries. In contrast, our multivariate approach yielded a sparser set of discriminative connections, highlighting the asymmetries of interhemispheric connections described above (Figure 3).
The example described in this paper was chosen to illustrate the implementation and use of generative embedding for fMRI. It is important to emphasize that this example does not represent the sort of clinical application that we envisage in the long term. Clearly, there are few diagnostic problems when dealing with aphasia and usually a clinical examination by the physician is sufficient. However, this example is useful for demonstrating the utility of generative embedding since the diagnostic status of each subject is known without doubt and the networks involved in speech processing are well characterized. In the future, we hope that our approach will be useful for addressing clinical problems of high practical relevance, for instance for dissecting psychiatric spectrum disorders, such as schizophrenia, into physiologically defined subgroups [47], or for predicting the response of individual patients to specific drugs. While an increasing number of studies have tried to describe neurobiological markers for psychiatric disorders [22], [107], [108], [3], [109], [110], [14], [15], we argue that these studies should be complemented by model-based approaches for inferring biologically plausible mechanisms. Such approaches will be useful in two domains of application: they can be used to decide between competing hypotheses (as in traditional applications of DCM and BMS); and they can harvest the potentially rich discriminative information encoded in aspects of synaptic plasticity or neuromodulation to build classifiers that distinguish between different subtypes of a psychiatric disorder on a physiological basis (using techniques such as generative embedding).
In the case of the illustrative dataset analysed in this paper, generative embedding yielded stronger classification performance than conventional methods, whether they were based on activations or regional correlations. One might think that this superior ability to accurately classify individual subjects determines the clinical value of the approach. Instead, we wish to argue that its clinical value will ultimately depend on whether patients that share the same symptoms can be differentially treated according to the underlying pathophysiology of the disorder. Generative embedding, using biologically plausible and mechanistically interpretable models, may prove critical in establishing diagnostic classification schemes that distinguish between pathophysiologically distinct subtypes of spectrum diseases and allow for predicting individualized behavioural and pharmacological therapy.
|
10.1371/journal.pgen.1007427 | Paired Immunoglobulin-like Type 2 Receptor Alpha G78R variant alters ligand binding and confers protection to Alzheimer's disease | Paired Immunoglobulin-like Type 2 Receptor Alpha (PILRA) is a cell surface inhibitory receptor that recognizes specific O-glycosylated proteins and is expressed on various innate immune cell types including microglia. We show here that a common missense variant (G78R, rs1859788) of PILRA is the likely causal allele for the confirmed Alzheimer’s disease risk locus at 7q21 (rs1476679). The G78R variant alters the interaction of residues essential for sialic acid engagement, resulting in >50% reduced binding for several PILRA ligands including a novel ligand, complement component 4A, and herpes simplex virus 1 (HSV-1) glycoprotein B. PILRA is an entry receptor for HSV-1 via glycoprotein B, and macrophages derived from R78 homozygous donors showed significantly decreased levels of HSV-1 infection at several multiplicities of infection compared to homozygous G78 macrophages. We propose that PILRA G78R protects individuals from Alzheimer’s disease risk via reduced inhibitory signaling in microglia and reduced microglial infection during HSV-1 recurrence.
| Alzheimer’s disease (AD) is a devastating neurodegenerative disorder resulting from a complex interaction of environmental and genetic risk factors. Despite considerable progress in defining the genetic component of AD risk, understanding the biology of common variant associations is a challenge. We find that PILRA G78R (rs1859788) is the likely AD risk variant from the 7q21 locus (rs1476679) and PILRA G78R reduces PILRA endogenous and exogenous ligand binding. Our study highlights a new immune signaling axis in AD and suggests a role for exogenous ligands (HSV-1). Further, we have identified that reduced function of a negative regulator of microglia and neutrophils is protective from AD risk, providing a new candidate therapeutic target.
| Alzheimer’s disease (AD) results from a complex interaction of environmental and genetic risk factors [1]. Proposed environmental risk factors include a history of head trauma [2–4] and infection [5–7]. In recent years, large-scale genome-wide association studies (GWAS) and family-based studies have made considerable progress in defining the genetic component of AD risk, and >30 AD risk loci have been identified [8,9,18–20,10–17].
A key role for microglial/monocyte biology in modulating risk of AD has emerged from analysis of the loci associated with AD risk. Rare variants of TREM2, a microglial activating receptor that signals through DAP12, greatly increase AD risk [11,14]. Beyond TREM2, a number of the putative causal genes mapping to AD risk loci encode microglial/monocyte receptors (complement receptor 1, CD33), myeloid lineage transcription factors (SPI1), and other proteins highly expressed in microglia (including ABI3, PLGC2, INPP5D, and PICALM).
The index variant for the Alzheimer’s disease risk locus at 7q21 is rs1476679 (meta P value = 5.6 x 10−10, odds ratio = 0.91)[15]. In addition to reduced disease risk, the C allele of rs1476679 has been associated with age of onset [21] and lower odds of pathologic AD (plaques and tangles) in the ROSMAP study [22]. In the 1000 Genomes project CEU population (phase 3 data), there were 6 variants in strong linkage disequilibrium (r2>0.9) with rs1476679 (S1 Table). None of the 6 variants were predicted to alter regulatory motifs that might influence gene expression (Regulome DBscore ≤ 4), but one variant (rs1859788) encoded a missense allele (G78R, ggg to agg transition) in Paired Immunoglobulin-like Type 2 Receptor Alpha (PILRA) protein. Using a cohort of 1,357 samples of European ancestry whole genome-sequenced to 30X average read-depth (Illumina), we confirmed the strong linkage between rs1476679 (in ZCWPW1 intron) and rs1859788 (G78R PILRA variant) (S1 Table).
We hypothesized that PILRA G78R was the functional variant that accounts for the observed protection from AD risk. As expected from the strong linkage disequilibrium (LD) between PILRA G78R and rs1476679 (Fig 1A), conditional analysis demonstrated that the 2 variants were indistinguishable for AD risk in individuals of European ancestry. In a cohort of 8060 European ancestry samples (a subset of samples described in 19), individuals homozygous for R78 (OR = 0.72) and heterozygous (OR = 0.89) for R78 were protected from AD risk relative to G78 homozygotes. We note that the allele frequency of PILRA G78R varies considerably in world populations. Indeed PILRA R78 is the minor allele in populations of African (10%) and European descent (38%) but is the major allele (65%) in East Asian populations [23].
The index variant in the 7q21 locus (rs1476679) has been associated with expression levels of multiple genes in the region, including PILRB [24,25]. However, the strongest cis-eQTL in the region is a haplotype tagged by rs6955367 which has a low coefficient of determination to rs1476679 (r2 = 0.085, D’ = 0.982) in Europeans and is more strongly associated with expression in whole blood of multiple genes in the region (PILRB, STAG3L5, PMS2P1, MEPCE) compared to rs1476679 [26]. Since the PILRB eQTL P value for rs1476679 is not significant (P = 0.31) after conditioning rs6955367 (S2 Table) in whole blood, we conclude that rs1476679 and rs1859788 are not significant causal eQTLs in the 7q21 region and the observed relationship of these SNPs with PILRB expression is due to the weakly correlated variant rs6955367 (S1 Fig). Of interest, the G allele of rs6955367 (increased expression of PILRB) is linked to rs7803454 (r2 = 0.83), a variant associated with increased risk of age-related macular degeneration and suggests the presence of independent effects in the PILRA/PILRB region [27].
Paired activating/inhibitory receptors are common in the immune system, with the activating receptor typically having weaker affinity than the inhibitory receptor toward the ligands. PILRA and PILRB are type I transmembrane proteins with highly similar extracellular domains that bind certain O-glycosylated proteins [28–31], but they differ in their intracellular signaling domains [32–34]. PILRA contains an immunoreceptor tyrosine-based inhibitory motif (ITIM), while PILRB signals through interaction with DAP12, which contains an immunoreceptor tyrosine-based activation motif (ITAM). Analysis of PILRA knockout mice suggests that PILRA is a negative regulator of inflammation in myeloid cells [35–37], with knockout macrophages showing increased production of cytokines (IL6, IL-1b, KC, MCP-1) in addition to increased infiltration of monocytes and neutrophil via altered integrin signaling.
PILRA is known to bind both endogenous (including COLEC12, NPDC1, CLEC4G, and PIANP) and exogenous ligands (HSV-1 glycoprotein B (gB)) [30,31,36,38]. Because the G78R (R78 (AD protective)) variant resides close to the sialic acid-binding pocket of PILRA, we tested whether the glycine (uncharged, short amino acid) to arginine (basic, long side chain amino acid) substitution might interfere with PILRA ligand-binding activity. All non-human PILRA sequences, as well as all PILRB sequences, encode glycine at this position. We also generated amino acid point variants in and around the sialic acid-binding pocket of PILRA. A residue conserved among PILR proteins and related SIGLEC receptors, R126 in PILRA, is well known to be essential for sialic acid interaction [29,31,38] and so was not further studied here. Based on their location in the crystal structure, evolutionary conservation [31], and involvement in binding HSV-1 gB [38], amino acids R72 and F76 were predicted to be important for ligand binding and were substituted to alanine as positive controls for loss-of-function [31]. In addition, S80, a residue outside of the sialic acid-binding pocket was substituted to glycine. The R72A, F76A, and S80G mutations have not been detected in human populations (dbSNP v147).
To study receptor-ligand binding, 293T cells were transfected with G78 (AD risk) PILRA or variants, and then incubated with purified NPDC1-mIgG2a protein (Fig 1B), followed by flow cytometry to detect PILRA and the NPDC1 fusion protein. Among known PILRA ligands, NPDC1 is expressed in the central nervous system and binds with high affinity to PILRA [31]. Expression of the PILRA variants on the transfected 293T cells was comparable to or greater than G78 (AD risk) PILRA (S2 Fig). G78 (AD risk) PILRA binding to NPDC1 was considered 100%. Both R72A and F76A mutations severely impaired NPDC1 binding (~20% of G78, p-value < 0.0001). The R78 (AD protective) variant also showed significantly reduced ligand binding (~35% of G78, p < 0.0005), while the G80 mutant was the least affected (~60% of G78, p < 0.0001) (Fig 1C and S3A and S3B Fig).
To further test the hypothesis that the AD protective PILRA R78 variant impacts ligand binding, NPDC1 or alternative PILRA ligands HSV-1 gB and PIANP were expressed on the cell surface of 293T cells, and the binding of purified PILRA protein variants was measured by flow cytometry. PILRA R78 showed reduced binding to the various ligands in these assays as compared to G78 (Fig 1D to 1G and S4A to S4G Fig). These data confirmed that the R78 variant impairs ligand-binding activity of PILRA.
A peptide motif for PILRA interaction has been established (Fig 2A) that includes an O-glycosylated threonine, an invariant proline at the +1 position, and additional prolines at the -1 or -2 and +3 or +4 positions [31,38]. Of note, PILRA is capable of binding murine CD99 and human NPCD1 (both contain the consensus motif), but not human CD99 or murine NPCD1 (both lack the consensus motif), suggesting divergence between human and mouse in the range of endogenous ligands bound by PILRA [31].
We sought to identify novel endogenous PILRA ligands by searching for human proteins with either the PTPXP, PTPXXP, PXTPXP or PXTPXXP motif. A total of 1540 human proteins carry at least 1 of these putative PILRA-binding motifs (S3 Table). Narrowing the search, we considered proteins with the motif that have previously been shown to be O-glycosylated in human cerebral spinal fluid [39], and measured the binding of these proteins to PILRA variants. By flow cytometry, complement component 4A (C4A) bound to G78 (AD risk) PILRA in a manner comparable to NPDC1, while APLP1 and SORCS1 showed relatively little interaction with PILRA (Fig 2B and S5A and S5B Fig). We further demonstrated that the PILRA R78 (AD protective) variant has reduced binding for C4A (Fig 2C and S5C Fig). We did not test C4B, but its putative PILRA-binding motif is identical to that of C4A.
To understand the conformational changes that might occur in the PILRA sialic acid-binding pocket during receptor-ligand interactions in the presence of G78 (AD risk) or R78 (AD-protective) variants, we evaluated available experimental crystal structures (Fig 3A to 3C) [38,40]. Structures of G78 (AD risk) PILRA reveal a monomeric extracellular domain with a single V-set Ig-like β-sandwich fold that binds O-glycan ligands (Fig 3B and 3C) [38]. By analogy to a molecular clamp, the sialic acid-binding site in PILRA undergoes a large structural rearrangement from an “open” to a “closed” conformation upon binding its peptide and sugar ligands simultaneously (Fig 3A to 3C). The essential R126 side-chain engages the carboxyl group of sialic acid (SA) directly in a strong salt bridge (Fig 3C). The CC’ loop which contains F76 and G78 undergoes a large conformational change where F76 translates ~15 Å to participate in key interactions with the peptide of the ligand and abut the Q140 side-chain of PILRA (Fig 3B and 3C). In this ligand-bound “closed” conformation of PILRA, Q140 helps to position R126 precisely for its interaction with SA (Fig 3C).
Notably, in the structure of R78 (AD protective) PILRA crystallized in the absence of any ligand [40], the long side-chain of R78 is observed to hydrogen bond with Q140 directly (Fig 3A). This unique R78-Q140 interaction has three important consequences: 1) it sterically hinders F76 from obtaining a ligand-bound “closed” conformation, 2) it affects the ability of R126 to interact with the carboxyl group of SA by altering the R126-Q140 interactions observed in G78 (AD risk) PILRA and, 3) it likely alters CC’ loop dynamics, (Fig 3B to 3C). Overall, the structure of the R78 (AD protective) variant shows that this single side-chain alteration appears to stabilize the “open” apo form of PILRA and likely alters the conformational sampling of the molecular clamp required to obtain its “closed” form to engage its ligands.
We therefore propose that in G78 PILRA (AD-risk associated), the engagement of SA by R126 and peptide by F76 is facilitated by G78 (Fig 3C). However, in the AD-protective PILRA variant R78, the R78 side-chain competes with the central R126-Q140 interaction and alters the positioning of F76 (Fig 3A), which leads to an overall decrease in PILRA ligand binding. This structure-based hypothesis is consistent with the reduced functional cellular binding observed for the R78 variant (Fig 1).
To further test this model, we generated two additional alanine mutants of PILRA at amino acids predicted to be essential (Q140) or non-essential (S141) for conformational changes associated with ligand interaction. 293T cells were transfected with G78 (AD risk), R78 (AD protective), Q140A and S141A variants of PILRA, and receptor-ligand interaction was measured after incubating cells with soluble NPDC1-mIgG2a. PILRA expression was comparable among variants, matching or exceeding G78 (AD risk) expression (S2 Fig). R78 (44% of G78, p = 0.02) and Q140A (22% of G78, p = 0.0004) variants showed significantly decreased binding to NPDC1, while S141A (117% of G78, p = 0.5) had no significant effect (Fig 3D and S6A and S6B Fig). These data are consistent with the experimental structural models that show the interaction of Q140 with R126 is important for productive sialic acid binding (Fig 3A to 3C). Consistently, the Q140A mutation has a strong effect because the Q140-R126 interaction network is completely abolished. By contrast, the AD-protective R78 variant likely has an intermediate effect since it only modulates the Q140 interaction with R126, which is expected to only alter the frequency or strength of relevant PILRA-ligand interactions.
We next investigated the interaction of PILRA variant and ligands in vitro using surface plasmon resonance (SPR). Human PILRA–Fc variants (G78, R78, or Q140A) were immobilized on a ProteOn GLC sensor chip and binding of NPDC1-mFc or a control mFc-tagged protein was measured. Qualitatively, NPDC1-Fc bound to the R78 (AD-protective) and Q140A (essential for R126 conformation) variants to a much lesser extent than to G78 (AD risk) PILRA, while control Fc-tagged protein showed no binding (Fig 3E).
To further probe the mechanistic basis of R78 (AD protective) function and phenotype, a more complete SPR characterization of NPDC1-His binding to PILRA variants was performed (S6C Fig). The affinity of NPDC1 toward R78 (AD-protective) PILRA (76.5 nM) was 4.5-fold weaker than the affinity toward G78 PILRA (16.8 nM). The on-rate constant kon for NPDC1-His binding to R78 (AD protective) (6.8×10+3 M-1s-1) was ~3-fold lower than binding to G78 (AD risk) PILRA (2.2×10+4 M-1s-1), while the koff rate constants were comparable (S6C Fig). These results are consistent with the idea that, once engaged, the affinity and disassociation rate of R78-ligand complexes are similar to G78 PILRA, but the frequency with which PILRA can productively engage with ligands is reduced in the R78 (AD protective) variant by R78 side chain interactions favoring the apo-state (Fig 3). Taken together, these data support a structural model in which R78 impairs PILRA-ligand interactions by altering the accessibility of a productive sialic acid-binding conformation in PILRA.
Given that PILRA is a known entry receptor for HSV-1 [41] and the R78 (AD protective) variant showed reduced binding to HSV-1 gB (Fig 1F), we next determined whether there were differences in HSV-1 infectivity based on PILRA genotype. We isolated and differentiated human monocyte-derived macrophages (hMDMs) from five pairs of healthy volunteers homozygous for either the G78 (AD risk) or R78 (AD protective) PILRA variants (matched for age, gender and ethnicity). hMDMs were infected with HSV-1 at different multiplicities of infection (MOI) (0.01, 0.1, 1 and 10), and infectivity was measured morphologically by light microscopy, by using an LDH cytotoxicity assay, by measuring intracellular viral DNA and in a viral plaque assay.
No notable cytopathic effects were observed in the first 6 h of infection, however at 18 hours post infection, extensive cytopathy was detected in G78/G78 PILRA-expressing hMDMs, including loss of cell shape, increased cell volume, birefringence, and formation of both cell aggregates and multinucleated giant cells (syncytia) (Fig 4A and S7 Fig). Cytopathic changes were less pronounced in R78/R78 (Alzheimer’s protective) homozygous hMDMs (Fig 4A and S7 Fig).
hMDMs from R78/R78 PILRA donors showed significantly less HSV-1-induced cytotoxicity at 18 hrs post infection in the LDH assay at 0.01, 0.1, or 1 MOI (Fig 4B and S4 Table). The difference was no longer significant at 10 MOI or if the infection was allowed to proceed for 36 hrs, except at the lowest MOI of 0.01 (Fig 4B, S8A Fig and S4 and S5 Tables).
hMDMs from R78/R78 donors showed 5–10 fold decreased amounts of HSV-1 DNA at 6 hrs at all MOIs (0.01, 0.1, 1 and 10), and at 18 hrs at lower MOIs (0.01 and 0.1), compared to those from G78/G78 donors (Fig 4C and S8C and S8D Fig). No significant differences in HSV-1 DNA were observed between the two genotypes at 18 hrs at higher doses (1 and 10 MOI) (Fig 4C and S8D Fig), or at 36 hrs for any dose of virus (S8B and S8E Fig).
Finally, we measured the amount of infectious HSV-1 virus by harvesting supernatants from HSV-1-infected hMDMs and measuring viral titer by plaque assays on Vero cells. Viral plaque forming units (PFUs) were significantly lower after 6 and 18 hrs of infection for all MOIs tested, and at 36 hrs for lower MOIs (Fig 4D and 4E, and S9 Fig). Taken together, these data indicate that R78/R78 macrophages were less susceptible to HSV-1 infection than G78/G78 macrophages.
We show here that PILRA G78R is a likely causal variant conferring protection from AD risk at the 7q21 locus. G78R alters the access to SA-binding pocket in PILRA, where R78 PILRA shows reduced binding to several of its endogenous cellular ligands and with HSV-1 gB. Reduced interaction with one or more of PILRA’s endogenous ligands (including PIANP and NPDC1) could impact microglial migration or activation [35–37]. In fact, microglia up-regulate the expression of the PIANP gene in the PS2APP, 5xFAD, and APP/PS1 mouse models of AD [42–44]. The identification of C4 as a novel PILRA-interacting protein is also intriguing, given the increased expression of C4 in mouse AD models [42], the increase in amyloid deposition observed when complement activation is inhibited [45], and the genetic association of complement receptor 1 with AD [46]. Finally, we note that both TREM2 and PILRB function as activating receptors and signal through DAP12 [32,34,47]. A reduction of PILRA inhibitory signals in R78 carriers could allow more microglial activation via PILRB/DAP12 signaling and reinforce the cellular mechanisms by which TREM2 is believed to protect from AD incidence [48]. The relevant ligands for PILRA/PILRB in vivo and the mechanism by which reducing PILRA-ligand interaction confers protection from Alzheimer’s disease remain to be elucidated.
A role for infection in accelerating AD has been proposed, but remains controversial [49]. HSV-1 is a neurotropic virus that infects a large fraction of the adult population and has frequent reactivation events. HSV-1 has been implicated in AD pathogenesis by several lines of evidence, including the presence of HSV-1 viral DNA in human brain tissue [50,51], increased HSV-1 seropositivity in AD cases [52–55], the correlation of high avidity HSV-1 antibodies with protection from cognitive decline [55], the binding of HSV-1 gB to APOE-containing lipoproteins [56], HSV-1-induced amyloidogenic processing of amyloid precursor protein (APP) [57–59], and preferential targeting of AD-affected regions in HSV-1 acute encephalitis [60]. In addition, HSV-1 gD receptors and gB receptor PILRA increase with age in multiple brain regions, including the hippocampus [61]. Additional AD risk loci have been proposed to play a role in the life cycle of HSV-1 [62], including CR1, which is capable of binding HSV-1 [63]. The reduced infectivity of HSV-1 in R78/R78 macrophages suggests that brain microglia from R78/G78 and R78/R78 individuals are less susceptible to HSV-1 infection and more competent for immune defense during HSV-1 recurrence.
These data provide additional evidence for a key role of microglia in AD pathogenesis and provide a mechanism by which HSV-1 may contribute to AD risk. Inhibiting the interaction of PILRA with its ligands could therefore represent a novel therapeutic mechanism to prevent or slow AD progression.
Blood samples and genotypes from healthy human volunteers from the Genentech Genotype and Phenotype program (gGAP) were used in this project. Written consent was obtained from all participants in the gGAP program. The study was reviewed and approved by the Western Regional Institutional Board (Study Number: 1096262, IRB Tracking Number: 20080040).
The conditional analysis between rs1476679 and rs1859788 was performed using the Genome-wide Complex Trait Analysis (GCTA) program’s Conditional & joint (COJO) analysis option. This program takes summary statistics as input. We used the summary statistics for rs1859788, rs1476679 from IGAP stage1 GWAS [15]. The COJO program also needs a reference population to calculate the LD and to perform the conditional analysis. For reference population analysis we used the raw genotype data from ADGC cohort. There were 22,255 individuals in this cohort that had the non-missing genotype for the rs1859788. The ADGC dataset was also used for the minor allele frequency calculations that are provided in the text.
The coding sequences (CDS) of full length PILRA (AJ400841), human herpesvirus 1 strain KOSc glycoprotein B (HSV-1 gB) (EF157316), and neural proliferation, differentiation and control 1 (NPDC1) (NM_015392.3) were cloned in the pRK neo expression vector. Several PILRA point mutations were generated, including R72A, F76A, G78R, S80G, Q140A and S141A. The PILRA variants were incorporated into a full-length G78 (AD risk) PILRA construct by site-directed mutagenesis as per the manufacturer’s recommendation (Agilent Cat. No. 200523) and sequences were verified. A full length myc-DDK tagged PIANP construct was purchased from Origene (Cat. No. RC207868). Full length complement component 4A (Rodgers blood group) C4A (NM_007293.2), extra cellular domain (ECD) of amyloid beta precursor like protein 1 (APLP1) (NM_005166) (1–580 aa) and ECD of sortilin-related VPS10 domain-containing receptor 1 (SORCS1) (NM_052918) (1–1102 aa) were fused with C-terminal gD tag (US6/gD, partial [Human alphaherpesvirus 1) (AAP32019.1) and GPI anchor in pRK vector. The ECD of all PILRA variants (1–196 aa) and NPDC1 (1–190 aa) were PCR amplified and cloned with C-terminal murine IgG2a Fc tag in a pRK expression vector.
ECDs of PILRA variants (G78 (AD risk), R72A, F76A, G78R, S80G, Q140A and S141A) and NPDC1 fused to the Fc region of murine IgG2a were expressed in a CHO cell expression system, supernatants collected, protein A/G affinity-purified and verified by SDS-PAGE and mass spectroscopy.
293T cells were transfected with lipofectamine LTX reagent (ThermoFisher) with various full-length constructs of PILRA variants (G78 (AD risk), R72A, F76A, G78R, S80G, Q140A and S141A). After 48 hours, the transfected cells were harvested and incubated with soluble mIgG2a-tagged ligand, NPDC1-mFc at 50 μg/ml (as described above) for 30 minutes on ice. Cells were then washed and stained with 1 μg/ml chimeric anti-PILRA antibody (mouse Fc region is substituted to human IgG1 backbone on anti-PILRA antibodies [31]) on ice for 30 min followed by APC-conjugated mouse anti-human IgG (BD Pharmingen Cat.No. 550931) and FITC anti-mouse IgG2a (BD Pharmingen Cat. No. 553390) secondary antibodies according to manufacturer’s instruction. PILRA-transfected 293T cells were examined by flow cytometry for binding of NPDC1 by measuring the frequency of APC and FITC double-positive cells. Double positive cells were gated on the WT sample and than the gates were overlaid on subsequent samples to maintain the same cell population throughout the experiment. For each PILRA variant, the mean percentage of the number of cells binding to NPDC1-mFC relative to the wild type PILRA binding for each experiment was calculated.
In the inverse experiment, 293T cells were transfected with lipofectamine LTX reagent [ThermoFisher] with known full-length PILRA ligand (NPDC1, HSV-1gB and PIANP) and predicted ligand constructs (SORCS1, APLP1 and C4A) (described above). After 48 hours, the transfected cells were harvested and incubated with soluble mIgG2a-tagged variants of PILRA (G78 (AD risk), R72A, F76A, G78R, S80G) (described above) 50 μg/ml for 30 min on ice. Cells were then washed and stained with FITC anti-mouse IgG2a (BD Pharmingen Cat. No. 553390) secondary antibody according to manufacturer’s instruction. PILRA ligand-transfected 293T cells were examined by flow cytometry for binding to PILRA variants by measuring the frequency of FITC-positive cells. The percentage of mean fluorescence intensity (MFI) of PILRA-mFC binding on ligand-transfected cells relative to the wild type PILRA binding for each experiment was calculated.
Binding of human NPDC1.Fc to PILRa-Fc variants was measured by SPR using a ProteOn XPR36 (Bio-Rad). PILRA-Fc WT and variants (G78R and Q140A) were immobilized on a ProteOn GLC sensor chip (Bio-Rad) by EDC/NHS amine coupling (2000–2400 RU’s) and the chip surface was deactivated by ethanolamine after immobilization. NPDC1-Fc diluted in PBST or a control Fc-tagged protein was injected at a concentration of 100 nM over the immobilized PILRA proteins at room temperature[31].
Healthy human volunteers from the Genentech Genotype and Phenotype program (gGAP) were genotyped for rs1859788 (PILRA G78R) using custom design ABI SNP genotyping assay with the following primers; Forward primer seq: GCGGCCTTGTGCTGTAGAA, Reverse primer seq: GCTCCCGACGTGAGAATATCC, Reporter 1 sequence: VIC- ACTTCCACGGGCAGTC-NFQ, Reporter 2 sequence: FAM- ACTTCCACAGGCAGTC-NFQ. To control for a possible effect of the eQTL for PILRB, all volunteers selected were homozygous AA (lower PILRB expression) for rs6955367 (http://biorxiv.org/content/early/2016/09/09/074450). Genotype for rs6955367 was determined using an InfiniumOmni2.5Exome-8v1-2_A.bpm. Peripheral Blood Mononuclear Cells (PBMC’s) were obtained by Ficol gradient from five pairs of homozygous donors for rs1859788 (one with each genotype AA/GG). The pairs of samples were matched for age [± 5 years], gender and self-reported ethnicity. Monocytes were purified from PBMC’s by negative selection using the EasySep Human Monocyte Enrichment Kit without CD16 Depletion (19058), as recommended by the manufacturer. Isolated monocytes were differentiated into macrophages in DMEM + 10%FBS + 1X glutaMax and 100 ng/ml MCSF media for 7–10 days. The gGAP program was reviewed and approved by the Western Regional Institutional Board.
Macrophages differentiated from healthy human monocytes were incubated with 10, 1, 0.1 and 0.01 multiplicity of infection (MOI) of HSV-1 virus at 37°C for 1 hour with gentle swirling to allow virus adsorption. Cells were washed after 1 hr of adsorption and infection was continued for 6, 18 and 36 hrs. Supernatant was harvested at 6, 18 and 36 hrs of infection and cell debris were removed by centrifugation at 3000 rpm for 5 min at 4°C. DNA was isolated from infected cells using the QIAamp DNA mini-kit (Qiagen Cat. No. 51304). Additional cells were fixed with 4% paraformaldehyde after infection and stained with DAPI for microscopy.
The CytoTox 96 Non-Radioactive Cytotoxicity Assay (Promega Cat. No. E1780) was performed on supernatant harvested from HSV-1-infected human macrophages as per manufacturer’s recommendations to measure cell toxicity after HSV-1 infection. For each sample, the percent cytotoxicity was calculated as the ratio of LDH released in culture supernatant after infection to completely lysed cells (maximum LDH release).
HSV-1 DNA was quantitated using a custom design ABI TaqMan gene expression assay, with the following primers: Forward primer seq: 5'-GGCCTGGCTATCCGGAGA-3', Reverse primer seq: 5'-GCGCAGAGACATCGCGA-3', HSV-1 probe: 5'-FAM-CAGCACACGACTTGGCGTTCTGTGT-MGB-3'. GAPDH DNA was quantitated using ABI endogenous control (Applied Biosystem Cat. No. 4352934E). Amplification reactions were carried out with 5 μl of extracted DNA from infected cells in a final volume of 25 μl with TaqMan Universal PCR Master Mix (Applied Biosystems Cat. No. 4304437) as per manufacturer’s recommendations. HSV-1 DNA (Ct values) was normalized to cell GAPDH (Ct values) to account for cell number.
Virus titers from HSV-1-infected cells were determined following a standard plaque assay protocol [64]. In brief, the plaque assay was performed using Vero cells (African Green Monkey Cells) seeded at 1x105 cells per well in 48-well plates. After overnight incubation at 37°C, the monolayer was ~90–100% confluent. Supernatants harvested from HSV-1-infected human macrophages were clarified from cells and debris by centrifugation at 3000 rpm for 5 minutes at 4°C. Virus-containing supernatants were then diluted from 10−1 to 10−8 in DMEM (1 ml total volume). Growth media was removed from Vero cells and 250 μl of supernatant dilution was transferred onto the cells, followed by incubation at 37°C for 2 hrs with gentle swirling every 30 min to allow virus adsorption, after which the virus-containing media was aspirated. The cells were then overlaid with 2% methylcellulose containing 2X DMEM and 5% FBS and incubated at 37°C. 48 hrs post-infection, plaques were enumerated from each dilution. Virus titers were calculated in pfu/ml.
|
10.1371/journal.pntd.0006335 | Group A Streptococcus pharyngitis and pharyngeal carriage: A meta-analysis | Antibiotic treatment of Group A Streptococcus (GAS) pharyngitis is important in acute rheumatic fever (ARF) prevention, however clinical guidelines for prescription vary. GAS carriers with acute viral infections may receive antibiotics unnecessarily. This review assessed the prevalence of GAS pharyngitis and carriage in different settings.
A random-effects meta-analysis was performed. Prevalence estimates for GAS+ve pharyngitis, serologically-confirmed GAS pharyngitis and asymptomatic pharyngeal carriage were generated. Findings were stratified by age group, recruitment method and country income level. Medline and EMBASE databases were searched for relevant literature published between 1 January 1946 and 7 April 2017. Studies reporting prevalence data on GAS+ve or serologically-confirmed GAS pharyngitis that stated participants exhibited symptoms of pharyngitis or upper respiratory tract infection (URTI) were included. Included studies reporting the prevalence of asymptomatic GAS carriage needed to state participants were asymptomatic.
285 eligible studies were identified. The prevalence of GAS+ve pharyngitis was 24.1% (95% CI: 22.6–25.6%) in clinical settings (which used ‘passive recruitment’ methods), but less in sore throat management programmes (which used ‘active recruitment’, 10.0%, 8.1–12.4%). GAS+ve pharyngitis was more prevalent in high-income countries (24.3%, 22.6–26.1%) compared with low/middle-income countries (17.6%, 14.9–20.7%). In clinical settings, approximately 10% of children swabbed with a sore throat have serologically-confirmed GAS pharyngitis, but this increases to around 50–60% when the child is GAS culture-positive. The prevalence of serologically-confirmed GAS pharyngitis was 10.3% (6.6–15.7%) in children from high-income countries and their asymptomatic GAS carriage prevalence was 10.5% (8.4–12.9%). A lower carriage prevalence was detected in children from low/middle income countries (5.9%, 4.3–8.1%).
In active sore throat management programmes, if the prevalence of GAS detection approaches the asymptomatic carriage rate (around 6–11%), there may be little benefit from antibiotic treatment as the majority of culture-positive patients are likely carriers.
| Treating sore throats caused by Group A Streptococcus infections (GAS pharyngitis) with antibiotics is important for preventing acute rheumatic fever (ARF). It is impossible to distinguish patients with true GAS pharyngitis infections from GAS carriers with pharyngitis caused by viral infections when throat swab culturing alone is used to diagnose GAS pharyngitis. Carriers are not likely to benefit from antibiotic treatment, but may receive treatment unnecessarily. Reported rates of GAS pharyngitis and carriage vary considerably depending on the setting. Thus it is difficult to ascertain which groups are likely to benefit significantly from active sore throat management programmes which treat GAS pharyngitis in order to prevent ARF. We performed a meta-analysis to estimate the prevalence of GAS pharyngitis and asymptomatic carriage in different settings. Approximately 10% of all children swabbed for a sore throat in clinical settings have true GAS pharyngitis, but this increases to around 55% if the children have GAS detected in their throat using swab cultures. In active sore throat management programmes, the prevalence of GAS detection is lower than in clinical settings and if it declines towards 8% (the asymptomatic carriage level), there may be little benefit in treating GAS culture-positive patients with antibiotics.
| Acute pharyngitis is a common cause of doctor’s visits across the world.[1] Most pharyngitis episodes (40–80%) are caused by self-limiting viral infections. Group A Streptococcus (GAS) infection is the most common bacterial cause of pharyngitis, responsible for approximately15-30% of cases.[2] In a small minority of patients (0.3–3%), untreated GAS pharyngitis may trigger acute rheumatic fever (ARF).[3–5] ARF and its sequela, rheumatic heart disease (RHD), remain important public health problems in low- and middle-income countries,[6–8] and persist in certain (predominantly Indigenous-minority) groups in high-income countries.[9, 10] Indigenous Australians and New Zealand Māori and Pacific populations have among the highest rates of ARF in the world.[1]
There is conflicting pressure on clinicians to either prescribe antibiotics to patients with pharyngitis to reduce the risk of ARF, or to withhold prescriptions and minimise antibiotic-related harms.[11–14] Most high-income countries have national clinical guidelines on antibiotic treatment of GAS pharyngitis, but guidelines differ markedly in their recommendations.[15] For example, in North America, Finland and France, throat swabbing and prescribing antibiotics to patients with GAS culture-positive (GAS+ve) pharyngitis is recommended.[15–17] Conversely, antibiotic treatment is discouraged in other high-income countries, notably the United Kingdom, Belgium and the Netherlands.[15, 18] In New Zealand and Australia, clinical guidelines restrict swabbing and treatment to patients at high-risk of ARF.[16, 17] Specifically, in New Zealand it is recommended that antibiotic treatment begin immediately after a symptomatic high-risk patient presents to a healthcare provider, but be discontinued if GAS is not cultured. In this instance, patients may be exposed to several days of unnecessary treatment while laboratory results are generated.[19, 20] Despite clinical guidelines, healthcare practitioners sometimes prescribe antibiotics to relieve symptom duration, regardless of the patient’s ARF risk.[21]
Accurate diagnosis of true GAS pharyngitis infection remains a major barrier to effective ARF prevention. Some individuals carry GAS in the throat, but have no symptoms of infection nor an antibody response.[11, 22] The Infectious Diseases Society of America makes a strong recommendation against routine antibiotic treatment of carriers. This recommendation is based on evidence in the literature indicating that carriers are unlikely to transmit GAS pharyngitis, and face little or no risk of developing complications (including ARF).[23] In addition, a previous review estimated the prevalence of asymptomatic pharyngeal GAS carriage at 12% amongst school-aged children.[24] When throat culture alone is used to diagnose GAS pharyngitis, many patients prescribed antibiotics are likely suffering viral pharyngitis with coincidental GAS carriage.[25–27] The reference standard for determining whether true GAS pharyngitis is present requires both throat culture and serological testing to identify GAS+ve patients with elevated antibodies targeting conserved GAS antigens, streptolysin-O and deoxyribonuclease-B.[28, 29] However, serological testing relies on obtaining patient blood samples and thus is not routinely performed in primary care. This situation has resulted in a major knowledge gap with respect to the prevalence of true GAS pharyngitis.
There is an important need to ensure that clinicians target high-risk individuals with effective, evidence-based treatment strategies, particularly in an era of increasing antimicrobial resistance. Accurate prevalence estimates of serologically-confirmed GAS pharyngitis across different geographic and population settings are therefore needed to inform clinical practice and policy. Accordingly, this study aimed to use a systematic literature review to determine: (i) the prevalence of GAS culture-positive pharyngitis in different settings and populations; (ii) the prevalence of serologically-confirmed GAS pharyngitis in symptomatic GAS+ve individuals; and (iii) the prevalence of asymptomatic pharyngeal GAS carriage.
No patient recruitment or other involvement in this study was required and consequently ethics approval was not needed. All data analysed were anonymised.
A systematic literature review was conducted and reported in accordance with PRISMA guidelines.[30] No pre-existing review protocols for were identified, but the methodology was loosely based around that of a previously published meta-analysis by Shaikh et al.[31] A total of 17 literature searches on Medline and EMBASE databases were performed to identify articles containing prevalence data on GAS+ve pharyngitis, serologically-confirmed GAS pharyngitis and asymptomatic pharyngeal GAS carriage published between 1 January 1946 and 7 April 2017. Search terms on Medline included MeSHs: ‘Streptococcal Infections’ AND ‘Pharyngitis’. Search terms on Embase included MeSHs: ‘Streptococcus Group A’ AND ‘Pharyngitis’, also ‘Streptococcal pharyngitis’ AND ‘Streptococcus Group A’. Keyword searches using the terms: ‘GAS pharyngitis’ OR ‘streptococcal pharyngitis’ AND ‘ASO’ OR ‘anti-streptolysin’ OR ‘anti-DNase-B’ were employed on both databases. Keyword searches using the terms: ‘ASOT’ OR ‘ADBT’ OR ‘ADB’ AND ‘streptococc*’ were also performed. Search findings were limited to ‘Humans’. Further details of the search strategy are listed in the Methodology Appendix. Publications were catalogued using Endnote X7. A researcher (JO) screened the search results and applied the eligibility criteria. Eligible literature, identified in title or abstract screening, was obtained for full screening. Where systematic reviews were identified, prevalence studies they referenced were searched by title on Google Scholar or Medline, if published, and searched by title using the Google search engine if not published. Non-English language papers were screened using Google translator.
Explicit a priori inclusion and exclusion criteria were applied to assess article quality and reduce bias. Only studies using throat swab culture with an agar plate and incubator to detect GAS were included. Because we were interested in the point prevalence of GAS, longitudinal studies in which participants were swabbed multiple times had data from the first swab included only. When investigating the prevalence of GAS+ve pharyngitis, all studies that stated participants exhibited symptoms of pharyngitis or upper respiratory tract infection (URTI) were included where they presented to health practitioners who decided to obtain a throat swab. GAS pharyngitis demonstrates a wide range of clinical presentations[23, 32] and we aimed to maximise the number of relevant studies included. For studies investigating serologically-confirmed GAS pharyngitis, the same criteria applied and prevalence data were abstracted where the authors considered their findings provided serological confirmation. Information on the method of confirmation was noted where available. For included studies investigating the prevalence of asymptomatic GAS carriage, the authors needed to state participants did not have symptoms of pharyngitis or URTI when the throat swab was obtained, otherwise were excluded. Only studies that reported the number of participants and the number (or proportion) that produced GAS in throat cultures were included (and where applicable, the number of participants that were serologically-confirmed as having GAS pharyngitis). The country or countries recruitment was conducted in was also required to be discernible for inclusion, as was the participant recruitment method (active/passive). Studies were excluded if they were not likely to be representative of the general population, notably those conducted in outbreaks and other distinct settings (for example, from closed communities such as detention centres). We excluded studies that could not be translated to English.
Demographic and prevalence data were abstracted and entered on a spreadsheet (JO). This included the study citation, participant age group/s, country study sample was drawn from, number of cases, sample size, and recruitment period (where available). A second reviewer (EMW) independently abstracted prevalence data. Where abstracted data differed between the two reviewers, the article was rechecked and remaining differences resolved through discussion between the study authors.
Results were stratified using up to five characteristics (Fig 1): (a) clinical outcome measured (GAS+ve pharyngitis, serologically-confirmed GAS pharyngitis, asymptomatic GAS carriage); (b) participant recruitment method (active or passive recruitment); (c) country income level (OECD or non-OECD member country); (d) age group; (e) where serologically-confirmed prevalence studies were reported, whether unequivocal case confirmation had occurred or otherwise.
‘GAS+ve pharyngitis’ was considered to occur when an individual with symptoms of pharyngitis or URTI received a throat swab which produced GAS when cultured. ‘Serologically-confirmed GAS pharyngitis’ was considered to occur when an individual with GAS+ve pharyngitis demonstrated an antibody reaction in response to GAS infection. ‘Unequivocal’ confirmation occurred when either a 0.2log10 or greater rise in ASO or ADB antibody titres was observed between acute and convalescent serum samples, or a four-fold increase in ASO titre occurred.[29, 33] ‘Asymptomatic GAS carriage’ occurred when individuals with no symptoms of pharyngitis or UTRI received a throat swab which produced GAS when cultured.
‘Passive recruitment’ was considered to occur where the study population presented to healthcare providers of their own accord and the practitioner obtained a throat swab. ‘Active recruitment’ occurred where a population had been sensitised to reporting pharyngitis or URTI symptoms to healthcare practitioners (e.g. by being asked about pharyngitis symptoms) and health services had been aligned to maximise accessibility to the participants (e.g. in terms of being close-by, involving home visits and/or offering tokens of thanks for study involvement). These distinctions were applied to pharyngitis studies. Prevalence studies of asymptomatic pharyngeal GAS carriage require active recruitment, as participants neither require nor present for treatment.
National Organisation for Economic Cooperation and Development (OECD) membership status was used as a means of classifying populations by socioeconomic position, as member countries tend to have high-income economies and very-high human development indexes[34] (Fig 1).
An ‘all ages’ analysis was performed, which included all studies in each category with no age restrictions applied. Other analyses were restricted to certain age groups: children aged <5 years old; children 5–19 years, all children aged < 20 years (‘children’) and ‘adults’ (generally including adults ≥20 years, but also allowing studies where this category started from ≥12 years if that was the adult category the authors used). If the study did not state the population age range, it was included in the ‘all ages’ analysis only. Exceptions were when the study population was described using terms such as ‘paediatric’, in which case it was pooled in the ‘Children’ category. The overall ‘children’ category included more studies than all the child subgroups put together. In order to be included in a more specific age category, the exact age range of the participants was required to be specified, and in many studies participants were simply described using terms such as ‘pediatric’ or ‘children under 16 years-old’. Similarly, studies which described their populations in such terms as ‘university students’ were grouped in the ‘Adult’ category.
Where individual articles did not state case data, but stated the prevalence of GAS and included the number of participants tested, the prevalence percentage was multiplied by the number of participants to find the number of cases.
A random-effects meta-analysis was used to produce pooled estimates for all outcome measures. Outcome measures were expressed as summary point prevalence percentages with 95% confidence intervals (CIs). Serologically confirmed GAS pharyngitis prevalence was calculated in two ways: serologically-confirmed patients as a proportion of the total number of symptomatic participants who had throat swabs; and as a proportion of those with GAS+ve throat swabs only. A sub-analysis of serological GAS studies was performed, including only those studies that met the unequivocal criteria. The programme R (version 3.4.1) was used throughout the analysis with the meta package.[35]
To estimate what proportion of total variation across study groups was due to heterogeneity rather than chance, the I2 statistic was used. Heterogeneity in pooled study groups was considered low if I2 <30%, moderate if 30–59%, substantial if 60–75% and considerable if >75%.[36]
In total, 4,022 articles were identified and 1,076 were selected for further investigation. Exclusion criteria are listed in Fig 2. Overall, 285 articles that reported prevalence data on GAS were included.
Included articles addressed GAS culture-positive pharyngitis (254 studies), serologically-confirmed GAS pharyngitis (21 studies) and asymptomatic GAS carriage (56 studies). Of studies that reported on culture-positive pharyngitis, only 22% (57 studies) reported on populations that did not live in OECD countries. Nine of these 57 studies used active recruitment strategies, as did nine OECD studies. All others used passive recruitment. Three-quarters of passive recruitment studies were based in community primary care settings, often general practitioner clinics. One quarter were based in hospital Emergency Departments. Approximately 10% recruited in both hospital and primary care clinics.
Considerable heterogeneity was observed within most pooled study groups with some exceptions. Pooled serological studies demonstrated moderate heterogeneity, and low heterogeneity when pooled by age group. Moderate heterogeneity was also observed in the pooled prevalence estimate for GAS carriage in children <5 years old. Further details of abstracted data in this review are presented in the S1 Appendix, with accompanying measures of heterogeneity.
A detailed breakdown of prevalence estimates with numbers of included studies and participants is provided in Table 1.
The overall ‘all age’ prevalence of GAS+ve pharyngitis was 22.7% (95% CI: 21.2–24.2%). Children (<20 years old) demonstrated the highest prevalence of culture-positive GAS pharyngitis: 25.2% (23.1–27.5%). When restricted to children aged <5 years old, a 16.6% (12.6–21.6%) prevalence was estimated. When restricted to children aged 5–19 years old, a prevalence of 24.3% (19.3–30.1%) was found. A prevalence of 13.7% (11.1–16.8%) was identified in adults. GAS+ve pharyngitis was more prevalent in OECD countries: 24.3% (22.6–26.1%) than in non-OECD countries: 17.6% (14.9–20.7%).
Passive recruitment, that is clinical settings where participants self-present to a healthcare provider, generally detected markedly higher prevalence estimates (overall prevalence: 24.1%, 22.6–25.6%) than active recruitment (overall prevalence: 10.0%, 8.1–12.4%), where a population was sensitised to reporting a sore throat. This discrepancy was especially marked in 5-19-year-old OECD children (with a pooled prevalence of 36.8% (30.9–43.1%) in clinical settings, compared with 11.6% (8.3–16.1%) in active sore throat management programmes). Similarly, for non-OECD 5-19-year-old children, the prevalence of GAS+ve pharyngitis was much higher: 37.4% (27.7–48.2%) in clinical settings than in active sore throat management programmes: 9.2% (4.9–16.6%, Table 1, Fig 3).
This review attached most weight to the pooled prevalence estimate from six reported studies that used the unequivocal criteria for detecting serologically-confirmed GAS pharyngitis. Only 12 of 21 studies investigating serologically-confirmed GAS pharyngitis provided data on the total number of symptomatic individuals swabbed to identify those that were GAS+ve, on whom serological investigation was undertaken (details in S1 Appendix). Of these 12 studies, six reported using the unequivocal criteria–that being a significant titre increase in paired sera. All six were conducted in OECD populations and only one used active recruitment.
The overall ‘all age’ prevalence of serologically-confirmed GAS pharyngitis was 9.4% (5.6–15.5%). Studies using the unequivocal criteria detected a higher pooled prevalence (with an overall ‘all age’ prevalence of 16.4%, 9.9–26.0%). Higher prevalence estimates (22.6%, 17.8–28.2%) were detected when active recruitment was used, however this estimate is based on a single study which included paired serology. By comparison, pooled studies which used passive recruitment with unequivocal confirmation detected an overall prevalence of 15.2% (8.1–26.7%). Where participants have GAS+ve pharyngitis, the proportion of serologically-confirmed patients is around 50%, and around 60% in 5-19-year-old children (Table 2, Fig 3).
Table 3 shows the prevalence of GAS carriage as well as numbers of included studies and participants. The overall prevalence of asymptomatic carriage was 7.0% (5.6–8.8%). When studies were pooled regardless of country income, the highest carriage was observed in children <20 years old (8.0%, 6.6–9.7%). A slightly lower overall prevalence was detected in non-OECD settings (6.4%, 4.6–8.9%, compared with 7.5%, 5.3–10.3%) in OECD settings (Table 3, Fig 3).
Pooled prevalence of GAS+ve pharyngitis, serologically confirmed pharyngitis and asymptomatic carriage are shown graphically in Fig 3A–3B for specific age groups and country income levels.
GAS+ve pharyngitis was the most prevalent manifestation of GAS. Higher levels were found in OECD countries. The overall prevalence of carriage was similar in high- and low-country income settings, however GAS carriage was twice as prevalent in children from OECD countries compared to children in non-OECD countries (Fig 3A). In passive recruitment OECD studies overall, the sum of the asymptomatic carriage prevalence and the serologically confirmed GAS pharyngitis prevalence approximately equals the prevalence of culture-positive GAS pharyngitis. This relationship was also observed, albeit with less certainty, when restricted to children <20 years old (Fig 3B). This relationship could not be explored in active recruitment settings as only one study in that category examined serologically confirmed GAS pharyngitis.
To our knowledge, this is the first comprehensive review of pharyngeal GAS detection that has assessed all three of its clinically relevant manifestations: GAS+ve pharyngitis; serologically-confirmed GAS pharyngitis; and asymptomatic GAS carriage. The prevalence of serologically-confirmed GAS pharyngitis for school-aged children, who have the highest risk of ARF, has been quantified. In high-income countries only one in 10 children with pharyngitis symptoms are likely to have serologically-confirmed GAS pharyngitis. Where participants are identified as having GAS +ve pharyngitis, the proportion that are serologically-confirmed is around 50–60%. This finding supports the use of throat swabbing in symptomatic children, rather than providing presumptive antibiotic treatment. The prevalence of serologically-confirmed GAS pharyngitis indicates how many children may go on to develop ARF as a complication,[19, 23, 37] which in turn indicates how effective a sore throat management programme is likely to be. A limitation here, as with all sore-throat management programmes, is that up to two-thirds of ARF cases do not appear to present with preceding pharyngitis, so other prevention strategies are necessary when seeking to remove the burden of disease.[38, 39]
The decision to obtain a throat swab was likely influenced by the healthcare practitioner’s suspicion for GAS pharyngitis and concern for possible complications. The Centor criteria gives an indication of the likelihood of a sore throat being due to bacterial infection. Practitioners may have been more inclined to swab patients presenting with symptoms suggestive of GAS pharyngitis, such as fever.[40, 41] GAS+ve prevalence is strongly affected by whether patients present to sore throat management programmes that actively recruit them (active recruitment), or present of their own accord to healthcare practitioners with manifestations of pharyngitis (passive recruitment). Passive recruitment strategies tend to detect a higher prevalence, compared with active recruitment methods–for example, in children less than 20 years old, a higher prevalence (28.5% in OECD countries and 23.1% in non-OECD countries) was observed in those presenting to primary healthcare providers, compared with those identified through specialised programmes (16.6% in OECD countries and 9.1% in non-OECD countries). This difference may be due to active recruitment studies including patients with less severe pharyngitis, who would not otherwise seek treatment for their symptoms. Given that an estimated 8% of children are carriers, if the prevalence of GAS detection in active recruitment studies approaches this level, then the majority of culture-positive patients are likely to have carriage, not true GAS pharyngitis. It is therefore important for active sore throat treatment programmes to monitor the prevalence of GAS detection and consider serological testing for a sample of patients.
Around 37% of 5-19-year-old children in passive recruitment settings have GAS+ve pharyngitis, both in OECD and non-OECD countries. Despite this, many non-OECD countries have much higher rates of GAS diseases and ARF.[1] The intercountry distribution of GAS pharyngitis does not therefore appear to reflect the very wide difference in ARF rates. This apparent discrepancy could be because GAS skin infections dominate in tropical climates where the highest burden of contemporary ARF occurs and may be a major driver of ARF in these settings.[42–44] ARF is also ecologically associated with poverty, overcrowding and potentially other environmental factors which vary markedly in time and place.[45, 46] International research has consistently noted associations between ARF and socioeconomic conditions, including poor housing conditions.[1, 45, 47–54]
Our review of 285 studies greatly extends the findings of a previous review by Shaikh et al. which included only 29 studies[24] and did not include data on GAS seroconversion, which is generally accepted as the key clinical outcome that triggers the autoimmune process driving ARF.[23, 37, 55] Shaikh et al. also reported a pooled prevalence estimate of 37% for GAS culture-positive pharyngitis in children, the same as our identified prevalence for 5–19 year-old children in passive recruitment settings (where the majority of included studies originated). The use of a less stringent inclusion criteria to meet our study aims is justified as sore throat management programmes will generally treat any GAS+ve patient with a self-reported sore throat.[56] The previous review reported a 12% asymptomatic carriage prevalence estimate, very similar to our carriage estimate for OECD children, but higher than our overall carriage estimate for children of 8%.
This review has several limitations. Firstly, as pooled studies span multiple continents, ecological bias is apparent. Pooling study data over time averages temporal variation in GAS distribution. Thus it is not possible to distinguish geographical or temporal trends in the pooled prevalence estimates. Stratifying according to OECD status was intended to reduce these effects, particularly that of disease determinants and ecological biases. Secondly, whether ASO and ADB titres are a valid means of identifying ‘true’ GAS pharyngitis remains a matter of debate, however an increase in antibody titre is a much more accurate indicator of GAS infection than a single titre result.[57] Serological techniques to determine titre may have varied across studies. Our use of the unequivocal criteria for serological confirmation attempted to minimise these biases. Thirdly, despite our best efforts, it is possible that relevant literature was not identified or was mistakenly excluded as our database access permitted articles dating back to 1946 to be obtained online. Finally, we did not attempt to systematically assess publication bias or the quality of included studies (beyond investigating heterogeneity). Hand-searches of previously published review reference lists were performed in an attempt to reduce publication bias. Risk of bias assessment tools can be problematic due to the broad nature of their specifications and the variable interpretation (and applicability) of criteria.[58] Studies in closed communities and disease outbreak settings were excluded in an attempt to reduce selection bias, as was our use of stringent inclusion criteria, and stratifying findings according to participants’ age range and country income level. Despite this, there is still considerable heterogeneity in most pooled study categories. This is likely due to differences in inclusion and exclusion criteria applied within the pooled studies. As a result, a range of clinical manifestations and severities are included in the final prevalence calculations. This feature can actually be considered a study strength, given the wide range of clinical presentations pharyngitis patients present to healthcare practitioners with.
Due to the collateral damage of antibiotic misuse on human health and the environment, there is a pressing need to target pharyngitis testing and treatment in the most effective and efficient way possible. School-aged children with symptomatic sore throats have a relatively low chance of having serologically-confirmed GAS pharyngitis, particularly in organised sore throat management programmes. ARF prevention programmes need to be carefully designed with this knowledge in mind and targeted to groups at high risk of ARF. Ultimately, reducing ARF is likely to depend on prevention programmes that address the underlying determinants of disease risk, such as income, housing conditions and access to primary healthcare. Further research should validate the main conclusions of this systematic review, particularly through collection of GAS serological data in low- and middle-income countries. It would also be useful to have more studies that measured all three clinically important GAS throat infection outcomes in the same populations at the same time to see how these states are related to one another.
|
10.1371/journal.ppat.1005701 | CRISPR/Cas9-Mediated Genome Editing of Herpesviruses Limits Productive and Latent Infections | Herpesviruses infect the majority of the human population and can cause significant morbidity and mortality. Herpes simplex virus (HSV) type 1 causes cold sores and herpes simplex keratitis, whereas HSV-2 is responsible for genital herpes. Human cytomegalovirus (HCMV) is the most common viral cause of congenital defects and is responsible for serious disease in immuno-compromised individuals. Epstein-Barr virus (EBV) is associated with infectious mononucleosis and a broad range of malignancies, including Burkitt’s lymphoma, nasopharyngeal carcinoma, Hodgkin’s disease, and post-transplant lymphomas. Herpesviruses persist in their host for life by establishing a latent infection that is interrupted by periodic reactivation events during which replication occurs. Current antiviral drug treatments target the clinical manifestations of this productive stage, but they are ineffective at eliminating these viruses from the infected host. Here, we set out to combat both productive and latent herpesvirus infections by exploiting the CRISPR/Cas9 system to target viral genetic elements important for virus fitness. We show effective abrogation of HCMV and HSV-1 replication by targeting gRNAs to essential viral genes. Simultaneous targeting of HSV-1 with multiple gRNAs completely abolished the production of infectious particles from human cells. Using the same approach, EBV can be almost completely cleared from latently infected EBV-transformed human tumor cells. Our studies indicate that the CRISPR/Cas9 system can be effectively targeted to herpesvirus genomes as a potent prophylactic and therapeutic anti-viral strategy that may be used to impair viral replication and clear latent virus infection.
| Herpesviruses are large DNA viruses that are carried by almost 100% of the adult human population. Herpesviruses include several important human pathogens, such as herpes simplex viruses (HSV) type 1 and 2 (causing cold sores and genital herpes, respectively), human cytomegalovirus (HCMV; the most common viral cause of congenital defects, and responsible for serious disease in immuno-compromised individuals), and Epstein-Barr virus (EBV; associated with infectious mononucleosis and a wide range of malignancies). Current antiviral drug treatments are not effective in clearing herpesviruses from infected individuals. Therefore, there is a need for alternative strategies to combat these pathogenic viruses and prevent or cure herpesvirus-associated diseases. Here, we have assessed whether a direct attack of herpesvirus genomes within virus-infected cells can inactivate these viruses. For this, we have made use of the recently developed CRISPR/Cas9 genome-engineering system to target and alter specific regions within the genome of these viruses. By targeting sites in the genomes of three different herpesviruses (HSV-1, HCMV, and EBV), we show complete inhibition of viral replication and in some cases even eradication of the viral genomes from infected cells. The findings presented in this study open new avenues for the development of therapeutic strategies to combat pathogenic human herpesviruses using novel genome-engineering technologies.
| Herpesviruses are large DNA viruses that cause widespread, lifelong infections; most adults carry multiple herpesviruses [1]. The herpesvirus family is divided into three subfamilies, the Alpha-, Beta- and Gammaherpesvirinae. The subfamily of Alphaherpesvirinae includes the herpes simplex virus type 1 and type 2 (HSV-1 and 2) and varicella zoster virus (VZV). HSV-1 causes cold sores and herpes simplex keratitis, a common cause of corneal blindness [2, 3]. HSV-2 is responsible for genital herpes. Primary infection with VZV results in chickenpox; reactivation may lead to herpes zoster or shingles [4]. The subfamily of Betaherpesvirinae includes the human cytomegalovirus (HCMV), which gives rise to serious complications in immuno-compromised individuals [5, 6]. Additionally, HCMV is the most common viral cause of congenital defects. The Gammaherpesvirinae include Epstein-Barr virus (EBV) and Kaposi's sarcoma-associated herpesvirus (KSHV). EBV induces infectious mononucleosis and is strongly associated with multiple malignancies, including nasopharyngeal carcinoma, Burkitt’s lymphoma, Hodgkin’s lymphoma, gastric carcinoma, and post-transplant lymphoproliferative disorders (PTLD) [7]. KSHV is a human tumor virus that is associated with Kaposi's sarcoma and two lymphoproliferative disorders occurring in AIDS patients: primary effusion lymphoma and multicentric Castleman disease [8].
Current treatment options to restrict the clinical manifestations of productive herpesvirus infections are limited and all approved antiviral agents target the viral DNA polymerase [9, 10] during the productive (lytic) phase of infection. Herpesviruses, however, are characterized by their ability to establish a quiescent, latent state [1]. During latency, herpesviruses express only few viral gene products allowing them to persist in the host without being effectively cleared by our immune system. During this stage, herpesviruses are not actively replicating their viral genomes by viral DNA polymerases, rendering antiviral treatments targeting these polymerases ineffective. Occasionally, herpesviruses reactivate, thereby producing new virus progeny; depending on the herpesvirus in question, reactivation may cause serious disease [1]. Examples are herpes simplex keratitis, genital herpes and herpes zoster, caused by HSV-1, HSV-2 and VZV, respectively. For EBV, pathology is mainly associated with latent infection which is associated with the occurrence of various tumors of lymphoid and epithelial origin [11]. There is a need to clear the latent infection, as this would allow the removal of herpesvirus pathogens from the human host; this would prevent future reactivation events and herpesvirus-associated diseases.
Archaea and bacteria have evolved an adaptive immune system comprised by CRISPR/Cas (clustered regularly interspaced short palindromic repeats (CRISPR)—CRISPR–associated (Cas) systems) that use short RNA molecules to induce degradation of foreign nucleic acids of viruses and other genetic elements [12–15]. Recently, this CRISPR/Cas9 system has been engineered to induce robust RNA-guided genome modifications in human cells [16]. By co-expressing a bacterial Cas9 nuclease with a guideRNA (gRNA), one can direct Cas9 to almost any site in the genome and induce cleavage of double stranded DNA (dsDNA) at the target site [16]. This cleaved DNA is subsequently ‘repaired’ by mammalian DNA repair mechanisms that are inherently error-prone, thereby inducing insertions and deletions (‘indels’) and mutations at the target site. Applications of the CRISPR/Cas9 system are currently revolutionizing the field of molecular biology. Researchers are now able to generate full knockouts for any gene in any cell-type [17, 18], induce specific editing/mutagenesis of loci [15, 17, 19], edit multiple alleles simultaneously [20, 21], in addition to many other applications [12, 15, 22]. The CRISPR/Cas9 system has also been engineered to selectively modify dsDNA viruses [23–30], positive-sense single stranded RNA viruses, such as Hepatitis C [31] and integrated HIV proviral DNA in human cells [32–36]. Therefore, CRISPR/Cas9 targeting of herpesviral genomes may provide a powerful strategy to combat these viruses.
In this study, we set out to explore whether the CRISPR/Cas9 system can be rewired to limit herpesvirus infection during the latent and productive stage of the viral life cycle. For this, we usurped the CRISPR/Cas9 system to induce efficient editing of the genomes of three prototypic members of the human herpesvirus family. By targeting gRNAs to essential viral genes, we show effective abrogation of HCMV and HSV-1 replication. Simultaneous targeting of HSV-1 with two gRNAs completely impaired the production of new infectious particles from human cells. Using the same approach, EBV can be efficiently cleared from latently infected EBV-transformed human tumor cells. On the contrary, CRISPR/Cas9 appears inefficient at targeting quiescent HSV-1 genomes whereas replication post virus reactivation can be efficiently abrogated. The data presented in this study indicate the potential of the CRISPR/Cas9 system as a new therapeutic strategy to combat pathogenic human herpesviruses.
We set out to explore whether the CRISPR/Cas9 system could be targeted to the genome of human herpesviruses to directly limit virus replication in human cells. We generated a lentiviral CRISPR/Cas9 vector and asked whether human herpesviruses within infected cells can be genetically modified using this system. We initially focused our efforts on EBV, whose dsDNA genome resides in the nucleus of infected cells [7]. EBV can establish latency in B-lymphocytes, where the EBV genome circularizes and resides as an episome in the nucleus. During latency, the virus does not produce virus progeny [7] and expresses a limited number of viral proteins, in addition to non-coding RNAs, including a large set of miRNAs. We assessed whether these viral miRNA genes were amendable to CRISPR/Cas9 editing (see S1C Fig for a graphical representation of the experimental setup). The latently infected gastric carcinoma cell line SNU-719 was transduced to express CRISPR/Cas9 gRNAs targeting the viral miRNAs BART5, BART6, or BART16. The relative miRNA activity of these specific miRNAs was assessed using miRNA sensor constructs in which a single miRNA target site was cloned downstream of the fluorescent mCherry reporter (S1B Fig). Introduction of the miRNA sensors in SNU-719 cells resulted in downregulation of mCherry expression (S1B Fig), confirming functional expression of the miRNAs in these cells. Expression of the sensors was partially restored upon treatment of the cells with gRNAs targeting the corresponding miRNAs (Fig 1A), indicating site-specific editing of the EBV genome. Indeed, sequencing of the genomic miRNA loci showed editing of the targeted regions (Fig 1B). Hence, the CRISPR/Cas9 system allows for direct editing of the genome of latent EBV in EBV-positive tumor cells.
Since EBV can serve as a target for CRISPR/Cas9 genome editing, we reasoned that targeting of viral genetic factors critically involved in EBV genome maintenance could render the virus unstable, causing its loss from latently infected cells. We designed gRNAs targeting the viral EBV nuclear antigen 1 (EBNA1) and several areas within the EBV origin of replication (OriP). The latter included EBNA1 binding sites and the dyad symmetry element, all of which are involved in EBV episome maintenance and replication [37, 38]. As a model system we used the Burkitt’s lymphoma Akata-Bx1 cells that carry a recombinant EBV expressing eGFP under control of the CMV promoter [39]. Hence, eGFP expression serves as a marker for the presence of EBV in these cells. Upon transduction of the Akata-Bx1 cells with gRNAs targeting EBNA1, we observed a loss in eGFP expression in 40–60% of the cells, which did not occur in control cells expressing gRNAs targeting cellular genes (Fig 2A and 2B). Also targeting the EBNA1 binding sites within the EBV OriP and the dyad symmetry element induced a clear loss in eGFP expression from Akata-Bx1 cells (Fig 2A and 2B), suggesting depletion of EBV from these cells. Sequential introduction of combinations of these active gRNAs induced an almost complete loss of eGFP from the majority of cells (Fig 2B, right bar diagrams). We assessed the EBV content in these gRNA-expressing Akata-Bx1 cells by qPCR and indeed detected a strong reduction of the EBV genome content in the cells carrying double gRNAs (Fig 2C). Targeting EBNA1 with two different gRNAs proved most efficient, inducing over 95% loss of EBV genomes. These results indicate that CRISPR/Cas9-mediated targeting of essential regions within the EBV dsDNA effectively reduces the viral genome content in latently infected cells.
Since CRISPR/Cas9 proved efficient in editing and clearing latent EBV infections, we next assessed whether we could impact herpesvirus replication in human cells. For this, we turned to a lytic infection model for HCMV, the most-commonly studied member of the Betaherpesvirinae. Replication of HCMV is dependent on a large number of viral replication factors. We selected seven of these essential genes and asked whether CRISPR/Cas9 targeting of these genes impact virus infection. We designed four gRNAs per gene for the viral polymerase UL54, the polymerase accessory protein UL44, the single strand DNA binding protein UL57, the primase UL70, the DNA helicase UL105, the major capsid protein UL86, and UL84, which is involved in the initiation of viral DNA replication [40, 41]. After lentiviral delivery of the gRNAs to MRC5 cells, the cells were challenged with eGFP-encoding HCMV derived from the TB40/E strain [42]. For each of the essential HCMV genes, one or more gRNAs were capable of almost completely impairing HCMV replication which resulted in survival of the cells and absence of eGFP expression as assessed by flow cytometry (Fig 3A). Unexpectedly, none of the gRNAs targeting UL84 were effective at limiting infection (Fig 3A). In cells transduced with control vector or vectors carrying gRNAs targeting host genes, the percentage of eGFP-positive cells was similar as observed for untransduced cells (Fig 3A). The gRNAs targeting the nonessential HCMV genes US6, US7, and US11 also did not interfere with HCMV replication. We assessed the percentage of HCMV sequences that were edited at the US7 and US11 target sites via CRISPR/Cas9 engineering, and observed alteration at these sites in 76 and 95% of cases, respectively. This shows that, although these genes were mutated, this did not interfere with virus replication.
The inhibitory capacity of the anti-HCMV gRNAs was also assessed towards a different strain of HCMV: AD169. Most gRNAs that were effective at limiting TB40/E replication also impaired AD169 replication (Fig 3B). Unlike TB40/E, AD169 was targeted successfully by two independent gRNAs directed against UL84 (Fig 3B). Hence, replication of AD169 crucially depends on UL84, whereas replication of TB40/E does not. This data agrees with experiments described by Spector and Yetming [43].
In conclusion, effective inhibition of HCMV replication can be achieved using the CRISPR/Cas9 technology; this requires targeting of essential HCMV genes.
Many single anti-HCMV gRNAs effectively impaired HCMV infection, preventing virus replication and spread up to 11 days post infection (Fig 3A). However, replication-competent virus emerged after prolonged culture of cells in the majority of gRNA expressing cells. This may be due to the outgrowth of virus variants that harbor CRISPR/Cas9-induced mutations that still allow for expression of functional proteins. To assess this, virus variants were isolated from two cultures of CRISPR-expressing cells that were infected at high MOI (0.5) and displayed outgrowth of virus after 21 days of culture (UL57 gRNA #1 and UL70 gRNA #4). Subsequent 454 sequencing of the targeted genes predominantly detected variants that maintained the correct reading frame of these two essential genes (Fig 4) as mostly deletions of complete codons (i.e. 3-6-9-12 bp, etc) were detected. We quantified the number of sequences with frameshift variants, and observed a clear depletion of these in the anti-HCMV gRNA expressing cells targeting essential genes (1,3% for UL57 #1, and 4,6% for UL70 #4) as compared to gRNAs targeting the nonessential genes US7 and US11 (83,5 and 85,8% respectively) (Fig 4A). Furthermore, the sequence complexity of the mutants selected upon UL57 and UL70 targeting was low (Fig 4B and 4C), suggesting selection of few ‘fit’ variants and subsequent expansion of these infectious mutants over time. To conclude, the CRISPR/Cas9-technology represents a promising strategy to impair HCMV replication; however, its successful application requires effective modification of the viral genome to avoid the emergence and selection of escape variants that bypass CRISPR/Cas9 editing.
Since CRISPR/Cas9 proved efficient in limiting productive infection of the slowly replicating HCMV virus, we next assessed whether also the fast replicating alphaherpesvirus HSV-1 can be inhibited using this approach. We designed four gRNAs/gene targeting twelve essential HSV-1 genes [44]: the terminase UL15, the glycoprotein B coding UL27, the major binding protein UL29, the DNA polymerase UL30, the tegument protein UL36, the capsid assembly protein UL37, the DNA polymerase processivity factor UL42, the DNA replication protein UL5, the DNA helicase/primase complex protein UL52 and associated protein UL8, the transcriptional regulation protein UL54, and the replication origin binding protein UL9. We included gRNAs targeting the nonessential protein kinase US3 and the surface membrane protein US8. These anti-HSV-1 gRNAs were introduced into Vero cells. Subsequently, the cells were infected with HSV-1-eGFP, and monitored for eGFP expression as a measure for HSV-1 infection after 2 days. High percentages of infected cells were detected in untransduced cells and in cells transduced with empty vector or with vectors carrying gRNAs directed against cellular genes (Fig 5A). Most gRNAs targeting essential HSV-1 genes impaired virus replication effectively. In contrast to our observations for HCMV, targeting of nonessential HSV-1 genes (US3 and US8) also reduced HSV-1 replication, albeit to a lesser extent as compared to targeting essential genes.
Infection of gRNA-expressing Vero cells with HSV-1 resulted in breakthrough of HSV-1 for most gRNAs at 3 dpi (Fig 5B). Interestingly, the gRNAs targeting UL29, and the primase-helicase complex genes UL8 and UL52 [45, 46] maintained their potency at 3 dpi. This effect was apparent at both high (0.5) and low MOI (0.05) and resulted in a concomitant loss of HSV-1 genomes from infected cells (Fig 5C). These results indicate that infection with the fast replicating HSV-1 can be limited considerably; this, however, requires the use of effective gRNAs.
When using single gRNAs, prolonged propagation of cells infected with herpesviruses at high MOIs may result in the selection of escape mutants that bypass CRISPR/Cas9 editing. Simultaneous targeting of multiple essential viral genes may impair viral replication more effectively and prevent the generation of escape mutants. To investigate this possibility, Vero cells expressing single gRNAs targeting the essential HSV-1 genes UL8, UL29, or UL52 were compared to cells carrying double gRNAs targeting combinations of these genes. Cells expressing single gRNAs were partially protected from challenge with HSV-1 at high MOI, but the cells were not able to clear HSV-1 at 3 dpi (Fig 6A). Expressing double anti-HSV-1 gRNAs, however, induced a gradual loss of HSV-1 infected cells and resulted in a cell population that was HSV-1 negative at 3 dpi (Fig 6A). Similar results were obtained when double anti-HSV-1 gRNAs were introduced in human MRC5 cells that were challenged with HSV-1 (Fig 6B). To monitor the impact of the gRNAs on the generation of HSV-1 progeny, we assessed virus titers in the supernatants of these single and double gRNA-expressing cells via plaque assay (Fig 6C and 6D). As expected, we observed a clear drop in the amount of infectious viral particles present in the supernatants of these gRNA-expressing cells. A single gRNA targeting UL52 resulted in a four-log reduction in virus titer (Fig 6C). Combining two gRNAs targeting UL29 and UL8 or UL29 and UL52, however, caused a complete loss of infectious viral particles, with more than six log reduction in progeny virus (Fig 6C). In the plaque assay performed with supernatants harvested from single gRNA-expressing cells, smaller ‘plaques’ appeared that were not visible by eye. This was especially apparent for viruses harvested from the anti-UL8 gRNA-expressing cells (Fig 6D, lower panel). We speculate that the reduced plaque size was caused by mutant, attenuated virus. Importantly, we did not observe any signs of infection in the plaque assay performed with supernatants harvested from double gRNA-expressing cells. Therefore, our results show that simultaneous targeting of multiple essential viral genes using CRISPR/Cas9 greatly improves the efficiency at which herpesviruses are cleared from infected cells as compared to single gene targeting.
Since prolonged expression of Cas9 and gRNAs in human cells could result in editing at off-target sites, we analyzed the activity of nine gRNAs to the top three predicted off-target sites within the human genome. We PCR-amplified these 27 loci from gRNA-expressing and control cells and subjected these to 454 sequencing (S3 Table) or conventional Sanger sequencing (S2 Fig). Importantly, no signs of CRISPR/Cas9-induced editing at these loci were detected, suggesting that overt editing at undesired sites did not occur.
Next, we assessed whether the CRISPR/Cas9 system can target the latent state of HSV-1 in infected cells. As a model, we adapted the in vitro HSV-2 quiescence model previously established by Russell and Preston [47, 48]. In short, MRC5 human lung fibroblast cells were infected with HSV-1-eGFP and immediately incubated at 42°C for 4 days. At this temperature, HSV-1 replication is non-permissive and the virus establishes a quiescent state resembling latency. Upon subsequent incubation of the culture at 37°C, HSV-1 remains in a quiescent stage for a prolonged period of time (S3A Fig). Upon subsequent superinfection with AD169 HCMV, HSV-1-eGFP reactivates from quiescence resulting in virus spread to neighboring cells which can be monitored by assessing CPE by microscopy or eGFP expression by flow cytometry (S3A and S3B Fig). Although the MRC5 cells contain quiescent HSV-1, the cells themselves are not quiescent and maintain the ability to divide. As there is a gradual loss of quiescent HSV-1 prior to reactivation caused by cell division, the timeframe to study the effect of anti-HSV-1 gRNAs is limited. Therefore, to increase lentiviral titers carrying anti-HSV-1 gRNAs, we expressed Cas9 from a separate lentiviral vector and generated an MRC5 line stably expressing the endonuclease (MRC5-Cas9 cells). Upon establishment of a quiescent HSV-1 infection in these cells, anti-HSV-1 gRNAs were introduced and selected to purity. We next assessed the effect of these gRNAs on HSV-1 reactivation by superinfection with HCMV. Untreated MRC5 cells containing quiescent HSV-1 showed no signs of HSV-1-eGFP replication (Fig 7A, cells alone), whereas superinfection of these cells with HCMV resulted in reactivation of HSV-1-eGFP and subsequent virus replication and spread (Fig 7A, cells alone, reactivated HSV-1). Quiescent HSV-1 was also efficiently reactivated in control cells transduced with empty gRNA vector, resulting in rapid spread of HSV-1-eGFP, (Fig 7A, vector control, reactivated HSV-1). However, cells expressing anti-HSV-1 gRNAs targeting UL8, UL52, or UL29 displayed abrogated HSV-1-eGFP replication.
The observed block in HSV-1-eGFP replication upon reactivation may be caused by direct targeting of the quiescent HSV-1 genome by the CRISPR/Cas9 system. The latter could result in destabilization of the quiescent genome or mutagenesis of essential genes. To assess the mechanism of the observed interference with HSV-1 replication, we isolated genomic DNA from gRNA-expressing quiescent MRC5-Cas9 cells prior reactivation and quantified their HSV-1 genome content by qPCR. No evident loss of HSV-1-eGFP genomes was observed in anti-HSV-1-gRNA expressing cells as compared to control cells (Fig 7B). We next assessed whether CRISPR/Cas9-induced editing of the quiescent HSV-1 genomes occurred at the gRNA target sites. For this, we PCR-amplified and deep-sequenced the gRNA target sites from quiescent cells in the presence or absence of anti-HSV-1 gRNAs (Fig 7C). No indels were present in any of the 5 lines receiving empty vector gRNA controls. Minor editing frequencies were detected in 2 out of 5 gRNA-expressing samples where 6 and 1% of sequences contained indels in a UL52 and UL8 gRNA-expressing line respectively (Fig 7C and S4 Fig). Intriguingly, most of the mutant UL52 reads were from a single 21 in-frame deletion that could result in functional UL52 protein, suggesting that it has been actively selected for. As no selection occurs during quiescence due to absence of virus replication, we speculate that the low percentage of CRISPR/Cas9 indel formation was not caused by targeting of the quiescent genome, but rather by editing of newly produced viral genomes in sporadic spontaneously reactivated cells. Indeed, the potent effect of gRNAs on reactivated HSV-1-eGFP (Fig 7A) cannot be explained by the ineffective indel formation measured in Fig 7C and points towards editing of newly produced viral genomes.
To conclude, our data suggest that anti-HSV-1 gRNAs can actively restrict replicating virus, but, in the model used, are inefficient at targeting the HSV-1 genome during quiescence when directing gRNAs to the genomic regions of UL8, UL29, and UL52.
Herpesviruses such as HSV-1, HSV-2, HCMV and EBV are a frequent cause of disease and treating these herpesvirus infections remains a major challenge. Current therapies are mainly geared towards limiting productive infections by targeting virus replication via antiviral drugs such as the guanosine analogues acyclovir and ganciclovir. Although these compounds effectively inhibit the viral DNA polymerase during replication, they have limited impact on the latent stage of herpesvirus infections that relies on host polymerases for viral genome maintenance in dividing cells [7]. Strategies to impact the latent stage of infection often depend on reactivation of a latent infection by using chemotherapeutic agents and subsequent targeting of these cells by e.g. ganciclovir [49]. Such strategies are not always effective and may result in loss of the target cell. Although this approach may be beneficial to clear EBV-driven cancer cells, other instances ask for sparing of the infected cell, for example in HSV-1, HSV-2 and VZV-infected neurons. Hence, there is a need for a direct and potent strategy to limit herpesvirus infections by specifically targeting the viral genome or its maintenance. Preferably, such approaches should impact both lytic and latent herpesvirus infections. Indeed, recently, the HSV-1 genome has been targeted by using meganucleases [50, 51]. Here, we performed proof-of-concept studies to show that the CRISPR/Cas9 genome engineering system can be efficiently used to limit or eradicate three of the most prevalent herpesviruses from human cells: HSV-1, HCMV and EBV.
EBV relies on EBNA1 for maintenance of the viral episome. EBNA1 is a well-studied protein that is consistently expressed in EBV-infected cells, including all EBV-associated malignancies [38, 52–54]. Indeed, selective EBNA1 inhibition via siRNAs [55], small molecule inhibitors [56, 57], or dominant-negative forms of the protein [58] impairs EBV latency and results in a loss of EBV episomes from infected cells. In the present study, we show that the CRISPR/Cas9 system can indeed target the EBNA1 protein-coding sequence or regulatory EBNA1 binding sites on the EBV genome, as a means to inhibit EBV episome maintenance. We show that the CRISPR/Cas9 system can target and edit EBV genes in naturally EBV-transformed cells of lymphoid and epithelial origin. Taken into account that the gastric carcinoma cell line SNU719 carries ±800 EBV genomes/cell [59], the efficiency of editing was high. Indeed, targeting of the essential EBNA1 gene using a single gRNAs in Akata-Bx1 cells resulted in 50–60% loss of latent EBV from cells. However, a proportion of the cells still contained EBV, which could be either wt unedited EBV, or mutant EBV that has received CRISP/Cas9-mediated mutations at the target region while retaining EBNA1 activity (e.g. by in an in-frame deletion). Subsequent introduction of a second gRNA targeting a different region on the EBNA1 gene, however, further enhanced the removal to up to 95% of EBV genomes. We hypothesize that simultaneous administration of multiple anti-EBV gRNAs into latently infected cells will further enhance the destabilization of EBV episomes and may facilitate complete eradication of the virus. This clearance could result in the loss of EBV-driven tumorigenic, cell cycle-promoting functions, anti-apoptotic features, and anti-inflammatory functions of EBV-encoded gene products and as such may serve as a new therapeutic strategy to combat EBV-associated malignancies. Indeed, an independent study showed that transient transfection of multiple anti-EBV gRNAs cleared EBV from a small subset of latently infected Raji B-cells, thereby inducing a proliferation arrest in these cells [29]. Our approach relies on stable expression of anti-EBV gRNAs via lentiviral vectors and induces an almost complete loss of EBV episomes, suggesting that a further optimization of this approach may allow for the complete removal of EBV from human cells.
CRISPR/Cas9 targeting of essential HCMV and HSV-1 genes efficiently abrogated virus replication in human cells. The mechanism of protection by CRISPR gRNAs is likely threefold. Introducing a dsDNA break in the viral genome may impair packaging of intact viral genomes, limiting the production of viral particles. Second, a dsDNA cut within an essential gene may impair the production of the targeted protein, and hence impact the biology of virus replication. Third, upon repair of the CRISPR/Cas9-induced dsDNA break by the hosts’ non-homologous end joining machinery, the target site is frequently mutated, resulting in the formation of virus mutants with (often) impaired protein function. In the latter case, the generation of new viral particles may be hampered directly, or mutated DNA may be properly packaged into new viral particles, but replication is subsequently halted in newly infected cells. In all three cases, the virus replication will be impaired. Indeed, for HSV-1 we observe a block in virus replication when we target either essential or nonessential genes. This indicates that a dsDNA cut within the HSV-1 genome alone is sufficient to impact virus replication, likely because less correctly assembled viral particles are generated, thereby blocking subsequent infection of naïve cells. By targeting essential HSV-1 genes, we observe an increased impairment of HSV-1 replication. Here, the effect is likely caused by a combination of genome destabilization and impaired expression of the essential genes. The latter effect is further enhanced by simultaneous targeting of two essential genes using two gRNAs. Here, the viral genome may be fragmented, which adds to the potency observed for single anti-HSV-1 gRNAs. Indeed, when targeting two sites in a given linear piece of DNA, large parts can be excised out of human [17, 60], dsDNA viral [23, 24, 29, 30, 61], and integrated HIV genomes [32–36]. Intriguingly, we did not observe reduced virus replication when targeting the nonessential HCMV genes US7, US10, or US11, whereas targeting essential genes did. This result was unexpected, and may be caused by a difference in replication kinetics between HCMV and HSV-1. Since HCMV is a slowly replicating virus, the DNA repair machinery may have sufficient time for complete repair of the cleaved viral genomes to occur. Since in vitro replication of the mutant HCMV is not affected by mutations in nonessential genes, the virus may replicate as efficiently as the wild type virus. However, for fast-replicating viruses such as HSV-1, the replication speed could outpace the time needed for complete dsDNA break repair. Indeed, nonhomologous end-joining (NHEJ) kinetic studies [62] showed that ±75% of dsDNA breaks in MRC5 cells are repaired within a two-hour timeframe, whereas complete repair requires approximately 24 hours to occur. Since HSV-1 has a replication cycle of less than 18 hours [63], it is possible that slow NHEJ kinetics impacts HSV-1 more profoundly than HCMV that has a replication cycle of ±72 hours [63].
Although the majority of gRNAs targeting essential genes of HSV-1 and HCMV were effective in abrogating virus replication, some showed limited to no activity (Figs 3A and 5A). Indeed, CRISPR/Cas9 activity can vary greatly between different gRNA sequences targeting the same gene [64]. The potency of gRNAs is partially dependent on sequence features of the expressed gRNA as well as the composition of the PAM sequence and downstream nucleotides in the target allele [64]. Additionally, local chromatin structure at the genomic target site impacts the ability of Cas9 to bind to DNA [65, 66], and thereby affect CRISPR/Cas9 activity. Since HSV-1 is a fast replicating virus, low activity anti-HSV-1 gRNA are likely ineffective in abrogating virus replication and spread as unaltered HSV-1 virus replication will quickly outcompete mutant viruses.
Using single gRNAs targeting genes essential for HCMV or HSV-1, we were able to severely impair replication of these viruses in human cells. Using a double gRNA–expression approach, the generation of new infectious HSV-1 was completely blocked, which caused a >106 drop in virus titers. We did observe the emergence and selection of escape mutants upon single gRNA treatments targeting essential genes for both HCMV and HSV-1. The HCMV mutants were derived from CRISPR/Cas9-induced genome editing events, as all selected variants displayed editing at the gRNA target cleavage site. There was a strong selection for variants that retained the reading frame of the targeted genes, as >95% of identified variants contained deletions of complete codons (changes of 3-6-9-12 bp etc) at the target site, whereas for gRNAs targeting the nonessential US7 and US11 genes such edits occurred at much lower frequencies (±15%). The ‘in frame’ variants likely retained (partial) functionality of the targeted proteins and were hence selected for during the infection cycle. Indeed, in the plaque assay performed with supernatants harvested from anti-UL8 gRNA-expressing cells, we noted the presence of small plaques indicative of presence of infectious, yet attenuated HSV-1 variants. Importantly, we did not observe any outgrowth of escape mutants when we targeted two essential HSV-1 genes. Apparently, the impact of HSV-1 genome fragmentation together with the loss of essential viral gene expression is too stringent to allow for the formation of infectious virus variants that bypass CRISPR/Cas9 targeting. Further increasing the number of anti-viral gRNAs expressed within a single cell may enhance this protective effect.
We observed broadly similar potencies of anti-HCMV gRNAs towards two commonly used HCMV strains (AD169 and TB40/E). However, CRISPR/Cas9 targeting of the HCMV UL84 gene protected human cells from infection by the AD169 strain, but not by the TB40/E strain. Previously, it has been shown that replication of HCMV strain TB40 is indeed UL84-independent [43]. If the CRISPR/Cas9 system is to be applied to combat human herpesviruses in a clinical setting, it would be important to target viral genes that are essential for all clinical virus isolates. Also, since the CRISPR/Cas9 system acts on dsDNA, directing gRNAs to well-conserved regions in these genes is essential to target the majority of viral strains. Nevertheless, CRISPR/Cas9 target cleavage does allow few mismatches between the target DNA and the gRNA [17], allowing also less conserved regions to be targeted. In our studies, we targeted gRNAs to the N-terminal coding regions of essential viral genes. In this approach, indel mutations causing frame-shifts will completely impair protein function. An alternative strategy to further enhance the potency of anti-viral CRISPRs is to target gRNAs to DNA coding for key amino acid residues within such essential proteins. This way, any substitution or indels at these sites could render the targeted protein non-functional and prevent the emergence of escape mutants. Unlike fast-evolving RNA viruses such as HIV and Influenza virus [67], dsDNA herpesviruses evolve at a slow rate [68]. Therefore, the emergence of escape mutants will be mainly governed by the stochastic process of CRISPR/Cas9-induced error-prone NHEJ, and not by mutagenesis events caused by the replication machinery. Indeed, the likelihood of generating in-frame indels at a target site is relatively high (±15% for HCMV US7 and US11 genes) resulting in a significant chance for selecting mutant virus that retain fitness. Thus, instead of using single gRNAs to target herpesviruses, simultaneous administering of multiple CRISPR gRNAs will further reduce the likelihood of generating escape mutants as the viral genome will not only be mutagenized at multiple essential sites, but will also become fragmented
The potential of the CRISPR/Cas9 system to induce off-target editing of the human genome has been a matter of concern [15, 69]. Even though our CRISPR-expressing cells continuously expressed the CRISPR/Cas9 system for several weeks, we did not identify any obvious off-target activity at the top three predicted off-target sites for the nine most potent gRNAs. However, it remains possible that these gRNAs do induce editing at other genomic sites. Clearly, if the CRISPR/Cas9 system would be applied to combat herpesviruses in a clinical setting, the potential off-target activity should be reduced to an absolute minimum. Much progress has been made to limit CRISPR/Cas9 off-target activity. For example, Ran et al. describe the use of paired gRNAs in combination with a Cas9 mutant that can only induce nicking in the dsDNA. NHEJ will only occur when two successful gRNA-directed cuts occur in close proximity [70]. Another study describes the use of tru-gRNAs, where shortened gRNAs retain their potent on-target activity, but show much reduced off-target editing [71]. A combination of these two approaches may further reduce the (already limited) off-target activity of the CRISPR/Cas9 system. Another strategy is to engineer Cas9 derivatives by mutagenesis that have reduced off-targeting potential, yet retain their on-targeting activity [72–74]. Alternatively, as many Cas9 endonuclease variants exist in nature, it is likely that new variants will be identified that hold potent on-target, yet low off-target activity. A combination of these approaches will further reduce the (already limited) off-target activity of the CRISPR/Cas9 system.
Our data show that the CRISPR/Cas9 system can efficiently target latent EBV, but not quiescent HSV-1 in the model used. Although we did identify minor CRISPR/Cas9-mediated editing of quiescent HSV-1 in MRC5 cells in 2 out of 5 sequencing samples, it remains unknown whether this activity was directed towards the quiescent genome or towards new progeny virus that derived from an early spontaneous reactivation event. The MRC5 quiescence model can display spontaneous reactivation of HSV-1, which results in rapid virus replication and spread. Although we did not detect any signs of spontaneous reactivation in the experiments as presented in Fig 7C, it is conceivable that an early reactivation event occurred, which was subsequently targeted by the expressed gRNA and identified in the deep-sequencing studies. The minor HSV-1 editing that is observed cannot explain the efficient abrogation observed in HSV-1 replication upon HCMV-induced HSV-1 reactivation (Fig 7A). Hence, anti-HSV-1 gRNAs are efficient in targeting actively replicating virus, whereas quiescent HSV-1 genomes are inefficient substrates for the gRNAs used in the model we studied. The mechanisms underlying latency differ for various herpesviruses. This may result in differences in sensitivity to CRISPR/Cas9-mediated editing of latent genomes. EBV resides as an episome in the nucleus of actively dividing cells, where EBNA1 mediates replication of the viral genome via the host replication machinery without the formation of progeny virus [7]. HSV-1 latency on the other hand, occurs in non-dividing cells in sensory neurons where no genome replication occurs [75]. During latency, the HSV-1 genome is heavily methylated and viral gene expression is restricted [75]. It is suggestive that the CRISPR/Cas9 system is inefficient in accessing the tightly repressed state of the HSV-1 quiescent genome, whereas the open structure of the EBV genome during replication is accessible. Indeed, studies using models in human fibroblasts show that the quiescent HSV-1 genome is present in a tightly repressed state which is not responsive to most reactivation stimuli that normally de-repress latent genomes in neuronal cells [76, 77]. Hence, the low activity of CRISPR/Cas9 towards quiescent HSV-1 as observed in our model may not recapitulate the effect of anti-HSV-1 gRNAs in vivo. Additional in vitro and in vivo model systems are needed to assess whether CRISPR/Cas9 can target HSV-1 latency.
Human herpesviruses cause a wide range of diseases. The CRISPR/Cas9 approach may provide an attractive new strategy to combat such complications by directly impairing replication and removing these viral invaders from infected cells. The potential applications are abundant; one could direct anti-EBV gRNAs to remove EBV from EBV-driven tumors as an anti-tumor treatment. Anti-EBV and HCMV gRNAs may be used to deplete virus from tissues prior to transplantation, preventing donor-derived infections in immunocompromised recipients. Upon optimization, anti-HSV-1 gRNAs may be used to target latent HSV-1 in trigeminal ganglia, thereby curing herpes simplex keratitis, a common cause of eye infection leading to corneal blindness. VZV-associated zoster and the related post-herpetic neuralgia, a frequent and serious complication, may be treated in a similar fashion. In all these cases, the development of potent CRISPR/Cas9 delivery systems is of key importance. We have performed our studies using lentiviral delivery vectors; such vectors have been used as delivery vehicle in 114 clinical trials (http://www.abedia.com/wiley/vectors.php) and have proven efficient at correcting genetic defects in humans [78]. Other viral delivery systems may rely on non-integrating adenoviruses or recombinant adenovirus-associated virus (rAAV) as vectors. These non-integrating viruses have been found to effectively deliver a wide-range of genetic factors into human cells and the first rAAV vector has recently been approved for the treatment of lipoprotein lipase (LPL) deficiency in Europe [79, 80]. Especially the neurotropic nature of rAAV vectors [81] make these viruses attractive vehicles to introduce anti-HSV-1, HSV-2 and VZV gRNAs into infected neuronal cells. Furthermore, alternative delivery methods, such as the direct delivery of proteins to cells [82, 83], may remove the risk of insertional mutagenesis.
In conclusion, we have shown that CRISPR/Cas9 technology can effectively clear herpesvirus infections from human cells in vitro. We observed highly efficient and specific clearance of EBV from latently infected tumor cells and impairment of HSV-1 and HCMV replication in human cells. By combining two gRNAs targeting two essential HSV-1 genes, we completely inhibited the generation of new infectious virus during a lytic HSV-1 infection in vitro. Although CRISPR/Cas9 was inefficient at directing genome engineering of quiescent HSV-1 in our in vitro model, virus replication upon reactivation of quiescent HSV-1 was efficiently abrogated using anti-HSV-1 gRNAs. These new insights may allow the design of effective therapeutic strategies to target human herpesviruses during both latent and productive infections.
MRC5 human lung fibroblast cells, 293T human embryonic kidney cells, and Vero African Green monkey kidney epithelial cells, were obtained from ATCC (American Type Culture Collection). SNU719 cells were kindly provided by Prof. Jaap Middeldorp and were originally obtained from the Korean Cell Line Bank. MRC5 and Vero cells were grown in DMEM (Lonza AG, Switzerland) supplemented with glutamine, penicillin/streptomycin and 10% FCS. 293T and SNU719 cells were grown in RPMI 1640 medium (Lonza AG, Switzerland) supplemented with glutamine, penicillin/streptomycin and 10% FCS.
The EBV-positive Akata-Bx1 cells were kindly provided by Prof. Lindsey Hutt-Fletcher (Louisiana State University Health Sciences Center, Shreveport, USA). Akata-Bx1 cells carry a latent recombinant EBV in which the nonessential thymidine kinase gene has been replaced with a neomycin resistance gene under control of the thymidine kinase promoter and a modified eGFP gene under control of the CMV promoter [39].
The AD169-eGFP recombinant virus was kindly provided by Prof. Robert Kalejta (University of Wisconsin, MA, USA). AD169-eGFP (referred to as AD169 in the text) carries a simian virus 40 early promoter-driven eGFP gene in place of the viral US4-US6 region [84]. The UL32-eGFP-TB40/E strain was kindly provided by Prof. Christian Sinzger, Universitätsklinikum Ulm, Germany). UL32-eGFP-TB40/E (referred to as TB40/E in the text) is a recombinant HCMV expressing eGFP fused to the C terminus of the capsid-associated tegument protein pUL32 (pp150) [42]. HCMV viruses were propagated in MRC5 cells using standard culturing and harvest techniques. HCMV titers were determined in MRC5 cells using the eGFP marker as readout.
The HSV-1-eGFP strain constructed by Prof. Peter O’ Hare (Imperial College, London, UK) was kindly provided by Dr. Georges Verjans (Erasmus Medical Center, Rotterdam, The Netherlands). HSV-1-eGFP is a derivative of HSV-1 strain 17 expressing a VP16-eGFP fusion protein [85]. HSV-1-eGFP virus was propagated in Vero cells using standard culturing and harvest techniques and viral titers were determined in Vero cells using the eGFP marker as readout.
HCMV strain AD169 was obtained from ATCC (VR-538). AD169 virus was used to induce reactivation of latent HSV-1 by superinfection of latent MRC5 cells. AD169 virus was propagated in MRC5 cells using standard culturing and harvest techniques.
We constructed a selectable lentiviral CRISPR/Cas vector by altering the lentiviral pSicoR vector [86] (Addgene plasmid 11579, Tyler Jacks Lab, MIT) to express a human codon-optimized nuclear-localized codon-optimized S. pyogenes Cas9 gene (taken from Addgene plasmid 42229 [17], Zhang lab) that was N-terminally fused to PuroR via a T2A ribosome-skipping sequence [87]. This cassette was expressed from the human EF1A promoter. Additionally, we replaced the mouse U6 promoter with a human U6 promoter which drives expression of a guideRNA (gRNA) consisting of a 18-20bp target-specific CRISPR RNA (crRNA) fused to the trans-activating crRNA (tracrRNA) [88] and a terminator sequence. We generated a second variant of this vector, replacing the PuroR gene for a BlastR gene. These vectors are called pSicoR-CRISPR-PuroR and pSicoR-CRISPR-BlastR respectively. The pSicoR-CRISPR-PuroR vector was previously successfully used in Van de Weijer et al [89].
crRNA target sequences were designed manually or using the crispr.mit.edu website (Zhang lab, MIT) by entry of the first 250 nucleotides downstream of the start codon for the selected genes. CRISPR gRNA sequences were selected by highest score for specificity and the least off-targets within the human genome, as provided by the online CRISPR design tool. For each virus gene target, multiple independent CRISPR gRNAs were selected and cloned into the pSicoR-CRISPR-PuroR vector and verified by sequencing. gRNA target sequences used in this study are presented as supporting information (S1 Table). For the HSV-1 latency studies, gRNAs were introduced in a variant of the pSicoR-CRISPR-PuroR vector in which the Cas9 gene was removed. In these studies, Cas9 is expressed from the pSicoR-CRISPR-ZeoR vector. This vector is identical to pSicoR-CRISPR-PuroR vector, although PuroR was replaced by ZeoR and the U6-gRNA cassette was removed.
Lentiviral EBV miRNA reporters were cloned in the pSicoR backbone (Addgene plasmid 11579, Tyler Jacks Lab, MIT). For this, the CMV promoter was replaced with a human EF1A promoter that now drives expression of a PuroR-T2A-mCherry cassette allowing for selection and tracking of lentivirally transduced cells. Downstream of the PuroR-T2A-mCherry cassette we cloned single perfect miRNA target sites for the following EBV-encoded miRNAs: BART5-5p (5’-CGATGGGCAGCTATATTCACCTTG-3’), BART6-3p (5’-TCTAAGGCTAGTCCGATCCCCG-3’), and BART16-5p (5’-AGAGCACACACCCACTCTATCTAA-3’). All vectors were sequence verified.
For lentiviral transductions, virus was produced in 24-well plates using standard lentiviral production protocols and third-generation packaging vectors. Due to the large size of the lentiviral pSicoR-CRISPR-PuroR and pSicoR-CRISPR-BlastR vectors, titers were often low. When necessary, lentiviruses were concentrated using Lenti-X (Clontech) ±10 fold prior use. Target cells were transduced with lentiviral supernatants in the presence of polybrene (4μg/ml) via spin infection at 1,000g for 90 minutes at 33 °C. Three days post infection, successfully transduced cells were selected via puromycin (2μg/ml) or blasticidin (10μg/ml) treatment. For Vero cells, we used 5μg/ml puromycin or 20μg/ml blasticidin respectively. Double gRNA expressing cell lines were generated by initial transduction using the appropriate pSicoR-CRISPR-PuroR vector, followed by complete puromycin selection and subsequent transduction with the appropriate pSicoR-CRISPR-BlastR vector and blasticidin selection.
±35,000 CRISPR expressing cells were seeded in a 48 well plate in triplicate and infected with indicated MOIs of HCMV-eGFP or HSV-1-eGFP. The percentage of herpesvirus-infected cells were monitored in time by subjecting half of the well to flow cytometric analysis (FACSCanto II, BD BioSciences) to detect expression of the eGFP marker after formaldehyde fixing (1%) of the cells. For each flow cytometric measurement at least 2000 events were counted, although for some samples late in lytic infection this number could not be reached due to massive cell death. Flow cytometry data were analyzed using FlowJo software.
EBV miRNA sensor experiments were performed in latent EBV-positive SNU719 cells. Cells were transduced with control or anti-EBV miRNA gRNAs (see S1 Table for sequences), selected by puromycin for 2 days, and allowed to recover for ±12 days. Subsequently, cells were transduced with mCherry EBV miRNA sensors for BART5-5p, BART6-3p, or BART16-5p and analyzed by flow cytometry 4 dpi to assess miRNA activity in these cells (FACSCantoII, BD). CRISPR-induced editing of the EBV genome was assessed in anti-BART5 and anti-BART16 gRNAs expressing cells by amplifying these loci by PCR amplification using primers 5’-CGGGCTATATGTCGCCTTAC-3’ and 5’-AGAGGGTGGTGATCTTGGTG-3’ for EBV BART5 and 5’-CCAGGTCAGTGGTTTTGTTTC-3’ and 5’-TGGACCAACCTTAAAGTACCAAC-3’ for EBV BART16. Subsequently, PCR products were cloned in the pCR2.1-TOPO vector according manufacturers’ recommendations (Life technologies), transformed in Turbo competent cells (New England Biolabs) and several clones were mini-prepped (Fermentas) and subjected to sequencing (Macrogen Inc) using the M13-reverse primer 5’-GGAAACAGCTATGACCATG-3’. Sequence analysis was carried out using the Lasergene DNASTAR software package.
To assess the effect of anti-EBV CRISPRs on the EBV latent genome, Akata-Bx1 cells were transduced in triplicate with indicated anti-EBV or control gRNAs expressed from the pSicoR-CRISPR-PuroR vector. Cells were selected by puromycin for 2 days and allowed to recover. The percentage of eGFP-expressing cells as marker for latent EBV infection was monitored by flow cytometry (FACSCanto II, BD) around ±21 days post transduction and analyzed using FlowJo software. For double gRNAs, cells were initially transduced and puromycin-selected with gRNAs expressed from the pSicoR-CRISPR-PuroR vector, and subsequently transduced and blasticidin-selected with gRNAs expressed from the pSicoR-CRISPR-BlastR vector. Double anti-EBV gRNAs were monitored for eGFP expression at ±21 days post infection with the second gRNA.
To study the frequency of off-target editing in the human genome, we assessed the potential occurrence of off-target sites for nine gRNAs via the CRISPR design tool of the Zhang lab (http://crispr.mit.edu/). For each gRNA, we selected the top three scoring potential off-target sites and designed oligos using the primer3web (http://bioinfo.ut.ee/primer3/) to allow amplification of these sites (see S2 Table). We subsequently amplified these sites from gRNA-expressing (cells were transduced with gRNAs at least two weeks prior genomic DNA isolation) and control cells using Platinum Taq DNA polymerase (Life Technologies), purified the samples using the GeneJet PCR purification kit (Thermo Scientific), and assessed the concentration of DNA by the Quant-iT PicoGreen dsDNA Assay kit (Invitrogen) or the Qubit dsDNA BR Assay (Life Technologies). PCR amplicons were pooled and sequenced via 454 sequencing. To allow discrimination between amplicons from control and CRISPR-expressing cells, we designed two primer-sets by adding a 454 linker and a unique barcode to the 5’ end of the primers. Reverse and forward primers for CRISPR-expressing cells contained a 5’-CGTATCGCCTCCCTCGCGCCATCAGAGACGCACTC-3’ and 5’-CTATGCGCCTTGCCAGCCCGCTCAGAGACGCACTC-3’ addition respectively. Reverse and forward primers for control cells contained 5’-CGTATCGCCTCCCTCGCGCCATCAGAGACGCACTC-3’ and 5’-CTATGCGCCTTGCCAGCCCGCTCAGAGACGCACTC-3’ addition respectively. Underlined sequences represent 454 adapter sequences, and barcodes are indicated in bold. Amplicons from CRISPR-expressing cells and control cells were mixed in a 2:1 ratio and amplified on beads using the GS Junior Titanium emPCR Kit (Lib-A, Roche) and the GS Junior Titanium Sequencing kit (Roche). 454 sequencing was subsequently performed on a GS Junior System (Roche) and sequences were analyzed by using GS Amplicon Variant Analyzer software (Roche) and/or Seqman Pro software (DNAstar, Lasergene).
Supernatants of HSV-1-eGFP infected (MOI 0.5) MRC5 cells expressing gRNAs targeting UL57 (gRNA #1) and UL70 (gRNA #4) were harvested after 21 dpi. Supernatants from gRNA expressing cells containing anti-UL7 and anti-US11 gRNAs were isolated at 2 dpi since these cultures were completely lysed after several days. The genomic target sites of the respective gRNA were amplified by PCR and subjected to 454 sequencing analysis essentially as described in the previous paragraph. The gene-specific regions of primers used for the amplification were (5’-3’): US7-fw: TTTTCCGGTAAACCGAATTG; US7-rev: TCGCTACACGTGTGGAAGAC; US11-fw: CCTCTAACGAGCTCCACAGG; US11-rev: CCGACGTCACTAGATCACCA; UL57-fw: CGCACAGAGACGCCGAAATC; UL57-rev: AATTGCTGGGATCGTTGCGG; UL70-fw: ATGGTGCTGTACTGGCCCTC; and UL70-rev: GTGAACAACGAAACGCTGCAG. Only variants with insertions and deletions at the target sites were used for analysis in Fig 4A, as 454 sequencing cannot accurately assess single nucleotide substitutions.
Real Time quantitative Polymerase Chain Reaction (RT-qPCR) was performed using TaqMan universal PCR master mix (Applied Biossystems, Life Technologies) to quantify the relative amount of HSV-1 genomic DNA present in the supernatants from HSV-1 infected cells and EBV genomic DNA present in latently infected Akata-Bx1 cells. For HSV-1 qPCRs, supernatants were heat inactivated and subjected to triplicate TaqMan measurements using forward primer 5’-TTCTCGTTCCTCACTGCCTCCC-3’, reverse primer 5’-GCAGGCACACGTAACGCACGCT-3’, and FAM-TAMRA HSV-1 specific probe 5’-CGTCTGGACCAACCGCCACAC-3’. For EBV qPCR, genomic DNA was isolated using the Wizard genomic DNA isolation kit (Promega), which was subjected to qPCR using forward primer 5’-GGAACCTGGTCATCCTTTGC-3’, reverse primer 5’-ACGTGCATGGACCGGTTAAT-3’, and FAM-TAMRA TaqMan probe 5’-CGCAGGCACTCGTACTGCTCGCT-3’. Virus-specific TaqMan qPCRs were normalized to RNAse P via the commercial TaqMan Gene Expression Assay according to the instructions of the manufacturer (Applied Biossystems, Life Technologies). The relative viral DNA amount was calculated according to the comparative Ct method (User Bulletin number 2, ABI Prism 7700 Sequence Detection System, P/N 4303859). Relative genome content values for each condition were averaged from triplicate technical measurements and standard deviations were calculated from biological triplicate measurement of the same condition.
To determine the amount of HSV-1 genomes present in latent MRC5-Cas9 cells, genomic DNA was isolated using the Quick-gDNA Miniprep kit (Zymo Research, USA) from gRNA expressing latent cells pre-reactivation (see HSV-1 Latency generation) and subjected to qPCR essentially as described above. For each independently transduced HSV-1 latent MRC5-Cas9 sample the qPCR was performed in technical triplicate. Error bars and average Ct values were determined by calculating the standard deviation across all Ct values from each anti-HSV-1 gRNA biological replicates. All HSV-1 values were normalized to RNAseP levels. The ‘empty vector’ condition was normalized to 1.0.
HSV-1 plaque assays were performed in triplicate for each biological sample. 75,000 Vero cells were seeded in a 24 well plate well in complete DMEM and infected the following day with various dilutions of supernatant harvested from HSV-1 infected CRISPR-expressing Vero or MRC5 cells. After 3 hours incubation, cells were overlaid with 0.5% agarose (Seakem LE Agarose, Lonza AG, Switzerland) solution in complete DMEM and cultured for 3 days to allow the formation of plaques to occur. Afterwards, cells were fixed overnight at room temperature with 1% formaldehyde and stained using 0.5% crystal violet solution. After 3 washes with water, plates were allowed to dry and the numbers of plaques were counted by eye. Virus titers were calculated as plaque-forming units/ml (pfu/ml). If for a condition in the lowest dilution no plaques were detected, this was scored as not detected (ND). Images of plaques were taken using the EVOS FL Cell Imager (Life Technologies).
The HSV-2 quiescency model from Russell et al. [47, 48] was adapted for HSV-1 in MRC5 cells. In short, MRC5-Cas9 cells were seeded in multiple 6-wells plates 3 days prior infection at 210.000 cells/well. After 3 days, the cells were infected with 105k viral HSV-1-eGFP particles/well at 37°C for one hour. The infected MRC5-Cas9 cells were subsequently washed twice with complete medium and incubated in a CO2 incubator at 42°C for 4 days. The culture medium was replaced every day. After 4 days, the cells were returned to regular cell culture conditions (37°C 5% CO2). Next, cells were monitored by fluorescence microscopy for 6 days to exclude wells displaying signs of spontaneously reactivation, which occurred in ± half of the wells. Cells were subsequently harvested and plated at 150,000 cells/well in 12-wells plates. The next day, wells not displaying signs of spontaneous reactivation were transduced with gRNA-carrying lentivirus and selected for after 3 days using puromycin (2μg/ml) for the duration of 3 days. MRC5-Cas9 cells containing quiescent HSV-1-eGFP were counted and 40,000 cells/well (48-well plate) were seeded in triplicate in 48 well plates. The remainder was pelleted and subjected to gDNA isolation using the Quick-gDNA Miniprep kit (Zymo Research, USA). The gDNA was subjected to qPCR to determine the relative HSV-1 genome content. Quiescent MRC5-Cas9 cells were infected the next day with AD169 HCMV (ATCC VR538) to trigger HSV-1-eGFP reactivation. Three days later, cells were harvested, fixed with formaldehyde (1%), and measured by flow cytometry to assess the percentage of eGFP-positive cells as a measure for cells with replicating HSV-1-eGFP.
gRNA target sites were amplified from gRNA-expressing quiescent MRC5 cells using the following gene-specific primers: UL8-fw 5’-TAGAAATCCCGCAGCTCCGTC-3’, UL8-rev 5’-GGGGCGGTGAACTTTAGCAC-3’, UL29-fw 5’-GAGGGCGTCAGTTTCAGGGAC-3’, UL29-rev 5’-GATTCATTCCCCAACCCCGGTC-3’, UL52-fw 5’-GCGCGGATCATCTCATATTGTTCC-3’ and UL52–rev 5’- GACGAACATGGGTCGGGTTC-3’. Derived amplicons were isolated from 2% agarose gels by gel extraction (GeneJET PCR Purification Kit, ThermoFisher Scientific) and subjected to a second PCR reaction to increase product yield and add sequence tags (lower case sequence) for subsequent barcoding and Illumina adapter introduction: UL8_flank_fw 5’-tcgtcggcagcgtcagatgtgtataagagacagGGCGTTGCGACATACAAAATAC-3’, UL8_flank_rev 5’-gtctcgtgggctcggagatgtgtataagagacagTATAAGTCTCGGGACCGCACTC-3’, UL29_flank_fw 5’-tcgtcggcagcgtcagatgtgtataagagacagGTGCGAGAACCCACGACCAC-3’, UL29_flank_rev 5’- gtctcgtgggctcggagatgtgtataagagacagCTCGGGAGACATACCTTGTCG-3’, UL52_flank_fw -tcgtcggcagcgtcagatgtgtataagagacagTCTCATATTGTTCCTCGGGGCG-3’ and UL52_flank_rev 5’-gtctcgtgggctcggagatgtgtataagagacagTCTTCGAACCTGTCTTGCTCCG-3’. Expand High Fidelity PCR system (Roche) was used for all PCR reactions according to manufacturer’s instructions.
Amplified DNA was isolated from 2% agarose gels (GeneJET PCR Purification Kit, ThermoFisher Scientific), barcoded and prepared for Illumina MiSeq sequencing using the 16S Metagenomic Sequencing Library Preparation kit (Illumina). DNA concentrations were determined by the Qubit fluorometer 2.0 (Life Technologies, USA) with the Qubit dsDNA High Specificity assay kit. DNA libraries were sequenced by Illumina MiSeq (250bp paired-end). Obtained sequences were quality-checked with FASTQC v0.11.3 (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/) and trimmed with seqtk version 1.0-r31. Sequences were aligned to the viral reference sequences by using Bowtie2 version 2.2.6 [90] using the sensitive-local alignment mode. Alignments were converted to bam format and indexed with samtools version 1.3 [91] and analyzed in Tablet version 1.14.04.10 [92] for CRISPR/Cas9 specific indels at the gRNA target site.
|
10.1371/journal.pgen.1005008 | Elevated In Vivo Levels of a Single Transcription Factor Directly Convert Satellite Glia into Oligodendrocyte-like Cells | Oligodendrocytes are the myelinating glia of the central nervous system and ensure rapid saltatory conduction. Shortage or loss of these cells leads to severe malfunctions as observed in human leukodystrophies and multiple sclerosis, and their replenishment by reprogramming or cell conversion strategies is an important research aim. Using a transgenic approach we increased levels of the transcription factor Sox10 throughout the mouse embryo and thereby prompted Fabp7-positive glial cells in dorsal root ganglia of the peripheral nervous system to convert into cells with oligodendrocyte characteristics including myelin gene expression. These rarely studied and poorly characterized satellite glia did not go through a classic oligodendrocyte precursor cell stage. Instead, Sox10 directly induced key elements of the regulatory network of differentiating oligodendrocytes, including Olig2, Olig1, Nkx2.2 and Myrf. An upstream enhancer mediated the direct induction of the Olig2 gene. Unlike Sox10, Olig2 was not capable of generating oligodendrocyte-like cells in dorsal root ganglia. Our findings provide proof-of-concept that Sox10 can convert conducive cells into oligodendrocyte-like cells in vivo and delineates options for future therapeutic strategies.
| Developmental or acquired defects of oligodendrocytes or their myelin sheaths impairs saltatory nerve conduction in the central nervous system and thus leads to severe neurological diseases. Strategies to regenerate or replace these cells require a deeper understanding of the regulatory processes that underlie their generation during development. Here we show in a Sox10 overexpressing mouse model that increase of the levels of a single transcription factor during embryogenesis efficiently converts the already Sox10 expressing satellite glial cells of the peripheral nervous system into oligodendrocyte-like cells by a mechanism that does not simply recapitulate developmental oligodendrogenesis but involves direct Sox10-dependent induction of the oligodendroglial differentiation network. Our study identifies mechanisms that may help to convert other cell types into oligodendrocytes and thus prove eventually useful for therapies of myelin diseases.
| Transcription factor-mediated reprogramming is currently the method of choice for the generation of induced pluripotent stem (iPS) cells [1]. It is also used to directly convert one cell type into another. Successful conversion depends on the choice of transcription factors, but is also influenced by the proteomic constitution of the targeted cell with some cells being more susceptible to acquiring a specific new identity than others [2]. Both reprogramming and conversion are usually performed in culture with low efficiencies and are rarely studied in vivo.
Recently, murine fibroblasts have been converted into oligodendrocyte precursor cells (OPC) which in turn had the capacity to differentiate into myelinating oligodendrocytes when transplanted into the brain of a myelin-deficient mouse mutant [3,4]. This feat is important as generation of oligodendroglial cells from iPS cells is relatively inefficient and time-consuming [5]. Once optimized and adopted to human cells, it offers a potential source for cell replacement strategies in the various demyelinating and dysmyelinating diseases.
The conversion to OPC was achieved by applying a cocktail of several transcription factors to fibroblasts. While one study settled on a set of eight transcription factors with the core group consisting of Sox10, Olig2 and Nkx6.2 [3], the other defined a three-factor mix of Sox10, Olig2 and Zfp536 [4]. Sox10 and Olig2 thus seem to represent the minimal common denominator for the conversion process. The key role of Sox10 and Olig2 is not unexpected as previous studies had shown the exceptional importance of both transcription factors for oligodendroglial development and myelin formation during embryonic and postnatal development [6,7,8,9,10]. Olig2 is largely restricted to oligodendroglial cells. The few other Olig2-expressing cell populations (i.e. neuroepithelial cells of the ventral ventricular zone, motoneuron precursors and a subset of astrocyte precursors) are transient and restricted to the embryonic and early postnatal central nervous system (CNS) [11]. Sox10, in contrast, additionally occurs in several other cell types outside the CNS which are mostly neural crest-derived, such as all glial cells of the peripheral nervous system (PNS) [12].
When tested as single factors for their ability to induce OPC features in fibroblasts, only Sox10, but not Olig2 was found to exhibit some activity [4]. An independent study on cultured human neural progenitor cells recently confirmed Sox10 as the principle and rate-limiting determinant of myelinogenic fate [13]. This prompted us to postulate that it might be possible to induce oligodendrocyte properties in vivo in an especially conducive cell type with Sox10 alone. Indeed we found that its overexpression in already Sox10-positive satellite glia of PNS dorsal root ganglia (DRG) is sufficient to generate oligodendrocyte-like cells in vivo. The available evidence indicates that a key element in this conversion process is the activation of Olig2 as the second essential oligodendroglial identity factor mediated by a Sox10-responsive evolutionarily conserved enhancer of the Olig2 gene. Interestingly, analogous overexpression of Olig2 is not sufficient to convert satellite glia into oligodendrocyte-like cells. Our findings provide proof-of-concept that Sox10 can be used to convert a conducive cell type into oligodendrocyte-like cells in vivo and delineates options for future therapeutic strategies.
For targeted and strictly controlled Sox10 expression in vivo a transgene was generated in which rat Sox10 cDNA was placed under control of a bidirectional tetracycline-responsive promoter (Fig. 1A). GFP expression from the same promoter and Sox10-tagging with an aminoterminal 9myc epitope were used for detection of transgene expression. Luciferase reporter gene assays in transiently transfected Neuro2A cells confirmed that the aminoterminal 9myc tag did not interfere with the ability of Sox10 to activate a series of its targets, including the promoters of the Mag (myelin associated glycoprotein), Mbp (myelin basic protein), Cx32 (connexin 32), Cx47 (connexin 47) genes and the intronic oligodendrocyte enhancer of the Plp1 (proteolipid protein 1) gene (Fig. 1B).
Of the founders obtained by pronucleus injection of this TetSox10 transgene one was expanded into a line. It contained less than 5 tandem copies of the transgene on the long arm of mouse chromosome 10 (10qD1). In this study it was mostly combined with a Rosa26stopflox-tTA [14] and a Sox10::Cre [15] allele to direct transgene expression to all cells that normally express Sox10 during development or in the adult (Fig. 1A). Brain extracts from Rosa26+/stopflox-tTA Sox10::Cre mice contained as much transgenic as wildtype Sox10 at embryonic day (E) 18.5 when they were hemizygous for TetSox10, and approximately three times as much when homozygous (i.e. 2TetSox10) (Fig. 1C, D). Transgenic Sox10 expression corresponded to sites of endogenous Sox10 expression (Fig. 1E, F). In spinal cord and other CNS areas, both endogenous and transgenic Sox10 were restricted to and present in the vast majority of Olig2-positive cells of the oligodendroglial lineage (Fig. 1G, H). In the PNS, both DRG and nerves were labelled similarly by an anti-Sox10 antibody in the wildtype and an anti-myc-tag antibody in the transgenic animal (Fig. 1E, F, I, J). However, while endogenous Sox10 was restricted to glial cells as previously shown [16], transgenic Sox10 was additionally found in a subset of DRG neurons as a relic of their ontogenetic history (Fig. 1I, J). DRG neurons stem from Sox10-positive neural crest precursor cells and therefore experience transient Sox10::Cre expression which triggers induction of the TetSox10 transgene. The continued presence of transgenic Sox10 in cell lineages that normally express the protein only transiently and the resulting developmental defects may be one reason for the very early postnatal death of Sox10::Cre induced, tTA expressing TetSox10 and 2TetSox10 mice.
It had previously been shown that Sox10 deletion leads to cell loss and disorganization within the PNS, including a dramatic reduction of DRG size [16]. The CNS is less affected and mainly suffers from absent oligodendroglial differentiation and myelination [8,9]. When transgenic Sox10 is expressed homozygously on an otherwise Sox10-deficient background (i.e. in 2TetSox10 under control of Sox10rtTA in Sox10rtTA/rtTA mice [17] following doxycycline treatment), these defects are rescued as indicated by a near normal DRG size and reappearance of myelin gene expression in the spinal cord of compound mutant embryos (compare Fig. 1K-O to Fig. 1P-T and Fig. 1U-Y). This confirms functionality of the 9myc-tagged transgenic Sox10 in vivo.
When analysing oligodendroglial development in late embryos that overexpress TetSox10 and 2TetSox10 under control of Sox10::Cre and Rosa26stopflox-tTA we observed an earlier appearance of myelin markers such as Plp1 and Mbp in spinal cord and other CNS regions. This may be indicative of a precocious oligodendrocyte differentiation. More intriguingly, Plp1 and Mbp expressing cells were also detected in substantial numbers in DRG of 2TetSox10 embryos at E18.5, whereas they were rare in DRG of TetSox10 mice and absent from age-matched wildtype embryos (Fig. 2B-D, F-H). Myelin gene expression in DRG of 2TetSox10 embryos went along with the selective presence of Myrf, Olig2, Olig1 and Nkx2.2 (Fig. 2E, I-L, P-R). These transcription factors are strongly associated with oligodendrocytes and oligodendroglial myelination, while absent from Schwann cells, the only cell type normally capable of myelination in the PNS. In contrast, markers of myelinating Schwann cells such as Oct6 and Krox20 transcription factors were not expressed in DRG and remained restricted to the peripheral nerves of 2TetSox10 embryos (Fig. 2N, O, T, U).
Dissociated cells from DRG of E14.5 2TetSox10 mouse embryos also gave rise to Mbp-positive cells with the typical morphology of myelinating oligodendrocytes at low frequency when co-cultured with rat DRG neurons under myelinating conditions (Fig. 2W-Y). Such cells were not observed when dissociated DRG of E14.5 wildtype embryos were used instead (Fig. 2V). We therefore conclude that the myelinating cells in DRG of 2TetSox10 mice closely resemble oligodendrocytes. As we cannot exclude the possibility that these cells retain differences to oligodendrocytes and as it is technically not feasible for us to collect enough of these cells for in-depth expression profiling and characterization, we will refer to them as oligodendrocyte-like cells.
Despite the strong expression of oligodendrocyte lineage and differentiation markers, oligodendrocyte precursor cell (OPC) markers such as Sox9, Pdgfra and NG2 were not detected in substantial levels in DRG of 2TetSox10 mice (e.g. Fig. 2M, S). This leads to the conclusion that these oligodendrocyte-like cells have not gone through a classical OPC stage and may be derived from another cell source.
A study of consecutive embryonic stages from E11.5 to E18.5 (Fig. 3A-H) revealed that ectopic Olig2 expressing cells in DRG of 2TetSox10 mice are not yet detectable at E11.5, but are already present at E12.5 (Fig. 3E, F) approximately the same time when OPC start to be generated in the ventral ventricular zone and emigrate from the pMN domain into the marginal zone of the spinal cord (Fig. 3A, B, E, F). This early appearance strongly argues for an origin of Olig2-positive cells in DRG of 2TetSox10 mice that is independent from OPC and outside the CNS.
Considering that most of the PNS is neural crest-derived and that Wnt1::Cre is widely active throughout the early neural crest, we exchanged Sox10::Cre for this Cre driver to induce 2TetSox10 expression. Analysis of GFP autofluorescence as well as direct detection of transgenic Sox10 by anti-myc antibodies confirmed the widespread activation and expression of the transgenic construct throughout the embryonic PNS (Fig. 3I, J). It went along with efficient generation of differentiating oligodendrocyte-like cells in DRG as evident from the induced expression of Olig2, Nkx2.2, Myrf, Plp1 and Mbp (Fig. 3K-P). We therefore conclude that the oligodendrocyte-like cells stem from neural crest-derived cells of the PNS.
Because boundary cap cells represent a versatile source for different neural crest-derived cell types in the PNS [18] and have been reported to give rise to oligodendrocytes after engraftment into the CNS [19], we checked whether these cells were the source of Olig2-positive cells in the DRG. A Krox20::Cre driver in combination with Rosa26stopflox-tTA allows a restricted induction of 2TetSox10 in this transient cell population which is localized at the dorsal root entry zone during early embryonic times (Fig. 4A, see arrows). However, such selective induction of transgenic Sox10 expression did not lead to the appearance of Olig2-positive cells in DRG (Fig. 4E). Furthermore, Olig2-positive cells were never observed in substantial numbers in the dorsal root entry zone of mice in which 2TetSox10 expression was under control of Sox10::Cre at times when they were already numerous in the DRG (compare Fig. 5A-G to Fig. 5H-M). This argues against a boundary cap derived-origin of the ectopic Olig2-positive cells.
Brn4::Cre is active throughout the early CNS and in DRG neurons [20,21]. When this Cre line was used to activate 2TetSox10 expression (Fig. 4B, C) we again failed to observe any Olig2-positive cells in the DRG (Fig. 4F, G). This finding not only provides additional evidence for an origin of the Olig2-positive cells outside the CNS, it also excludes DRG neurons as source. The latter finding is also supported by the fact that there was no co-labelling of Olig2-positive cells with NeuN or Islet1 as markers for PNS neurons in DRG of mice in which 2TetSox10 expression was under control of Sox10::Cre (Fig. 5O-R, U-X). Instead, we observed a substantial overlap of Olig2 and Fabp7 staining in DRG of 2TetSox10 mice at E13.5 (Fig. 5N). Considering that Fabp7 is the only reliable early marker for peripheral glia, we conclude that Olig2 induction occurs in glial cells of the DRG. Interestingly, Fabp7 co-staining was only observed in cells with weak, but not with strong Olig2 immunoreactivity arguing that co-expression is transient and restricted to the phase of Olig2 induction.
Finally, we employed Dhh::Cre in combination with Rosa26stopflox-tTA to induce 2TetSox10 expression. Dhh::Cre is active in the Schwann cell lineage from the precursor stage onwards. Although we efficiently activated transgene expression in Schwann cells, for instance in spinal nerves in the immediate vicinity of the DRG (Fig. 4D, see arrowheads), no Olig2-positive cells were generated in the DRG itself (Fig. 4H). Considering (i) that the Olig2-positive cells are derived from PNS cells other than boundary cap cells, Schwann cells or DRG neurons and (ii) that they are glial in origin, satellite glia within the DRG remain as sole source. We thus conclude that overexpression of Sox10 in satellite glia leads to the generation of differentiating and myelinating oligodendrocyte-like cells. This conversion seems specific as we failed to obtain any evidence for a simultaneous generation of astrocytes or spinal cord neurons in DRG upon Sox10 overexpression in 2TetSox10 mice (Fig. 5, S, T, Y, Z)
It seemed reasonable to assume that one of the earliest events during the conversion of satellite glia into oligodendrocyte-like cells should be the Sox10-dependent activation of Olig2 as an essential determinant of oligodendroglial identity. During oligodendrocyte specification in the CNS, Olig2 is genetically upstream of Sox10 and appears to be a direct activator of Sox10 gene expression [22,23,24]. However, it has also been proposed that later on during oligodendrocyte development, Sox10 may in turn help to maintain Olig2 expression [25]. An increase of overall Sox10 levels in satellite glia upon transgene expression could thus be sufficient to activate Olig2 expression and thereby establish a key circuit of the oligodendrocyte regulatory network. To study this hypothesis, we searched for evolutionarily conserved non-coding regions (ECR) in the vicinity of the Olig2 gene. One such ECR was recently shown to be active in the early spinal cord, but was not analyzed at times relevant for oligodendrocyte development [26]. This 2.7 kb Olig2 ECR (OLE) is localized approximately 33 kb upstream of the transcriptional start of the mouse Olig2 gene (Fig. 6A). It furthermore exhibited a robust response to the presence of Sox10 in transiently transfected Neuro2A cells and allowed a 25-fold Sox10-dependent activation of a luciferase reporter gene (Fig. 6B). When split into a more distal (OLEa) and a more proximal part (OLEb), OLEa retained Sox10 responsiveness and even elicited an increased activation of the luciferase reporter, whereas OLEb failed to do so arguing that OLEa may contain the core elements for Sox10 induction. In agreement, chromatin immunoprecipitation (ChIP) experiments on three-week old mouse brain and oligodendrocytes differentiated in culture for 6 days found a specific enrichment of OLEa in chromatin precipitated with α-Sox10 antibodies (Fig. 6C, D) arguing that the effect of Sox10 on OLEa is direct.
Bioinformatic analysis of the OLEa sequence revealed the presence of 14 potential Sox binding sites, labelled I through XIV (Fig. 6E and Fig. 7). In electrophoretic mobility shift assays (EMSA) six sites were found to exhibit strong affinity for Sox10. These were sites I, II, IV, IX, X and XIV (Fig. 6F-P). Site V had a weaker affinity. Sites I and II were closely spaced and allowed binding of a Sox10 dimer. So did sites IX and X, whereas sites IV, V and XIV interacted with a Sox10 monomer (Fig. 6F, H, I, M, P). Each of the sites was mutated in such a way that Sox10 binding was no longer possible (Fig. 8A, C-R) and mutations for the high-affinity sites were introduced into the context of OLEa. Luciferase reporter gene assays in transiently transfected Neuro2A cells showed that mutation of the dimer site IX/X had the largest impact on Sox10 responsiveness among the single site mutations and reduced activation rates from 59-fold to 15-fold (Fig. 8B). The remaining activation rates were even further reduced when mutation of site IX/X was combined with additional mutations of the other sites such as site I/II and site IV. These in vitro studies therefore confirm that Sox10 binds and acts through multiple sites in OLEa.
To analyse whether the identified ECR is active as an oligodendroglial enhancer in vivo, we used lacZ reporter gene constructs containing the 2.7kb OLE or its subfragment OLEa in front of a hsp68 minimal promoter and the reporter gene cassette to generate transgenic mice (Fig. 9A). Five stably transmitting founders were obtained for the OLE-lacZ and the OLEa-lacZ reporter each (Fig. 9B). Despite some variability among the established lines (Fig. 9B), all exhibited staining in the spinal cord that was compatible with predominant expression in cells of the oligodendrocyte lineage (Fig. 10A-D for OLE-lacZ and Fig. 10E-H for OLEa-lacZ). Outside the CNS, there was weak transgene expression in the DRG, sometimes accompanied by faint staining in cartilage or vasculature (Fig. 9B).
IHC at E18.5 confirmed the predominantly oligodendroglial expression of the transgene as the majority of β-galactosidase expressing cells were positive for Olig2 and Sox10 at perinatal times (Fig. 10I, K). In contrast, only a small fraction of β-galactosidase expressing cells colabelled with NeuN as a neuronal marker (≤ 5%) or glutamine synthetase and GFAP as astrocytic markers (≤ 20%) (Fig. 10M-P). The substantial overlap between β-galactosidase and Mbp furthermore argues that reporter gene expression is not restricted to OPC but also found in differentiating oligodendrocytes (Fig. 10J, L). Nevertheless, there were some differences between OLE-lacZ and OLEa-lacZ lines (Fig. 10A-H). The OLE-lacZ reporter was on average more widely expressed throughout the oligodendroglial population than the OLEa-lacZ reporter. In addition to this better coverage, only OLE, but not OLEa was strongly active in the pMN domain (compare Fig. 10A, B to Fig. 10E, F). This confirms that OLEa contains the key regulatory elements for oligodendroglial activity, but may need to be modified by additional elements present in the larger OLE to faithfully recapitulate the complete developmental expression pattern of Olig2.
We also placed the OLE-lacZ and OLEa-lacZ transgenes on a background in which 2TetSox10 was expressed under Sox10::Cre-induced tTA control and investigated reporter gene activation in the DRG (Fig. 10Q-T”). Both transgenic constructs were strongly activated in a subpopulation of cells within the DRG of 2TetSox10 mice (Fig. 10R-R”, T-T”). In agreement with efficacy of transgene expression in the CNS, induction rates varied between the two transgenes and OLEa-lacZ transgenic animals reproducibly showed a lower amount of lacZ-expressing cells in their DRG than OLE-lacZ transgenic animals. Importantly, lacZ expression was restricted to a subset of the Olig2-expressing cells in the DRG (Fig. 10R, T). This supports the notion that Olig2 expression in DRG of Sox10 overexpressing mice involves the identified Olig2 enhancer.
Finally we asked whether the presence of Olig2 in DRG glia is sufficient to induce oligodendrocyte-like cells. For that purpose we exchanged the TetSox10 transgene by an analogously constructed TetOlig2 transgene [27] and probed the DRG of 2TetOlig2 mice at E18.5 for the expression of the myelin genes Mbp and Plp1 and Myrf as a marker for differentiating oligodendrocytes. Unlike 2TetSox10 mice, 2TetOlig2 mice were indistinguishable from the wildtype in that oligodendrocyte markers were not expressed (compare Fig. 11C-E, G-I to Fig. 2C-E, G-I). Therefore Olig2 cannot convert satellite glia into oligodendrocyte-like cells.
Using a genetic strategy that allows transgene expression in all cells that normally express Sox10, we have shown in this study that overexpression of Sox10 in DRG satellite glia is sufficient to directly convert these cells into cells that strongly resemble differentiating oligodendrocytes. Evidence that the reprogrammed cells are oligodendrocyte-like is manifold. These cells express the typical markers and regulatory network components of myelinating oligodendrocytes, including Olig2, its relative Olig1, Nkx2.2 and Myrf. Additionally we find expression of myelin genes such as Mbp and Plp1, and when cultured with DRG neurons, some of these cells acquire the typical morphology of myelinating oligodendrocytes. The presence of Plp1 furthermore shows that the cells are not Schwann cells, which express Mpz instead. Similarly, characteristic regulatory network components of myelinating Schwann cells such as Oct6 and Krox20 were missing from these cells.
Despite their clear oligodendrocyte character these cells were not of CNS origin as evidenced by the fact that CNS-specific overexpression of Sox10 failed to give rise to these cells. Their early appearance in DRG at E12.5 also argues against a CNS origin as it is difficult to imagine that the newly generated OPC could have migrated from the pMN domain all the way through the spinal cord parenchyma into the DRG during this extremely short time window. We also failed to detect substantial levels of Sox9, Pdgfra and NG2 in reprogrammed cells at any time of their development arguing that these cells did not go through a classic OPC stage.
The PNS origin of these oligodendrocyte-like cells was also supported by their appearance after neural crest-wide overexpression of Sox10. Among neural crest-specific cellular sources for these oligodendrocyte-like cells within the PNS boundary cap cells and immature Schwann cells could as much be ruled out as DRG neurons. Instead, Fabp7-positive resident glia within the DRG were identified as the cells in which Olig2 induction occurred. Fabp7 is to date the only reliable marker for satellite glia during embryonic development [16]. We therefore conclude that the oligodendrocyte-like cells arise from satellite glia. However, we are aware that this assignment is based on a single marker and should be revisited once additional markers for embryonic satellite glia become available.
Satellite glia represent a poorly characterized cell population. They are closely apposed to neuronal somata and appear to supply them with nutrients, neurotrophins and other essential molecules. Their intense communication with neurons and strong coupling by gap junctions has led to the assumption that they may be the PNS counterpart to CNS astrocytes [28]. Satellite glia furthermore appear to represent a persistent precursor cell population. They are slowly dividing in the adult and respond to noxious stimuli and inflammation by enhanced proliferation [29]. When taken from their normal environment and placed in culture they have been reported to display plasticity and give rise to various types of PNS and CNS glial cell types [30].
Our finding that satellite glia are prone to reprogramming may thus at least in part be attributable to their precursor cell characteristics and plasticity. To us, the frequency with which satellite glia are converted into oligodendrocyte-like cells upon Sox10 overexpression is particularly noteworthy. With standard in vitro conversion rates (for review, see ref. 2), we would have had little chance to observe this process in vivo. One reason for this phenomenon may actually be found in the microenvironment of satellite glia, including their close proximity to neurons which may supply instructive signals for oligodendrocyte development. However, it is probably also important that satellite glia already express some amount of endogenous Sox10. Considering that Sox10 may function as a pre-patterning factor [31,32], its presence may help to keep those chromatin regions in a poised state that need to be activated during the direct conversion of satellite glia into oligodendrocyte-like cells. It is this activity as pre-patterning factor that makes Sox10 especially suitable for reprogramming strategies. A valuable further property of Sox10 may be its capacity to induce many of the factors that it needs to cooperate with during oligodendroglial cell fate decisions and differentiation processes such as Nkx2.2 and Myrf [8,25].
Equally noteworthy is the fact that reprogramming is achieved by a change of dose rather than introduction of a novel factor. Sox10 amounts are tightly regulated and its functions are concentration-dependent during normal development in mouse and human [16,33,34,35].
One of the results of the increased Sox10 levels in satellite glia is the additional activation of Olig2 as a second essential factor for oligodendrogenesis and oligodendrocyte differentiation. This activation furthermore appears to be direct and mediated by an ECR in the distal upstream region of the Olig2 gene which is not only active in oligodendroglial cells, but also responds to the presence of Sox10 and is bound by this factor in vitro as well as in vivo. The multiplicity of Sox10 binding sites and the complicated structure of a core and adjacent accessory elements makes this ECR an ideal element for a Sox10 dosage-dependent enhancer that normally comes under Sox10 control in oligodendrocytes when amounts of this transcription factor increase with the onset of differentiation [9], or artificially in satellite glia when Sox10 levels increase by overexpression. This ability of high levels of Sox10 to induce and maintain Olig2 expression is likely a central element in the conversion process as it establishes a key circuit in the corresponding regulatory network. However, our results also indicate that Olig2 induction is not sufficient to convert satellite glia into oligodendrocyte-like cells. This argues that additional Olig2-independent processes are set in motion by high Sox10 levels in satellite glia. The previously reported induction of Myrf expression may be one of them [8].
It is intriguing to assume that other Sox10 expressing neural crest-derived cells with precursor cell characteristics may similarly be convertible into oligodendrocyte-like cells. These include melanocyte stem cells and enteric glia [36]. Especially the latter are similar to DRG satellite glia in their close apposition and functional interaction with neurons, as well as in their maintenance of precursor cell characteristics and plasticity that allows enteric glia to respond to injury with increased proliferation and production of enteric neurons [37]. The presence of melanocyte stem cells and enteric glia in the adult and their relatively easy accessibility may make them amenable to isolation and Sox10-dependent conversion as a realistic source of oligodendrocytes for future applications like cell replacement strategies.
All Plasmids were generated by standard cloning procedures. Expression plasmids for 9myc-tagged Sox10 were based on pCMV5 and pBI-EGFP. Reporter plasmids for transgenic animals were generated by cloning the respective ECR fragments upstream of an Hsp68 minimal promoter followed by a lacZ cassette [38]. For luciferase assays the respective ECR fragments were cloned upstream of a β-globin minimal promoter followed by a luciferase cassette [22], Sox binding sites were mutated using the QuikChange XL site-directed mutagenesis kit (Stratagene). Mouse Neuro2a neuroblastoma cells and rat primary oligodendroglia were kept in culture as described [39]. Transient transfections of Neuro2a cells, luciferase assays and EMSA followed standard procedures [22]. ChIP was performed as reported [31] with the following modifications: Chromatin was prepared from primary oligodendrocytes kept under differentiating conditions for six days and from brain tissue of three week old mice. Fixation was with 1% formaldehyde in PBS. For precipitation of sheared chromatin, anti-Sox10 antiserum and corresponding preimmune serum were used in combination with protein G magnetic beads (Cell Signaling Technology). A list of primers for cloning and detection of genomic fragments in PCR experiments, including their sequence and position is available upon request.
All animal experiments were carried out with permission and in compliance with animal policies of the local authorities and governmental agencies. Mice transgenic for TetSox10, OLE-lacZ or OLEa-lacZ were obtained by microinjecting the respective linearized DNA into male pronuclei of fertilized oocytes according to standard techniques. Mice transgenic for TetOlig2 have been described [27].
Expression of TetSox10 was achieved by combining one or two copies of the transgene with the Sox10rtTA allele and administration of doxycycline [17]. Alternatively, TetSox10 expression was induced by a combination of the Rosa26stopflox-tTA allele [14] and any of the following Cre alleles: Sox10::Cre [15], Wnt1::Cre [40], Krox20::Cre [41], Brn4::Cre [20] or Dhh::Cre [42]. If not otherwise stated, analysed animals contained two copies of the TetSox10, and one copy of the Rosa26stopflox-tTA and the Cre allele each. Expression of the TetOlig2 transgene was similarly achieved by combining two copies with one copy of the Rosa26stopflox-tTA and the Sox10::Cre allele.
After genotyping, material from staged embryos was processed for X-Gal staining [9], ISH with probes specific for Mbp, Plp1 and Myrf, or IHC using primary antibodies against Sox10 (guinea pig antiserum in 1:1000 dilution) [43], Sox9 (guinea pig antiserum in 1:500 dilution) [44], Glast (guinea pig antiserum in 1:500 dilution, Millipore), Olig1 (rabbit antiserum in 1:10000 dilution, Millipore), Olig2 (rabbit antiserum in 1:1000 dilution, Millipore), Oct6 (rabbit antiserum in 1:2000 dilution) [31], Fabp7 (rabbit antiserum in 1:300 dilution, Millipore), Krox20 (rabbit antiserum in 1:200 dilution, Covance), Mbp (rabbit antiserum in 1:200 dilution, NeoMarkers), β-galactosidase (rabbit antiserum in 1:500 dilution, ICN; goat antiserum in 1:500 dilution, Biotrend), myc-tag (goat antiserum in 1:200 dilution, Abcam), Nkx2.2 (mouse monoclonal in 1:5000 dilution, Developmental Studies Hybridoma Bank, University of Iowa), NeuN (mouse monoclonal in 1:500 dilution, Millipore), Gfap (mouse monoclonal in 1:100 dilution, Millipore), GlnS (mouse monoclonal in 1:1000 dilution, BD Transduction Laboratories), Hb9 (mouse monoclonal in 1:50 dilution, Developmental Studies Hybridoma Bank), Islet1 (mouse monoclonal in 1:1000 dilution, Developmental Studies Hybridoma Bank), and GFP (rat monoclonal in 1:2000 dilution, Nacalai Tesque). Olig1 and Nkx2.2 immunoreactivity were detected with the TSA Plus Cyanine 3 system (PerkinElmer). Source and working concentration of fluorophore-labelled secondary antibodies were as described [10,21,39,45]. Nuclei were counterstained with Dapi.
Coculture experiments were performed as described [46] with the difference that dissociated DRG cells from wildtype or 2TetSox10 mice were added instead of OPC to cultured rat DRG neurons. To this aim, DRG were dissected from E14.5 mouse embroys and dissociated with papain (4 U/ml), DnaseI (40 μg/ml) and L-cysteine (240 μg/ml) at 37°C for 60 min. The dissociated cells were added at a density of 150,000 per well in a 12-well plate and incubated under myelinating conditions for 4 weeks. After fixation, cells were stained with antibodies directed against Mbp (rat monoclonal in 1:750 dilution, Serotec) and Nf165 (mouse monoclonal in 1:3000 dilution, Developmental Studies Hybridoma Bank).
For western blots, brain extracts were prepared as described [47], and proteins were detected with antibodies against Sox10 and Gapdh (Santa Cruz Biotechnology).
Mice experiments were in accord with animal welfare laws and approved by the responsible local committees and government bodies (Regierung von Mittelfranken and Behörde für Gesundheit und Verbraucherschutz Hamburg).
|
10.1371/journal.pntd.0002488 | Barriers to Treatment Access for Chagas Disease in Mexico | According to World Health Organization (WHO) prevalence estimates, 1.1 million people in Mexico are infected with Trypanosoma cruzi, the etiologic agent of Chagas disease (CD). However, limited information is available about access to antitrypanosomal treatment. This study assesses the extent of access in Mexico, analyzes the barriers to access, and suggests strategies to overcome them.
Semi-structured in-depth interviews were conducted with 18 key informants and policymakers at the national level in Mexico. Data on CD cases, relevant policy documents and interview data were analyzed using the Flagship Framework for Pharmaceutical Policy Reform policy interventions: regulation, financing, payment, organization, and persuasion. Data showed that 3,013 cases were registered nationally from 2007–2011, representing 0.41% of total expected cases based on Mexico's national prevalence estimate. In four of five years, new registered cases were below national targets by 11–36%. Of 1,329 cases registered nationally in 2010–2011, 834 received treatment, 120 were pending treatment as of January 2012, and the treatment status of 375 was unknown. The analysis revealed that the national program mainly coordinated donation of nifurtimox and that important obstacles to access include the exclusion of antitrypanosomal medicines from the national formulary (regulation), historical exclusion of CD from the social insurance package (organization), absence of national clinical guidelines (organization), and limited provider awareness (persuasion).
Efforts to treat CD in Mexico indicate an increased commitment to addressing this disease. Access to treatment could be advanced by improving the importation process for antitrypanosomal medicines and adding them to the national formulary, increasing education for healthcare providers, and strengthening clinical guidelines. These recommendations have important implications for other countries in the region with similar problems in access to treatment for CD.
| Chagas disease is a vector-borne disease caused by the parasite Trypanosoma cruzi. The disease is most frequently transmitted by triatomine insects but can also be passed through blood donation or from mother to child at birth. Experts estimate that 8 million people are infected with Chagas disease globally and that 1.1 million of these infections are found in Mexico. Most public health programs for Chagas disease focus on preventing new infections through vector control and screening the blood supply. However, in recent years there has been a greater focus on treating the disease with one of two available medications, benznidazole or nifurtimox. This study explores access to these two drugs in Mexico. The study shows that less than 0.5% of those who are infected with the disease received treatment in Mexico in years. The study also identified important factors that limit access in Mexico, including the exclusion of both drugs from the national health insurance program and problems importing these medications. Finally, the paper suggests ways that these problems can be overcome in Mexico, while providing helpful insight for other countries that struggle with similar problems in treating this disease.
| Chagas disease is a vector-borne, parasitic disease with a prevalence of 8 million infections globally. The disease is responsible for as many as 15,000 deaths per year [1], [2], largely concentrated among the poor in Latin America, and a recent study found that the disease is also responsible for substantial losses in productivity and a large economic burden, especially in high prevalence countries [3]. Trypanosoma cruzi, the etiologic agent of Chagas disease, is most often transmitted by contact with infected triatomine insects, though transmission can also occur congenitally and through blood transfusion or organ transplantation [4]. In 2009, it was estimated that less than 1% of those infected with T. cruzi received treatment for the disease globally [5].
According to prevalence estimates for 2006 from the World Health Organization [6], approximately 1.1 million people are infected with T. cruzi in Mexico. However, limited published information exists on how many patients receive treatment in Mexico and what obstacles may hinder access to treatment. This study sought to determine for Mexico: (1) the extent of treatment access for Chagas disease; (2) the national level barriers to access to treatment for Chagas; and (3) strategies that could be used to overcome these barriers and increase access to treatment for Chagas disease.
This study uses an existing health systems framework, the Flagship Framework for Pharmaceutical Policy Reform [7], to analyze the barriers to treatment access for Chagas disease in terms of five policy interventions – regulation, financing, payment, organization, and persuasion. Based on this analysis, we also suggest strategies to increase access.
Chagas disease is clinically manifested in two stages – an acute stage and a chronic stage. The acute stage lasts for approximately 4–8 weeks and is characterized by flu-like symptoms or a characteristic local swelling at the site of parasite entry [8], [9], following which an infected person enters the indeterminate form of the chronic phase of infection. Among those with the indeterminate chronic form, about 20–30% of patients progress to the chronic cardiac or digestive forms of Chagas disease [10]. The most common course of Chagasic cardiomyopathy includes conduction system abnormalities early in the disease, resulting in heart failure. In all phases, serological tests such as the enzyme-linked immunosorbent assay (ELISA) test, the indirect haemagglutination assay (IHA), and the indirect immunofluorescent antibody test (IIF) are used for diagnosis [4], [9], [11]. Because these tests can be difficult to interpret, the WHO recommends the use of two concomitantly positive tests to make a confirmed diagnosis [11], [12].
Currently, benznidazole and nifurtimox are the only antitrypanosomal medicines available to treat T. cruzi infection. Antitrypanosomal therapy is strongly recommended by WHO for acute, congenital or reactivated infections, and for chronic infection in children under the age of 18 [13], [14], [15]. Recent scientific evidence about the clinical effectiveness of these medications has led to the expansion of treatment indications to include adults in the chronic phase of the disease without advanced cardiomyopathy [1], [11], [16], [17], [18], [19]. Though no randomized controlled trial has directly compared the two medications [11], WHO guidance and the clinical literature place greater emphasis on the use of benznidazole [4] as a first-line therapy because there is more clinical evidence for its efficacy, and it has a more favorable side-effect profile and is better tolerated by adult patients [9], [15], [16], [17], [18], [20]. A randomized clinical trial of benznidazole is underway to determine its efficacy in slowing progression of disease among patients with early to moderate stage Chagasic cardiomyopathy [21], [22].
Both benznidazole and nifurtimox have undergone changes to their global supply chains over the past decade. Benznidazole was manufactured by Roche until 2003, at which time the rights and manufacturing technology were transferred to the Pernambuco state pharmaceutical laboratory in Brazil, Laboratorio Farmaceutico do Estado Pernambuco (LaFepe) [23], [24]. Between 2004 and 2006, LaFepe produced several batches of benznidazole using active pharmaceutical ingredient that was donated by Roche [24]. Then, after a period of no production, LaFepe resumed production of benznidazole in late 2011 and the medicine is now distributed by several entities including LaFepe, WHO, and Masters Pharmaceuticals. Nifurtimox is manufactured by Bayer HealthCare in El Salvador. In 2007 Bayer reached an agreement with WHO for Bayer to donate nifurtimox to WHO and for WHO to distribute the medicine through the WHO-Bayer Nifurtimox Donation Program [25].
Access to treatment for Chagas disease in Mexico must be considered in the context of the Mexican health system and its recent reforms. Mexico has three major national insurance schemes, the Instituto Mexicano del Seguro Social (IMSS), Instituto de Seguridad y Servicios Sociales de los Trabajadores del Estado (ISSSTE), and Seguro Popular (SP) [26]. IMSS and ISSSTE together offered coverage to approximately 42.6 million private sector (IMSS) and public sector (ISSSTE) employees in 2010 [26]. As of 2011, SP, a social health insurance program started in 2003, offers a package of 284 essential services to approximately 51.8 million Mexicans, according to the Mexican government [10], [27], [28], [29]. Affiliation with SP requires a fixed family contribution that is based on a progressive scale by income, though individuals and families who fall in the lowest two income deciles are exempt from payment of a premium [26], [27].
The national Program on Onchocerciasis, Leishmaniasis and Chagas Disease within the Mexican Secretary of Health's National Center for the Prevention and Control of Diseases (CENAPRECE) is the unit responsible for establishing guidelines and coordinating national activities for Chagas disease control. The State Secretaries of Health report patients who are diagnosed by ISSSTE, IMSS and SP systems to the national Program, which then turn provides medicines to treat confirmed cases. Figure 1 shows the process of case registration for a patient with Chagas disease.
IRB exemption was obtained from Harvard School of Public Health (Protocol# 21514-101) and the National Institute for Public Health (INSP) located in Cuernavaca, Mexico. Oral informed consent was obtained from all interviewees.
Table 2 provides a list of national level obstacles to treatment access for Chagas disease, based on our analysis of data collected in this study. The list includes all obstacles that were mentioned during interviews and could be triangulated using a second data source.
This study provides evidence regarding the extent of treatment access for Chagas disease in Mexico and the barriers that influence the level of access. In particular, the study demonstrates that the number of Chagas disease cases registered at the national level in Mexico since 2007 is approximately 0.41% of expected cases and that 120 registered, eligible cases were awaiting treatment at the time of the study. These findings also indicate that Mexico has made an effort to register new cases and provide treatment at both the state and national level and thus show an increased commitment to addressing this disease in Mexico.
Our findings also demonstrate that epidemiologic surveillance for Chagas disease remains a challenge in Mexico and that the complexity of the case registration system may delay or limit registration. Evidence from national data shows that problems in the supply chain of medicines make it difficult to ensure timely access to treatment as cases are registered and further, that the medicine provided by the national program since 2009 has exclusively been nifurtimox, a medicine that has been identified in the clinical literature and international guidelines as second-line therapy [20].
The lack of awareness and understanding of the disease and its treatment among both physicians and populations at risk was another important challenge related to the persuasion policy intervention area [46]. Patient and provider awareness of the disease has implications for efforts to strengthen epidemiologic surveillance and the willingness of physicians to treat infected patients when medicines are available. Additionally, access to treatment for Chagas disease has until 2012 been further weakened by its exclusion from the package of health interventions that are covered under SP [27], [28]. While its addition to the CAUSES in 2012 represents an important step (organization) toward increasing access to treatment, clinical information about the disease is still lacking in this document and neither benznidazole nor nifurtimox is listed as a treatment for the diseases in this category.
In addition to these barriers, it is important to acknowledge the role of international actors and policies as barriers to access to treatment for Chagas disease in Mexico and potentially in other countries as well. The global shortage of benznidazole in 2011 and the challenges in obtaining nifurtimox through WHO exist outside the Mexican context but directly affect efforts by the Mexican national and state control programs to increase access to treatment [23], [24].
These findings provide new information on the state of treatment for Chagas disease in Mexico and the barriers that prevent more widespread access. Previous work on this subject has suggested that efforts to control and treat Chagas disease in Mexico are insufficient [36], [53] but no study has previously measured the gap in access to treatment or analyzed related obstacles. In addition, a recent study estimated the economic burden associated with Chagas disease to exceed seven billion dollars globally and several studies have described the need for increased treatment globally [3], [5], [33], [53], [54]. This study is one of the first to examine the multiple complex factors within the health system that prevent more widespread treatment access in a particular country setting. It is important to note, however, that the state of Morelos did successfully procure benznidazole and offers an important case for showing how a state can take significant initiative in improving access to treatment for Chagas disease.
Some of the findings from the Mexican experience may be relevant to treatment access for Chagas disease in other countries in the region. For instance, reliance on nifurtimox as a first-line therapy in both the 2010 Mexican guidelines for vector-borne diseases and in procurement of medicines at the national level raises questions about the reasons for this choice and whether other countries may also choose to procure nifurtimox through the donation program now or in the future instead of purchasing benznidazole through the private market. In the case of Mexico, the regulatory status of the drugs, especially the lack of commercial permits for them, and the exclusion of antitrypanosomal therapies for Chagas disease from the Mexican national formulary have severely limited sources of financing to buy benznidazole, causing the national program to instead rely on the free nifurtimox. However, little information exists about whether other countries also rely on nifurtimox as a first-line therapy and if so, why. Though clinical guidelines overwhelmingly suggest that benznidazole is better tolerated and that the clinical evidence of its efficacy is more robust, clear international consensus guidelines for the treatment of Chagas disease have not been published and relatively limited data are available about the use and clinical outcomes for the two drugs by different countries around the world.
There are several limitations to this study. First, data on the prevalence of Chagas disease are limited both in Mexico and globally. This constitutes an important challenge to efforts to address this disease in Mexico. In this analysis, we use the official 2010 prevalence estimate from the Mexican Secretary of Health because it is more conservative than the most recent WHO estimate and because the WHO estimate does not have a clear evidence base. This choice may result in our analysis showing greater access to treatment (as a proportion of total infected cases) than may actually exist in Mexico. Some actors within the Mexican Secretary of Health have argued that the epidemiology of Chagas disease in Mexico is focal and that states with a high burden of disease should undertake activities to address this disease at a state level, while others have maintained that the prevalence of Chagas disease is substantial across much of the country and that the disease should be a national priority, especially given the migration of populations from endemic areas both within Mexico and from neighboring countries to Mexico [36], [50]. To provide a more reliable estimate of national prevalence, a nationally representative epidemiologic survey could be conducted, both nationally and by state. This would advance efforts by both the state and national programs to make more informed decisions about the priority and resources that are warranted for Chagas disease treatment.
A second limitation is that we consider benznidazole as the first line antitrypanosomal medicine, despite the lack of definitive international consensus on this issue. We made this decision because benznidazole is being used exclusively as the reference treatment regimen in clinical trials of new drugs, is named as the first line therapy in the treatment guidelines of several non-governmental organizations [20], and is cited as such in the vast majority of the clinical literature [9], [17], [18]. It is worth noting, however, that there is some diversity on treatment regimens within Mexico. Although the national program has used nifurtimox from the WHO donation program, the state of Morelos in Mexico has purchased benznidazole for its treatment program. Morelos registered 263 cases between 2007 and 2011, and treated 148 cases with benznidazole and 4 with nifurtimox.
This study was also limited by lack of data availability at the national and global levels. At the national level in Mexico, this included a lack of national treatment guidelines or data prior to 2010, a dearth of information about treatment eligibility or patient refusal of treatment, and a lack of data on treatment dose, completion or clinical outcomes. In particular, it was difficult to determine what proportion of patients would be treatment eligible according to the guidelines given that no data were available on co-morbidities or patient clinical history that would allow a more thorough analysis of patients in whom treatment may be contraindicated. Furthermore, there is limited evidence about access to treatment in other countries to provide a comparison for assessing Mexico's achievement in this area. Of note, however, a recent study estimated that less than 1% of those infected with T. cruzi receive treatment globally, suggesting that the extent of access in Mexico is likely to be similar in other countries [5].
Based on these findings, there are three important strategies that could be undertaken to increase access to treatment for Chagas disease in Mexico.
First, under regulation, an effort could be made to ease the importation process for these drugs. Ideally, this could be accomplished by securing COFEPRIS approval for both medicines and adding them to the national formulary, which could require actions by the relevant producers of benznidazole and nifurtimox. However, as noted above, benznidazole and nifurtimox are not approved by the United States Federal Drug Administration or the European Medicines Agency, in part because full clinical trials have not been completed for either drug. This lack of approval from two leading regulatory bodies may affect the willingness of other national regulatory bodies to approve the medicines. That said, both medications are included on the WHO Essential Medicine List [55]. In addition, clinical evidence continues to accumulate in favor of these drugs and efforts by institutions such as the Drugs for Neglected Diseases Initiative are being made to register the drugs in countries such as Colombia, Paraguay and Bolivia. In other contexts, alternative regulatory approaches such as investigational protocols are being utilized to make the drugs available [18]. Also with respect to regulation, countries with a high burden of Chagas disease may consider instituting laws that mandate rigorous epidemiologic surveillance and health education as well as prevention, diagnosis and treatment of the disease. For instance, Argentina offers a model for such legislation in National Law No. 26281. This law requires, among other things mandatory diagnostic testing and reporting for Chagas disease in all pregnant women and in newborns in the first year of life born to infected mothers.
Second, under persuasion, efforts could be expanded to provide disease-specific health education programs on Chagas disease for physicians, healthcare providers and populations at risk. Increased awareness of the disease and a better understanding of appropriate treatment methods is a critical aspect of strengthening case registration and access to treatment. In addition, health education activities have been emphasized in other national control programs such as those in Guatemala [56] and the Southern Cone initiative and have been used alongside vector control to increase awareness of the disease in high risk communities and among physicians and health workers. Increased awareness of the disease and of treatment methods is a critical aspect of strengthening case registration and access to treatment. Given the importance of this programming, the WHO and PAHO also play a potentially important role in terms of encouraging these programs and providing guidance on their design and implementation.
Third, under organization, it is important to strengthen existing guidelines in Mexico for the diagnosis and treatment of Chagas disease and information availability about the supply chains for these two medicines. This includes the addition of a clinical description of Chagas disease and the two medicines to its entry in the CAUSES and the creation of a clinical guide for diagnosis and treatment as this information is critically important to strengthen awareness of treatment for Chagas disease and information for practitioners about how to diagnose and treat the disease. In addition, better public reporting of medicines released and used at the state, national and global levels is needed.
In conclusion, this study found that access to treatment for Chagas disease in one high burden country (Mexico) is limited in important ways and identified three critical obstacles to treatment access: regulatory barriers to importation, a lack of understanding of the disease and its treatment, and a dearth of clinical guidelines [5]. Several of these barriers are likely to affect access in other countries as well, especially the lack of regulatory approval and registration of benznidazole and nifurtimox and the lack of publically available information on their supply chains. Finally, the study proposed a series of actions that could be taken in Mexico, based on a general analytical framework, to improve access to treatment for Chagas disease. These recommendations have important implications for other countries in the region with similar problems in access to treatment for Chagas disease.
|
10.1371/journal.ppat.1005685 | Biosynthesis of Antibiotic Leucinostatins in Bio-control Fungus Purpureocillium lilacinum and Their Inhibition on Phytophthora Revealed by Genome Mining | Purpureocillium lilacinum of Ophiocordycipitaceae is one of the most promising and commercialized agents for controlling plant parasitic nematodes, as well as other insects and plant pathogens. However, how the fungus functions at the molecular level remains unknown. Here, we sequenced two isolates (PLBJ-1 and PLFJ-1) of P. lilacinum from different places Beijing and Fujian. Genomic analysis showed high synteny of the two isolates, and the phylogenetic analysis indicated they were most related to the insect pathogen Tolypocladium inflatum. A comparison with other species revealed that this fungus was enriched in carbohydrate-active enzymes (CAZymes), proteases and pathogenesis related genes. Whole genome search revealed a rich repertoire of secondary metabolites (SMs) encoding genes. The non-ribosomal peptide synthetase LcsA, which is comprised of ten C-A-PCP modules, was identified as the core biosynthetic gene of lipopeptide leucinostatins, which was specific to P. lilacinum and T. ophioglossoides, as confirmed by phylogenetic analysis. Furthermore, gene expression level was analyzed when PLBJ-1 was grown in leucinostatin-inducing and non-inducing medium, and 20 genes involved in the biosynthesis of leucionostatins were identified. Disruption mutants allowed us to propose a putative biosynthetic pathway of leucinostatin A. Moreover, overexpression of the transcription factor lcsF increased the production (1.5-fold) of leucinostatins A and B compared to wild type. Bioassays explored a new bioactivity of leucinostatins and P. lilacinum: inhibiting the growth of Phytophthora infestans and P. capsici. These results contribute to our understanding of the biosynthetic mechanism of leucinostatins and may allow us to utilize P. lilacinum better as bio-control agent.
| Purpureocillium lilacinum, a well-known bio-control agent against various plant pathogens in agriculture, can produce antibiotic leucinostatins—peptaibiotic with extensive biological activities, including antimalarial, antiviral, antibacterial, antifungal, and antitumor activities, as well as phytotoxic. We have sequenced the genomes of two P. lilacinum isolates, and compared them with other fungi, focusing on their bio-control characteristics. We discovered a rich repertoire of CAZymes, proteases, SMs and pathogenesis related genes. We also identified a gene cluster containing 20 genes involved in the leucinostatins A and B biosynthesis by gene deletion, qRT-PCR and RNA-seq analyses. A transcription factor in the pathway was overexpressed, resulting in the upregulation of the related genes and a 1.5-fold increase in leucinostatins A and B. A new bioactivity of leucinostatins, inhibition of the growth of the notorious Phytophthora, was identified in this study by confronting incubation with P. lilacinum. These results provided new strategies for the agricultural development of leucinostatins and improving P. lilacinum strains.
| Plant parasitic nematodes with wide host ranges cause enormous crop and economic losses amounting to $157 billion annually worldwide [1, 2]. Biological control by fungi has become increasingly popular due to nematicides’ risks of environmental toxicity and adverse effects on human health [3]. One of the most promising and commercialized agents, Purpureocillium lilacinum, has been evaluated to assess its bio-control activity against plant nematodes in a number of studies [2, 4]. In particular, P. lilacinum has been reported to effectively control such species as the cotton aphid Aphis gossypii [5], the greenhouse whitefly Trialeurodes vaporariorum, the glasshouse red spider mite Tetranychus urticae [6], and the leaf-cutting ant Acromyrmex lundii [7].
The genus Purpureocillium was recently proposed for of Ophiocordycipitaceae, based on the internal transcribed spacer (ITS) and translation elongation factor 1-α (TEF) sequences of P. lilacinum, although it was originally classified in the genus Paecilomyces [8]. P. lilacinum is commonly isolated from soil, plant roots, nematodes and insects, and it occasionally infects people. This fungus employs flexible lifestyles, including soil-saprobes, plant-endophytes and nematode pathogens. Opportunistic infection occurs when nematode eggs encounter P. lilacinum; therefore, parasitism can be a mechanism for nematode bio-control (Fig 1A). It has now been confirmed that a serine protease [9], a cuticle-degrading protease [10] and chitinase [11] play important roles in infection by degrading nematode eggshells.
Recently, the production of SMs has been shown to be a mechanism that kills nematodes. For example, culture filtrates of P. lilacinum, in which leucinostatins were produced, caused strong mortality and inhibited nematode reproduction [12]. In addition to leucinostatins, a few other SMs have been isolated from P. lilacinum. The novel pyridone alkaloid paecilomide, an acetylcholinesterase inhibitor, was produced when this fungus was co-cultured with Salmonella typhimurium [13]. Two xanthone-anthraquinone heterodimers, acremoxanthone C and acremonidin A, were isolated in the course of a search for calmodulin ligands [14].
The leucinostatins (Fig 1B) are a family of lipopeptide antibiotics isolated from P. lilacinum [15], Paecilomyces marquandii [16–18] and Acremonium sp. [19]. Leucinostatin A contains nine amino acid residues, including the unusual amino acid 4-methyl-L-proline (MePro), 2-amino-6-hydroxy-4-methyl-8-oxodecanoic acid (AHyMeOA), hydroxyleucine (HyLeu), α-aminoisobutyric acid (AIB), β-Ala, and a 4-methylhex-2-enoic acid at the N-terminus as well as an N1,N1-dimethylpropane-1,2- diamine (DPD) at the C-terminus. Twenty-four different structures have been described in the leucinostatin series[20]. Leucinostatin A significantly suppressed prostate cancer growth in a coculture system in which prostate stromal cells stimulated the growth of DU-145 human prostate cancer cells through insulin-like growth factor I [21]. When screening for antitrypanosomal compounds among several peptide antibiotics, leucinostatins showed the most potent activity against trypanosomes. Trypanosome infection causes human African trypanosomiasis, which is one of the world’s most neglected diseases lacking satisfactory drugs [22]. Furthermore, leucinostatins have displayed broad bioactivity against bacteria and fungi. These antibiotics’ functions are based on their ability to inhibit ATP synthesis in the mitochondria as well as different phosphorylation pathways [23]. These findings drew our attention to the relationships between the bio-control function of P. lilacinum and leucinostatins. Furthermore, genetic and molecular information regarding the biosynthesis of this family of lipopeptide antibiotics, of which little was known to date, could contribute to increasing its production and screening for more efficient derivative compounds.
Genome sequences have shed light on the mechanism of the endoparasitic lifestyle or nematode control beyond biological research. During the preparation of our manuscript, the genome sequence of P. lilacinum was published [24]. Two other plant nematode endoparasitic fungi, Pochonia chlamydosporia [25] and Hirsutella minnesotensis [26], were recently sequenced. Genome sequencing revealed that P. chlamydosporia encoded a wide array of hydrolytic enzymes and transporters expressed at the mRNA level, which supported its multitrophic lifestyle, and H. minnesotensis, which mainly invades juvenile stage cyst nematodes, putatively conducted its parasitic process through lectins, secreted proteases and SMs. Thus, the genome sequence of P. lilacinum provides an opportunity to better understand its mechanism in controlling plant nematodes, and it would be useful to enhance its capabilities as a bio-control agent. At the same time, the genome sequence has the potential to solve the biosynthetic puzzle of leucinostatins as well as to detect novel genes and metabolites that might be of value in agriculture and medicine.
Here, we present the results of genome sequencing of the PLBJ-1 and PLFJ-1 strains of the bio-control agent P. lilacinum, and we increased our knowledge of its bio-control capabilities by comparing the sequences of P. lilacinum with those of other fungi. The genome revealed a repertoire of SM-encoding genes that illustrated the potential for using this fungus to discover natural products. Furthermore, we identified the leucinostatin gene cluster (lcs cluster) and proposed a hypothetical pathway for biosynthesis through genetic manipulation. In the course of screening for new activities of leucinostatins, we found that they inhibited the most notorious oomycetes P. infestans, which causes potato late blight and results in global yield losses of 16% [27].
Two P. lilacinum isolates, PLBJ-1 and PLFJ-1, were sequenced to ensure the accuracy of the genome information and the subsequent analysis. PLBJ-1 and PLFJ-1 were assembled into 144 and 163 scaffolds, respectively, with total sizes of 38.14 and 38.53 Mb, while the published TERIBC I was assembled into 301 scaffolds with a total size of 38.82 Mb (Table 1). The comparative genome sizes of related fungi species are listed in S1 Table. A total of 11,773 and 11,763 gene models were predicted in both genomes, respectively, parallel to other ascomycetes fungi (S1 Table). BLASTN analysis was performed between the two genomes and demonstrated that 83.56% of the PLBJ-1 genome and 82.79% of the PLFJ-1 genome shared high synteny (Fig 2A). According to the syntenic relationship of PLBJ-1 and PLFJ-1, we reconstructed 10 super-scaffolds (S2 Table), which illustrated the physical ubieties of the assembled scaffolds; e.g., scaffold 00006, scaffold 00016 and scaffold 00015 in PLFJ-1 were combined into a super-scaffold (Fig 2B). The overall syntenic relationship of PLBJ-1 and TERIBC 1 showed that 76.52% of the PLBJ-1 genome and 75.12% of the TERIBC 1 genome shared high synteny (S1 Fig). Approximately 6.07% of the repeat sequences that included transposon elements (TEs) (~4.37%) and tandem repeats (~1.70%) were identified in PLBJ-1. The Class I (retrotransposons) and Class II (DNA transposons) TEs occupied ~1.80% and ~0.76% of the genome, respectively. The PLFJ-1 isolate harbored a similar number of repeat sequences (6.00%). The distribution of the TE families was similar in the two isolates, with the exception of certain families, e.g., I, Gypsy, Penelope, Tc1-Mariner and hAT (S3 Table). In total, the two isolates of P. lilacinum contained a larger number of retrotransposons than DNA transposons. P. lilacinum exhibited expansion of repeat content comparable to other ascomycetes fungi, with the exception of H. minnesotensis, Ophiocordyceps sinensis and Fusarium oxysporum (fol), in which the repeat sequences accounted for more than one quarter of the genome (S1 Table). In TERIBC 1, approximately 1.68% of the genome sequence was identified as repeat content.
Among the predicted genes of PLBJ-1, 90.4% were supported by RNASeq data from mycelia cultured in PDB. Both strains exhibited a consistent KOG pattern. Except for the category “General function prediction only”, which was ambiguously sorted to a certain group, the most abundant KOG categories were “Signal transduction mechanisms”, “Posttranslational modification, protein turnover, chaperones”, “Lipid transport and metabolism”, and “Intracellular trafficking, secretion, and vesicular transport” (S2 Fig). A signal peptide analysis showed that 1,410 genes of PLBJ-1 and 1,448 genes of PLFJ-1 encoded putatively secreted proteins.
CAZymes that cleave and build polysaccharides could be required when P. lilacinum degraded the structural polysaccharide armor of nematode eggshells, such as chitin, during the course of its parasitism. The protease could stop the development of nematode eggs and drastically alter the eggshell structures when applied individually or in combination with chitinases [28, 29]. A detailed examination of the CAZymes and proteases of P. lilacinum was performed and compared with other fungi, including nematode parasitic fungi (P. chlamydosporia and H. minnesotensis), nematode-trapping fungi (Arthrobotrys oligospora and Monacrosporium haptotylum), entomopathogenic fungi (T. inflatum, Beauveria bassiana, Cordyceps militaris, Metarhizium robertsii, and O. sinensis), a mycoparasitic fungus (T. ophioglossoides), a saprotrophic fungus (T. reesei) and a plant pathogenic fungus (F. oxysporum). We identified 53 families containing 239 genes in PLBJ-1 and 55 families containing 253 genes in PLFJ-1 that encoded glycoside hydrolases (GH), which was more than the other fungi (an average of 213) (S4 Table). The most abundant family in PLBJ-1 and PLFJ-1 was GH18, which was represented by 32 and 41 chitinases, respectively, that degrade the chitin present in the chitin protein complex of the nematode eggshell [30]. Consistent with GHs, PLBJ-1 and PLFJ-1 contained relatively more carbohydrate-binding modules (CBMs) (59 and 64, respectively) (S5 Table), which were frequently appended to the enzymes involved in polysaccharide depolymerization. A series of carbohydrate esterase (CE)-encoding genes were also detected in the P. lilacinum genomes (33 and 32, respectively), including the most abundant sterol esterases (CE10) and cutinases (CE5), which are virulence factors of some plant pathogens [31] (S6 Table). Another major class of CAZymes, the glycosyltransferases (GT), establish natural glycosidic linkages across a broad range of small and macromolecules, and they were represented in the PLBJ-1 genome with 115 members in 32 families and in the PLFJ-1 genome with 124 members in 32 families (S7 Table). These enzymes’ classification demonstrated that they exhibited less variability in ascomycetes than did GHs, a trend that was maintained in a previous analysis [32]. The P. lilacinum genome contained more proteases (430 and 443, respectively) than other fungi (an average of 396). The largest category of proteases encoded in PLBJ-1 and PLFJ-1 were serine proteases (194 and 198, respectively) (S8 Table), 76 and 81 of which were secreted proteins, respectively. Among the serine proteases, we identified 34 subtilisins (S8) and ten serine carboxyproteases (S10) in the PLBJ-1 genome (36 and 11, respectively, in PLFJ-1), which were reported to be involved in infection and the lethal activity of nematodes [28, 33]. The metalloprotease (108 in PLBJ-1 and 109 in PLFJ-1) and cysteine protease (66 in PLBJ-1 and 68 in PLFJ-1) families also accounted for a significant proportion of the proteases.
A whole genome analysis was conducted against the pathogen-host interaction (PHI) gene database to identify potential virulence-associated genes, under the assumption that the homologue of an experimentally validated pathogenic gene suggested that it played a pathogenic role [34]. We demonstrated that 2,844 (24.1%) and 2,892 (24.6%) proteins of PLBJ-1 and PLFJ-1, respectively, showed sequence similarity to those in the PHI database. Among these proteins, 299 and 317 proteins of PLBJ-1 and PLFJ-1, respectively, were classified as putatively secreted proteins. The KOG functional class distribution of genes related to PHI showed a similar pattern to the whole genome KOG analysis (S2 Fig). The PHI database search yielded 195 CAZymes in PLBJ-1 and 217 in PLFJ-1, 28 and 36 of which were chitinases (GH18), respectively. Of the proteases, 125 in PLBJ-1 and 132 in PLFJ-1 were pathogenic genes according to the PHI database, of which 64 and 72 were identified as secreted proteins, respectively, and these proteins were more likely to function during the infection process [35].
A phylogenomic tree was constructed based on 855 single-copy orthologues of P. lilacinum and 34 other filamentous fungi, with Saccharomyces cerevisiae as the outgroup. The results verified that P. lilacinum belongs to Ophiocordycipitaceae, as described by Jennifer Luangsa-ard [8], and it formed a clade with T. inflatum, T. ophioglossoides, O. sinensis [36], O. unilateralis [37] and H. minnesotensis (Fig 3A). The inferred phylogeny illustrated that T. inflatum and T. ophioglossoides were most closely related to P. lilacinum, and they diverged after their split with O. sinensis, H. minnesotensis and O. unilateralis. This phylogeny also reinforced the previous analysis that found that the split between Cordycipitaceae (including B. bassiana and C. militaris) and Clavicipitaceae (including P. chlamydosporium and M. anisopliae) occurred before Ophiocordycipitaceae diverged from Clavicipitaceae (Fig 3A). The three nematode parasitic fungi P. chlamydosporium, H. minnesotensis and P. lilacinum clustered with insect pathogens, indicating that nematode and insect pathogens might share a common ancestor.
A comparative genomic analysis was performed between P. lilacinum and other nematode-related fungi (the nematode parasites P. chlamydosporia and H. minnesotensis and the nematode-trapping fungi A. oligospora and M. haptotylum). A total of 17,995 orthologous clusters consisting of 76,151 proteins were identified, of which 4,652 clusters containing 35,972 proteins were mapped to all four of the fungi types (Fig 3B). On the whole, the nematode-trapping fungi, which capture nematodes through an entirely different mechanism compared to P. lilacinum [26], possessed the largest number of unique gene clusters, although they had a more distant phylogenetic relationship with the other fungi in Hypocreales (Fig 3B). P. lilacinum contained a large number (3651) of species-specific clusters, while P. lilacinum shared 7,700, 6,673 and 5,253 clusters with P. chlamydosporia, H. minnesotensis and the nematode-trapping fungi, respectively.
Lineage-specific expansions could provide material for the evolution of a specific functional system or adaptation in eukaryotes [38]. To study gene family expansions in P. lilacinum, a comparative genomic analysis of 15 fungal species (PLBJ-1, PLFJ-1, P. chlamydosporia strain 123, P. chlamydosporia strain 170, H. minnesotensis, A. oligospora, M. haptotylum, T. inflatum, B. bassiana, C. militaris, M. robertsii, O. sinensis, T. ophioglossoides, T. reesei, and F. oxysporum) was performed. In total, 1,963 gene families with more than one gene expansion were identified in both PLBJ-1 and PLFJ-1, of which 1,761 gene families were only present in P. lilacinum, and some gene families with significant expansion are listed in S9 Table. However, most families were annotated as reverse transcriptases and transposases, and the others were related to transporters or lyases. When the nematode parasitic fungi P. chlamydosporium and H. minnesotensis were considered, 2,936 orthologous clusters showed expansion in the five isolates. The largest paralogous expansion contained protein families associated with SMs, such as cytochrome P450s, oxidoreductases, and transporters. In addition, these families also contained transcription factors, glycosyl hydrolases, the hAT family, the majority of which are listed in S10 Table.
To evaluate the capability of P. lilacinum to produce SMs, we searched the genome of PLBJ-1 and PLFJ-1 for biosynthetic genes encoding the four classes of the main SM-associated synthetases, including polyketide synthase (PKS), non-ribosomal peptide synthetase (NRPS), terpene synthase (TS) and dimethylallyl tryptophan synthase (DMATS) [26]. A uniform SM profile with parallel categories and numbers was presented in the two genomes (S11 Table). In total, 13 PKSs, 10 NRPSs, two PKS-like enzymes, 10 NRPS-like enzymes, one DMATS, 4 TSs and one PKS-NRPS hybrid were identified in the PLBJ-1 genome, as described in S11 Table. Compared to sequenced species in Ophiocordycipitaceae, the number of SMs in P. lilacinum (41) was similar to the 45 SMs in T. ophioglossoides, 39 SMs in O. unilateralis, more than 30 SMs in Ophiocordyceps sinensis, fewer than 55 SMs in T. inflatum [39], and 101 SMs in the nematode endoparasitic fungus H. minnesotensis [26]. These core backbone genes were dispersed among 39 clusters with other enzymes, such as transcriptional regulators, P450s and transporters, as predicted by antiSMASH (antibiotics and Secondary Metabolite Analysis SHell) [40] (S11 Table). According to the BLAST results from the NCBI NR database, no homologues of functionally characterized SMs were detected. Among them, we detected the expression of 29 core genes with FPKM (fragments per kilobase of transcript per million mapped fragments) values > 0.5, using an RNA-seq analysis of PLBJ-1 cultured in PDB medium for 8 days.
A phylogenetic tree was constructed based on the KS domain amino acid sequence of the PKSs in P. lilacinum and the products of known PKSs, which were divided into three main clades: non-reducing (NR) PKSs, partially reducing (PR) PKSs and highly reducing (HR) PKSs (S3 Fig). VFPBJ_05021, VFPBJ_09342, VFPBJ_09755, and VFPBJ_10843 were predicted as NR PKS-encoding genes, and they shared the highest homology with the non-reducing biosynthetic genes, such as citrinin [41] and griseofulvin [42]. VFPBJ_00212, VFPBJ_02527, VFPBJ_02532, VFPBJ_03442, VFPBJ_05962, VFPBJ _06473, VFPBJ _07567 and VFPBJ_09314 were distributed in the HR PKS clade in close relationship with HR polyketides, such as fumonisin synthase Fum1p [43]. The phylogenetic analysis was consistent with the domain structure analysis of degree of reduction, in which the HR PKS contained the reductive domains KR (keto-reductase), ER (enoyl reductase) and DH (dehydratase), while the NR PKS did not contain these domains (S3 Fig, S11 Table). VFPBJ_05021 and VFPBJ_09342 were grouped with the antibiotics griseofulvin and citrinin with a bootstrap value of 100%, and they shared a common domain structure. This finding suggested that griseofulvin/citrinin or structurally related compounds could be produced by P. lilacinum. However, we did not detect these compounds when P. lilacinum was cultured in PDB for 8 days.
Among the 10 NRPSs, six contained one module or an incomplete module, which could encode products with one amino acid. Four NRPSs were multi-module enzymes, which could encode products composed of more than one amino acid. To examine the potential NRPS orthologues of P. lilacinum and to detect the feasible NRPS evolutionary mechanism in the family Ophiocordycipitaceae, a genealogy was created based on the A-domains from the NRPS of fungi in Ophiocordycipitaceae and several functionally characterized products (S4 Fig). The tree depicted an intricate evolutionary relationship for the NRPS genes. A general trend throughout the tree was that, in Ophiocordycipitaceae, many A-domains clustered with orthologues in other species than with in the same protein. Notably, the 11 A-domains of the cyclosporine synthetases from T. inflatum clustered separately (S4 Fig, node 3), indicating that other species were incapable of encoding cyclosporine and that its evolution occurred after T. inflatum diverged from these fungi in Ophiocordycipitaceae.
This phylogenetic analysis of the A-domains for P. lilacinum detected a series of homologous A-domains: four of the mono-module NRPSs had functionally uncharacterized homologues. VFPBJ_05068 was identified as siderophore synthetase, of which three of the A-domains were grouped with homologues to form a sub-clade (S4 Fig, node 2). The three A-domains of VFPBJ_06596 were grouped with TINF2556, annotated as an ergot alkaloid in T. inflatum, while TINF2556 contained four modules.
The peptaibiotics, a class of linear NRPSs that are abundant of AIB [44], were clustered into one sub-clade (S4 Fig, node 1), mainly including the peptaibiotics from T. ophioglossoides [45], T. inflatum and P. lilacinum. The ten A-domains from VFPBJ_02539 (identified as the leucinostatin biosynthetic gene lcsA in this study), clustered with the ten A-domains from the peptaibiotic TOPH_08469 in T. ophioglossoides, with bootstrap values of 100%, and a global BLAST analysis revealed that the sequence identity of the two homologues was 65%. Neither orthologue was identified in other species of Ophiocordycipitaceae. The single A-domain of lcsA was scattered in the peptaibiotic sub-clade of the tree, while A2, A5 and A6, which activated Leu or related amino acids, were identified in the subsequent study and were grouped together with a bootstrap value of 60%, suggesting that both lineage-specific changes and module duplication contributed to the evolution of the leucinostatin metabolites. In the previous study, A4, A7 and A8 of TOPH_08469 were distributed in a sub-clade enriched in A-domains encoding AIB [45], and our study demonstrated that A4, A7 and A8 of lcsA were encoded for AIB.
In T. ophioglossoides, the TOPH_08469 gene cluster was predicted to contain 28 genes from TOPH_08452 to TOPH_08478 that were located in an ~124 kb region [45]. A comparative analysis of genes surrounding lcsA and TOPH_08469 cluster revealed a high synteny (Fig 4A). VFPBJ_02521 (designed as lcsG) shared 68% sequence identity with TOPH_08452, and lcsA shared 66% sequence identity with TOPH_08469. Interestingly, no homologues of the genes next to the cluster, VFPBJ_02510 to VFPBJ_02520 and VFPBJ_02540 to VFPBJ_02550, were identified in the T. ophioglossoides genome. Within the lcs cluster, two genes, cytochrome P450 lcsI and a protein with unknown function, lcsM, did not possess homologues in the TOPH_08469 cluster, while all of the leucinostatin biosynthetic genes in T. ophioglossoides (TOPH_08452 to TOPH_08469) had homologues within the lcs cluster. These results suggested that this nearly 100 kb region might have been horizontally transferred from other fungal or bacterial species. However, leucinostatins have not been reported to be produced by T. ophioglossoides to date.
The lipopeptide leucinostatin A contains ten amide bonds that divide the molecule into 11 moieties, including 4-methylhex-2-enoic acid, 9 amino acid residues and DPD. The property of the mixture of the polyketide and peptide moieties in the leucinostatins indicated a PKS, NRPS or hybrid PKS-NRPS origin. It is logical to consider that a single reducing PKS encodes the 4-methylhex-2-enoic acid, and a NRPS enzyme encodes the remaining portion, as in the models for emericellamide synthesis in Aspergillus nidulans[46] and pneumocandin in Glarea lozoyensis[47]. Among the multi-module NRPSs in P. lilacinum, VFPBJ_05068 contains 13 domains grouping into 3 modules, VFPBJ_06596 contains seven domains grouping into three modules, and VFPBJ_11400 contains six domains grouping into two modules. These enzymes were insufficient for the assembly of nine amino acids of leucinostatins. Thus, VFPBJ_02539 was left as the only plausible candidate. VFPBJ_02539 (LcsA) consists of 11,872 amino acids and was encoded by a gene with five introns. The domain structure of LcsA was comprised of 10 C-A-PCP modules and carried the correct number of amino acids for the assembly of leucinostatins. The NRPSpredictor2 [48] offered little insight into the substrates except that the substrates of A1 and A3 were proline and leucine, respectively (S12 Table). Two PKSs, lcsB and lcsC, located not far upstream of lcsA, which could encode 4-methylhex-2-enoic acid, indicated that this cluster is responsible for leucinostatin production. Furthermore, lcsD (VFPBJ_02533), located between lcsA and the PKSs, was annotated as an acyl-CoA ligase, offering a conceivable route for connecting the fatty acid and peptide.
To verify the associations between the putative lcsA and leucinostatins, a gene deletion method was developed for P. lilacinum based on the previous method for Fusarium oxysporum[49], with the G418 sulfate-resistance gene neo as the selection marker. A portion of lcsA (2,613 bp, including 236 bp upstream of the ORF) was knocked out by double homologous deletion cassettes with the neo marker via PEG-mediated transformation, and the resulting G418 sulfate-resistant isolates (S5A and S5B Fig) were verified by diagnostic PCR, using the primers in neo and outside the knockout cassette (S5C Fig) (S13 Table). Finally, one mutant (ΔlcsA) of PLBJ-1 was isolated with correct PCR amplification products from 320 G418 sulfate-resistance mutants (S5C Fig), and the remaining isolates resulted from ectopic integration of the neo gene cassette into the genome. The wild type of P. lilacinum and the ΔlcsA mutant of PLBJ-1 were cultured in PDB medium for 8 days, and the ethyl acetate extracts were analyzed by HPLC-MS. The MS spectrum of the wild type displayed two overlapping peaks at 15.6 and 16.0 min, with m/z [M+H]+ of 1218.9 and 1204.9, respectively, which were assigned to leucinostatins A and B and were absent in the ΔlcsA mutant (Fig 5, S6 Fig). A comparison with the authentic standard confirmed that the missing compounds of the ΔlcsA mutant were indeed leucinostatins A and B (Fig 5, S6 Fig). As expected, these results demonstrated the essential roles of lcsA in the biosynthesis of the leucinostatins.
Different boundaries of the lcs cluster were defined by the SMURF and antiSMASH programs (Fig 4A). Nine genes flanking lcsA from VFPPL_02532 to VFPPL_02540 spanning 62 Kb were predicted to be in the cluster by SMURF (Secondary Metabolite Unique Regions Finder) [50], while a larger cluster comprising 26 genes from VFPBJ_02521 to VFPBJ_02546, spanning 120 Kb, was predicted by antiSMASH. Therefore, it was necessary to explore the genes that were involved in the pathway using a biological approach.
Changes in the culture medium could impact the general metabolic profile of an organism, based on the “OSMAC” (one strain-many compounds) hypothesis[51]. Indeed, we found that P. lilacinum produced leucinostatins A and B when cultured with our lab recipe of PDB but did not produce leucinostatins when cultured in PDB-BD (see the Materials and Methods section). This result provided clues to identify the boundary of the lcs cluster using producing versus non-producing media. qRT-PCR analysis was conducted to compare the expression patterns of genes flanking lcsA when PLBJ-1 was grown in the two types of media for 8 days. Furthermore, RNA-Seq of PLBJ-1 under leucinostatin-inducing conditions (PDB medium) was performed.
As expected, the expression level of NRPS lcsA when P. lilacinum was grown in leucinostatin-inducing medium was upregulated 95-fold, compared to those grown in non-inducing medium (Fig 4B). The genes downstream of lcsA, including the putative transporter ABC gene VFPBJ_02540, did not display a higher expression level in the leucinostatin-inducing medium, indicating that they were not involved in the leucinostatin biosynthesis pathway. Correspondingly, the RNA-Seq expression profile during leucinostatin production showed a low FPKM value of VFPBJ_02540 (2.01) (S7A Fig), while the FPKM value of lcsA was 65.4. These results indicated that the 3’ edge of the cluster was lcsA. The genes upstream of lcsA from VFPBJ_02520 to VFPBJ_02538 (lcsT) were upregulated at different levels in the leucinostatin-inducing medium. A 16- to 2692-fold increase in expression was observed (Fig 4B), except for three genes, VFPBJ_02520, LcsM, and lcsQ, which showed less than 10-fold increase and low FPKM values in the transcriptional data (S7A Fig). VFPBJ_02520 was annotated as a phosphohydrolase that appeared to be involved in nucleic acid metabolism and signal transduction, instead of secondary metabolism[52]. Thus, we speculated that the 5’ boundary of the cluster was VFPBJ_02521 (lcsG). To support this hypothesis, the expression patterns of the genes flanking the cluster were analyzed using qRT-PCR analysis in wild type PLBJ-1 and ΔlcsA grown in leucinostatin-inducing medium. We observed an increase in the expression of wild type P. lilacinum ranging from four- to 79-fold (S7B Fig). Thus, a series of genes from VFPBJ_02521 to VFPBJ_02539, designated as lcsA to lcsT, included the core enzymes, modifying enzymes and transporter enzymes coding for the biosynthesis of leucinostatins (Fig 4A, Table 2).
Considering the structural similarities of leucinostatin A with emericellamide A [53] and pneumocandin [47], we reasoned that a similar biosynthetic mechanism might be required to form the skeletons of lipopeptides and peptides. As reported, a single module polyketide synthase iteratively catalyzes the formation of the linear polyketide chain; in daptomycin [54] and echinocandin B [55], acyl-CoA ligase converts the fatty acid to fatty acyl CoASH; in compound W493 B [56], a thioesterase was proposed to hydrolyze the thiol bond and shuttle the product to the first module of NRPS. To determine whether the same enzymes play critical roles in the leucinostatin biosynthesis pathway, we disrupted the PKS (lcsC), ligase (lcsD) and thioesterase (lcsE)-encoding genes in the cluster by homologous recombination (S5A Fig) and verified the mutants by PCR amplification (S8A Fig). After culturing the fungi in PDB medium and comparing the extracts with the PLBJ-1 wild type and ΔlcsA by HPLC-MS, we showed that leucinostatins A and B disappeared in ΔlcsC, ΔlcsD and ΔlcsE, similar to ΔlcsA (Fig 5, S6 Fig).
A powerful approach to enhancing the production of leucinostatins was to express transcription factors constitutively that were used for other SMs [57]. lcsF encodes a putative transcription factor with a bZIP domain structure, and it is associated with secondary metabolism [58]. To assess the function of lcsF, we cloned it into the KSTNP vector under the control of the TrpC promoter. The resulting plasmid, KSTNP-OElcsF, was randomly integrated into the genome of wild type P. lilacinum (S8B Fig). The positive transformants were screened by G418 sulfate and were diagnosed by PCR amplification of the expression cassette (S8C Fig). Transformants with an intact overexpression cassette were cultured in leucinostatin-inducing PDB medium for 8 days. The expression level of lcsF in the mycelia was analyzed by qRT-PCR, and six of ten transformants demonstrated more than 20-fold upregulation. Finally, three transformants without changes in their physiological indices were selected for the downstream test. As expected, all 20 genes in the cluster were upregulated to some extent by lcsF, with the exception of lcsL and lcsP, which were downregulated three- and five-fold (S9A Fig). In addition to the 30-fold increase in lcsF expression, the expression of the three PKS/NRPS synthase encoding-genes (lcsB, lcsC and lcsA) were increased by ~3- to 4-fold. For O-methyltransferase (lcsG), ABC transporter (lcsH and lcsO), thioesterase (lcsE), epimerase (lcsT) and the unknown function genes lcsM and lcsS, we observed ten-fold or higher upregulation. The other genes in the cluster displayed a two- to ten-fold increase in expression. Genes adjacent to the lcs cluster, VFPBJ_02520 and VFPPL_02540, were downregulated three-fold. After the wild type and OE::lcsF P. lilacinum were grown in PDB medium with shaking for 8 days, the resulting HPLC profile showed that the titers of leucinostatins A and B were elevated by at least 50% (S9B Fig). These results provided evidence that the pathway-specific transcription factor lcsF was capable of regulating the entire gene cluster and leucinostatin biosynthesis, further verifying the boundary of this cluster.
The deletion and overexpression of the genes in the lcs cluster had no apparent effects on the fungal hyphae or spore phenotypes of P. lilacinum and did not cause any growth defects. It is well known that leucinostatins are antibiotics used to combat fungi and bacteria. Here, we found that leucinostatins contributed to the inhibition of oomycetes, which had not previously been reported. The growth of P. infestans and P. capsici was inhibited in a confronting incubation with wild type P. lilacinum and OE::lcsF, while the inhibition disappeared when they were grown in a confronting incubation with ΔlcsA (Fig 6). Similar to ΔlcsA, P. infestans could grow normally in a confronting incubation with ΔlcsC, ΔlcsD and ΔlcsE (S10 Fig). The results indicated that leucinostatins A and B inhibited the growth of some oomycetes. A gradient inhibitory zone was explored with leucinostatins A and B in different concentrations to find quantitative evidence of inhibition against P. infestans (S11 and S12 Figs).
P. lilacinum is one of the most important endo-parasites of plant nematodes. We obtained the genome sequence of two P. lilacinum strains and compared them with other nematode parasites, nematode-trapping fungi, insect parasites, a mycoparasitic fungus, a saprotrophic fungus and a plant pathogen. This method provided insights into the life strategy and evolution of nematode endoparasites.
Major gene families (GH, protease, SMs) could corroborate each other for the three P. lilacinum strains (PLBJ-1, PLFJ-1 and the published TERIBC 1). However, TERIBC 1 was predicted to encode more CEs and fewer PHI genes in contrast with PLBJ-1 and PLFJ-1 (Table 1). The lcs cluster was also detected in the TERIBC 1 genome. The genomic sequence identity of the lcs cluster (lcsG to lcsA) between PLBJ-1 and TERIBC 1 was 98.0%, and the sequence identity was 99.0% between PLBJ-1 and PLFJ-1, with the syntenic relationships shown in S13 Fig.
Although fungi have been screened for activity as bio-control agents against P. infestans, the biological control of late blight is dominated by bacterial antagonists. Microbial compounds known as biosurfactants [59] are believed to participate in the process. For example, the cyclic lipopeptide massetolide A produced by Pseudomonas fluorescens exhibited destructive effects on the zoospores of oomycetes [60]. The inhibition of P. infestans by leucinostatins provided the basis for their chemical application in agriculture and for further biological studies of the antagonist P. lilacinum on oomycetes, which had not been researched previously. Leucinostatins also demonstrated inhibition against P. capsici [61], another oomyceteous plant pathogen, while the antagonism of P. lilacinum against P. capsici seemed inferior to that against P. infestans, as shown in Fig 6.
The phylogenomic analysis revealed that P. lilacinum was a member of Ophiocordycipitaceae, Hypocreales, which includes fungi engaged in various lifestyles, and it was not related to the previously considered Paecilomyces in Sordariales. This species’ closest relatives, T. inflatum and T. ophioglossoides, are insect and fungal parasites (Fig 3A), supporting the viewpoint that parasitism might occur due to the formation of novel genes that could be acquired through horizontal transfer or gene duplication and could play specific roles during host infection [62]. Moreover, these results indicated that the nematode pathogens had a strong link with insect pathogens and were distantly related to nematode-trapping fungi, as previously described in [25] and [26]. The large number of hydrolytic enzymes, particularly GHs and proteases, putatively secreted proteins and pathogenesis-related proteins in P. lilacinum support its various lifestyles as it encounters diverse nutrient resources [63]. Chitin and proteins comprise a significant proportion of the nematode and insect surface, the degradation of which requires serine proteases and chitinases.
The development of natural compounds from bio-control fungi have recently attracted considerable interest because the production of nematode-toxic SMs could also be a strategy for fungi to infect nematodes [64]. Next-generation sequencing technologies are becoming an essential tool for identifying novel genes for metabolite biosynthesis in fungi. The genome sequence of P. lilacinum revealed the potential to produce a rich repertoire of SMs, including 41 core enzymes. Most of the PKS enzymes could be clustered with the PKS enzymes encoding bioactive polyketides when the PKS tree was constructed based on the KS domains (S3 Fig). The NRPS enzyme tree based on the A-domains demonstrated that some of NRPSs of P. lilacinum possessed homologues in closely related species (S4 Fig). However, none of the enzymes’ biosynthetic function have been validated previously. The polyketide compounds acremoxanthone C and acremonidin A belong to the xanthone−anthraquinone heterodimer that was recently isolated from P. lilacinum. Structurally related heterodimeric compounds, such as acremoxanthones A and B [65] and xanthoquinodin B3 [66], have been isolated from the genera Acremonium and Humicoma. This category of compounds and their derivatives have remarkable biological and medicinal activities, and their total synthesis has garnered attention worldwide. Unfortunately, there still exist limitations in the current synthesis methods [67], and metabolite regulation based on molecular biosynthesis is not available because the biosynthetic genes have not been identified. One or more non-reducing PKSs (S3 Fig) might be involved in the biosynthesis of acremoxanthone C and acremonidin A, according to their structures.
By analyzing the genes located in the leucinostatin biosynthetic cluster, combined with the HPLC-MS analysis of gene deletion mutants (Fig 5), we were able to propose a putative biosynthetic pathway for leucinostatins (Fig 7). This hypothetical biosynthesis initiated with the assembly of 4-methylhex-2-enoic acid by a reducing PKS. However, two reducing PKS encoding genes with 38% sequence identity are present in the cluster, and both contained KS, AT, DH, cMT, ER, KR and ACP domains. We excluded the possibility of a partnership between the two PKSs in a sequential manner or a convergent manner, as has been reported for asperfuranone [68] and azaphilone A [69], based on their structures. Moreover, in the biosynthesis of chaetoviridins and chaetomugilins from Chaetomium globosum, a PKS cazF (KS-AT-DH-cMT-ER-KR-ACP) encoded an intermediate 4-methylhex-2-enoic acid [70]. The protein sequence identities between cazF and lcsB/lcsC were both 29%; thus, we could not estimate which PKS was responsible for 4-methylhex-2-enoic acid. The results of RNA-seq and qRT-PCR indicated that both genes contributed to leucinostatin synthesis. The deletion of lcsC interrupted leucinostatin biosynthesis, which confirmed that lcsC is essential for the synthesis of leucinostatins. Due to the difficult genetic manipulation, we failed to obtain lcsB deletion mutant.
The lipopeptide pathways and organizations of their clusters have some striking commonalities. We got clues from the lipopeptides echinocandin B [55], pneumocandin [47] and emericellamide [53]. The polyketide residue might be transferred to the NRPS LcsA, mediated by two additional putative enzymes, acyl-CoA ligase (LcsD) and thioesterase (LcsE). The linear polyketide carboxylic acid, which was released from PKS, was converted to a CoA thioester by LcsD, and then LcsE hydrolyzed the thiol bond and shuttled the polyketide intermediate to LcsA. 4-Methylhex-2-enoic acid was not detected in the culture of the ΔlcsD isolate, indicating that the triketide might be sticked in the PKS enzyme to prevent its release until the ligase is added for the reaction.
The phylogenetic analysis of the ten A-domains of LcsA revealed that LcsA_A2, LcsA_A5, LcsA_A6 and LcsA_A3 were grouped into one clade, and LcsA_A4, LcsA_A7 and LcsA_A8 were grouped into another clade (S14 Fig). The conserved domains are believed to have evolved through module duplication, and they activate similar amino acid structures [71]. In the plausible model for leucinostatin synthesis, A5 and A6 incorporated leucine, A2 incorporated AHyMeOA, the structure of which is equal to a hydroxyl-3-pentone extending at the leucine, and A3 incorporated 3-hydroxyl leucine. A4, A7 and A8 incorporated AIB. Thus, the structural similarity of amino acids activated by conserved A-domains verified that the 4-methylhex-2-enoic acid moiety in the leucinostatins was assembled by a discrete enzyme, instead of LcsA. The C domain of the first module catalyzed the condensation of 4-methylhex-2-enoic acid and MePro carried by domain A1, followed by successive condensations of nine amino acids to trigger the elongation of the linear peptide. Next, the peptide scaffold would be released by the NAD(P)H-dependent R domain (thioester reductase) at the C-terminal region of LcsA.
In the leucinostatin biosynthetic pathway, it is intriguing that the DPD residue at the C-terminus of leucinostatin A was neither an amino acid nor a carboxy acid, which are incapable of being activated by the A-domain and converting to amino acyl adenylate [72]. The DPD seems to be a modified form of amino acids, whereas the primary form of this moiety cannot be determined based on the domain sequence. However, we could deduce a possible pathway for the modification of the last amino acid according to the structure and the function of the genes in the cluster (Fig 7). Originally, an Ala was likely incorporated into the decapeptide skeleton by the A10 domain, which was attached to LcsA via a thioester bond; subsequently, the R domain released this intermediate product. The NAD(P)H-dependent R domains reductively catalyzed to produce linear aldehyde 1 by off-loading peptide thioesters, following completion of the peptide skeleton, as presented in previous studies [73, 74]. Linear aldehydes frequently occurred as intermediates and underwent subsequent reactions, such as macrocyclization, to yield the imine product koranimine [75]; further reduction yielded myxochelin A or transamination to form an amine myxochelin B by aminotransferase [76]. Regarding the leucinostatins, we speculated that aldehyde 1 would go through a transamination reaction to form compound 2, which was accomplished by the putative aminotransferase lcsP.
In this pathway, the unhydroxylated leucine of intermediate 2 undergoes hydroxylation to form compound 3. Three putative cytochrome P450-encoding genes (lcsI, lcsK and lcsN) within the cluster alternatively might catalyze this modification. Another scenario equivalent to this pathway for leucinostatin A synthesis was that a leucine was hydroxylated prior to its incorporation into the peptide. In all likelihood, the varying extents of methylation of compound 3 catalyzed to form leucinostatins A and B. It is worth mentioning that, had the methylation reaction not occurred, compound 3 might be the ultimate precursor of leucinostatin C, which is compound in leucinostatin family isolated from P. marquandii.
The AHyMeOA in leucinostatin A activated by the A2 of lcsA was regarded as a ramification of leucine because leucine is located at this position in leucinostatins C, T, F, D and H, although these compounds were not detected in the PLBJ-1 culture. Based on its structure and the presence of redundant PKSs within the cluster, alternative PKS could be involved in synthesizing the carbon chain. In addition, the leucinostatins contained the nonproteinogenic MePro, incorporated in the synthesis of nostopeptolides in Cyanobacteria [77]. A zinc-dependent dehydrogenase, nosE, and a P5C reductase, nosF, were involved in the oxidation and subsequent cyclization of leucine to form MePro, and the presence of nosE and nosF recently led to screening for novel MePro-containing peptides [78]. In the pneumocandins from Zalerion arboricola, feeding experiments established that leucine was cyclized to produce 3-hydroxy-4-methylproline, whereas MePro might be an intermediate [79]. It was reasonable to assume that the MePro in the leucinostatins originated from leucine cyclization. Although homologues of nosE nor nosF were not present in the lcs cluster, it was plausible that MePro biosynthesis, engaged in a separate pathway, was independent of leucinostatin synthesis. Another nonproteinogenic amino acid, β-Ala, was present in leucinostatins and activated by the A9 of lcsA. A previous study of the destruxins in Metarhizium proposed that the aspartic acid decarboxylase dtxS4 triggered the decarboxylation of aspartic acid into β-Ala, as a substrate for the assembly line [80]. A genome-wide blast search for genes encoding aspartic acid decarboxylases in PLBJ-1 revealed the presence of two candidate genes, VFPBJ_01400 and VFPBJ_10476, with 68% and 61% sequence identity to dtxS4, respectively, which could have catalyzed the biosynthesis of β-Ala in leucinostatins.
The genomes of P. lilacinum strains PLBJ-1 and PLFJ-1 were sequenced, completely assembled, annotated, and comparatively analyzed with related fungi. Phylogenomic analysis showed that P. lilacinum was most closely related to T. inflatum and T. ophioglossoides, and the cluster of nematode parasitic fungi and insect pathogens indicated their common origin. PKS and NRPS-encoding genes were thoroughly characterized and analyzed by phylogenetic analysis, from which we found that lcsA was specific to P. lilacinum and T. ophioglossoides. Furthermore, lcsA was proved to be responsible for leucinostatin biosynthesis by homologous deletion. The boundary of the lcs cluster was identified by comparison of gene expression levels when P. lilacinum was cultured in leucinostatin-inducing and non-inducing medium as well as RNA-Seq analysis. Disruption of lcsC, lcsD and lcsE demonstrated the critical roles of PKS, acyl-AMP ligase and thioesterase in the biosynthetic pathway of leucinostatins. Overexpression of the transcription factor lcsF increased the production of leucinostatins A and B through regulated expression levels of genes in the lcs cluster. We also demonstrated that leucinostatins could enable the fungus with antagonistic activity against the oomycetes.
The leucinostatin-producing P. lilacinum strain PLBJ-1 (CGMCC3.17492) was isolated from tomato roots in Beijing, China, and PLFJ-1 (CGMCC3.17493) was isolated from tomato roots in Fujian, China. Both strains were sequenced to obtain the common features of P. lilacinum and to ensure the information accuracy of the lcs cluster. PLBJ-1 was used as the wild type recipient for the subsequently genetic manipulations because PLFJ-1 was insensitive to the antibiotics used as selection markers. P. infestans and P. capsici were maintained at the Chinese Academy of Agricultural Sciences. The pKOV21 vector used for homologous deletion and the KSTNP vector used for overexpression came from Prof. Youliang Peng, China Agricultural University. The leucinostatin A standard came from Bioaustralis, Inc. (NSW, AUS). G418 sulfate was purchased from Amresco, Inc. (OH, USA).
The non-inducing PDB-BD medium (Potato Dextrose Broth) came from Becton, Dickinson and Company (NJ, USA). Leucinostatin-inducing PDB medium was prepared in the lab. Briefly, 200 g of potatoes were boiled for 30 min, and then 20 g of glucose were dissolved into the filtrate and diluted to 1 L. The rye agar medium contained 50 g of crushed rye, 20 g of sucrose and 15 g of agar per liter. PDB cultures with 1×105 conidia per mL of PLBJ-1 were grown at 28°C on a shaker at 150 rpm for 8 days before DNA/RNA isolation.
The mycelium tissues of the PLBJ-1 and PLFJ-1 isolates were harvested via filtration. Genomic DNA was isolated using a Qiagen DNeasy kit, according to the manufacturer’s protocol. The PLBJ-1 tissue for RNA isolation was grown in the leucinostatin-inducing medium. RNA was extracted using TRIZOL reagent (Invitrogen, USA) following the manufacturer’s protocol.
The raw sequencing data (Illumina HiSeq 2000) from the PLBJ-1 and PLFJ-1 strains were generated by BGI-Shenzhen (China) and Berry Genomics Co., Ltd. (China), respectively. A total of 13.27 Gb bases for the PLBJ-1 strain from three libraries, with average insert sizes of 165 bp, 758 bp and 5,490 bp, were obtained, and 5.88 Gb bases for the PLFJ-1 strain from two libraries with the average insert sizes of 175 bp and 4,760 bp were obtained. Both of the genomes were assembled using ALLPATHS-LG revision 42305 [81]. The repeat sequences were identified as previously described [82], based on de novo and homology methods. For the de novo method, Piler [83] and RepeatScout, version 1.0.5 [84], were used to construct the repeat sequence families; then, RepeatMasker, version 4.0.5, was used for repeat analysis. For the homology method, the sequence families from Repbase, version 19.06 [85], were used for annotation by performing RepeatMasker analysis.
For gene prediction, the Augustus algorithm, version 2.7 [86], identified 11,404 and 11,554 complete genes for the PLBJ-1 and PLFJ-1 strains, respectively, and the GeneMark-ES algorithm, version 2.3f [87], discovered 11,001 and 11,070 complete genes for the PLBJ-1 and PLFJ-1 strains, respectively. The comparison showed that 9,509 and 9,562 genes of the PLBJ-1 and PLFJ-1 strains were predicted by both Augustus and GeneMark-ES. These consensus genes were considered to be high quality predicted genes and were used in this study. The additional 2,264 and 2,201 genes of the PLBJ-1 and PLFJ-1 strains were obtained according to the method in [82]. EuGene, version 4.1 [88], was used to integrate multiple sources, including transcription start sites identified by Netstart [89], homologous proteins identified from the Swiss-Prot database, version 2015-07-22, by BLAST, version 2.2.26, the assembled transcripts generated by IDBA-tran, version 1.1.1 [90], and the exon junctions identified from RNA-seq by Tophat, version 2.0.13 [91]. The gene expression values were presented by the expected FPKMs using Cufflinks, version 2.2.1 [92], based on the Tophat [91] analysis.
Proteins were annotated by aligning their sequences to the NCBI fungi refseq, version 2015-07-10, and SwissProt, version 2015-07-22, with an E-value cutoff of 1e-5 using BLASTP. In addition, the Pfam database, version 27.0, was used for domain annotation by HMMER, version 3.1b1 (http://hmmer.janelia.org/). The putative proteins were further classified by Gene Ontology (GO) [93], using Blast2Go [94], and the euKaryotic Clusters of Orthologous Groups (KOG) [95], using BLAST (E-value of 1e-5). The Web server of the CAZymes Analysis Toolkit (CAT) [96] was used to identify CAZymes in P. lilacinum with E-values ≤ 1e-50.
The proteases were discovered by the MEROPS batch BLAST online server [97]. Proteins with sequences that matched the cytochrome P450 genes [98] with E-values ≤ 1e-50 were annotated as P450 enzymes. Candidate pathogenic factors were predicted by sequence alignment against the Pathogen Host Interactions (PHI) database, version 3.5 [99], with E-values ≤ 1e-50. In addition, the secretomes were identified based on recognizing the signal peptide and transmembrane sequences. Proteins were considered to be secreted proteins if the signal peptides were identified by at least two methods among SignalP, version 4.0 [100], TargetP, version 1.1 [101], Phobious, version 101 [102], and Predisi [103], and transmembrane sequences were not identified by at least one of the methods among SignalP, Phobious and TMHMM, version 2.0c [104].
Orthologous groups of genes from P. lilacinum and the other fungi listed in S14 Table were detected by OrthoMCL, version 2.0.9 [105], and then were filtered to identify the single copy orthologues. The single copy orthologues were aligned with MUSCLE [106]. The poor alignment regions of the concatenated sequences were removed using Gblock, version 0.91b [107], and then the high quality sequences were used for the maximum likelihood phylogeny analysis with the Dayhoff model implemented in the TREE-PUZZLE program [108]. Bootstrap support value was calculated by analyzing 1,000 replicates.
The secondary metabolite genes were discovered by performing SMURF [50] and antiSMASH [40] analyses. PKS and NRPS domain structures were characterized by antiSMASH and Pfam, or were visually identified by multiple alignments. The KS domains extracted from PKS and the A-domain from NRPS were aligned by MUSCLE [106], and then a maximum likelihood phylogeny was constructed by treeBeST (http://treesoft.sourceforge.net/treebest.shtml) using 1,000 bootstrap replicates.
Three biological replicates were performed for each analysis of the relative expression levels. The cDNAs were synthesized with a TIANScript Ⅱ RT Kit (TIANGEN, China). The cDNA was analyzed by qRT-PCR using SYBR Premix Ex Taq (TAKARA, Japan) on a BIO-RAD CFX96 (BIO-RAD). The housekeeping actin gene designed from VFPBJ_07912, which was similar to the reported GU299860.1, was used for normalization. The relative expression values were calculated using the 2-∆∆Ct method. The primers are listed in S13 Table.
Polyethylene glycol-mediated protoplast transformation of PLBJ-1 was performed as previously reported [49, 109], with the following modifications: the protoplast was produced by 20 gL−l Driselase (Sigma) digestion for 4 h at 31°C. The regeneration medium was PDA medium containing G418 sulfate (400 μg/L), supplemented with molasses (10 g/L), saccharose (0.6 M), yeast extract (0.3 g/L), tryptone (0.3 g/L), and casein peptone (0.3 g/L) [10]. The construction of knockout and overexpression plasmids originated from pKOV21 and KSTNP, and the primers are listed in S13 Table. A quick method for isolating the fungal genomic DNA was developed to screen for a large number of transformants. Briefly, a nip of mycelia was transferred to 50 μL of NaOH (50 mM) and was incubated at 95°C for 20 min. The solution was directly used for PCR amplification after 5 μL of Tris-HCl (1 M) were added to neutralize the base.
Cultures of 1×105 conidia per mL of P. lilacinum and its mutants were grown in leucinostatin-inducing PDB medium at 28°C on a shaker at 150 rpm for 8 days. Culture medium (7.5 L) was extracted with the same volume of EtOAc three times (each 1 h) ultrasonically. The combined EtOAc extracts were concentrated to afford a crude extract (0.4 g), which was subjected to reversed-phase ODS column chromatography eluting with MeOH-H2O (from 40% to 100%) to afford 6 fractions (Fr.A–Fr.F). Fr.E (40 mg) was passed through a Sephadex LH-20 column (MeOH) and yielded mixtures of 5.0 mg of leucinostatins A and B. The structure of the mixtures was further identified by standard substance using LC-MS analysis. Approximately 200 mL of culture medium were used for comparative LC-MS analysis between PLBJ-1 and its mutants. LC-MS was performed on an Agilent Accurate-Mass-QTOF LC/MS 6520 instrument. HPLC analysis was performed on a Waters HPLC system (Waters e2695, Waters 2998, Photodiode Array Detector) using an ODS column (C18, 250 × 4.6 mm, YMC Pak, 5 μm). The ODS (50 μm) column was produced by YMC Co. Ltd. (Kyoto, Japan). The Sephadex LH-20 was purchased from GE Healthcare. Analytical HPLC was conducted with a Waters HPLC system (Waters e2695, Waters 2998, Photodiode Array Detector) using an ODS column (C18, 250 × 4.6 mm, YMC Pak, 5 μm) with a flow rate of 1 mL/min. The fresh extracts were dissolved in methanol before being separated on a linear gradient of MeOH:H2O (0.1% formic acid) at a flow rate of 1 mL/min. Fresh extracts of mutant strains were detected for 30 min using a linear gradient of 20% to 100% (0–20 min), 100% MeOH (20–25 min), and 20% MeOH (25–30 min). The LC-MS analysis method was consistent with analytical HPLC.
Confronting incubation of P. lilacinum (wild type, ΔlcsA and OE::lcsF) with P. infestans was performed on rye agar medium in 9 cm Petri plates, incubated simultaneously and cultured at 28°C for 24 h and then at 18°C, the optimum temperature for P. infestans, for 9 days, while confrontation with P. capsici was performed on lab-made PDA medium, cultured at 28°C for 7 days. For the inhibitory zone experiment, freshly produced sporangia of P. infestans was suspended in sterile water at a concentration of 2×105 sporangia/mL. One milliliter of the suspension was smeared on 15 cm Petri plates, followed by Oxford cups with a diameter of 1 cm being placed. From 5 to 60 μg (increment 5 μg) of leucinostatins A and B dissolved in 20% methanol were added to the Oxford cup, and 20 μL of 20% methanol were used a control. Then, the Oxford cups were removed after the solution was absorbed by media. Five days later, the area was calculated by drawing circles of the inhibitory zone on metric graph paper and counting the number of square millimeters within the circle [110, 111]. Three biological replicates were performed. At the same time, the effects of 50 μg, 33 μg, 17 μg and 8.5 μg of leucinostatins are demonstrated in S11 Fig.
The genome sequences of PLBJ-1, PLFJ-1, and P. chlamydosporium strain 170 used for comparative analysis have been deposited at GenBank under the accession numbers LSBH00000000, LSBI00000000 and LSBJ00000000, respectively.
|
10.1371/journal.pbio.1001456 | Disease Ecology, Biodiversity, and the Latitudinal Gradient in Income | While most of the world is thought to be on long-term economic growth paths, more than one-sixth of the world is roughly as poor today as their ancestors were hundreds of years ago. The majority of the extremely poor live in the tropics. The latitudinal gradient in income is highly suggestive of underlying biophysical drivers, of which disease conditions are an especially salient example. However, conclusions have been confounded by the simultaneous causality between income and disease, in addition to potentially spurious relationships. We use a simultaneous equations model to estimate the relative effects of vector-borne and parasitic diseases (VBPDs) and income on each other, controlling for other factors. Our statistical model indicates that VBPDs have systematically affected economic development, evident in contemporary levels of per capita income. The burden of VBDPs is, in turn, determined by underlying ecological conditions. In particular, the model predicts it to rise as biodiversity falls. Through these positive effects on human health, the model thus identifies measurable economic benefits of biodiversity.
| While most of the world is thought to be growing economically, more than one-sixth of the world is roughly as poor today as their ancestors were hundreds of years ago. The extremely poor live largely in the tropics. This latitudinal gradient in income suggests that there are biophysical factors, such as the burden of disease, driving the effect. However, measuring the effects of disease on broad economic indicators is confounded by the fact that economic indicators simultaneously influence health. We get around this by using simultaneous equation modeling to estimate the relative effects of disease and income on each other while controlling for other factors. Our model indicates that vector-borne and parasitic diseases (VBPDs) have systematically affected economic development. Importantly, we show that the burden of VBPDs is, in turn, determined by underlying ecological conditions. In particular, the model predicts that the burden of disease will rise as biodiversity falls. The health benefits of biodiversity, therefore, potentially constitute an ecosystem service that can be quantified in terms of income generated.
| Despite long-term economic growth trajectories for most countries, extreme poverty persists for more than one-sixth of the world. The distribution of wealth and poverty has a clear geographic signature. Along with 93% of the global burden of vector-borne and parasitic diseases (VBPDs), the tropics host 41 of the 48 “least developed countries” and only two of 34 “advanced economies” (Figure 1) [1]–[3].
The latitudinal gradient in income is highly suggestive of underlying biophysical drivers. Latitudinal gradients are found among an extraordinarily wide range of intra- and inter-specific biological processes, from the evolution of animal body size to species diversity, and have served as centerpieces of a number of over-arching paradigms in evolutionary and ecological theory [4]–[11]. These common patterns suggest an opportunity for natural scientists to contribute to a more unified understanding of the role of biological processes in economic development [12]–[15].
Among the many potential biological drivers, the burden of VBPDs stands out as fundamental to explaining geographic distributions of income. VBPDs continue to be among the leading causes of morbidity and mortality of poor populations. Unlike directly transmitted diseases, VBPDs spend much of their life cycle outside of the human host, in other host species or in free-living stages, and are thus especially dependent on external environmental conditions. There is now a consensus among many economists that at least some VBPDS, such as malaria and hookworm, have systematically influenced economic growth [13],[14],[16]–[18].
However, intense debate remains on the relative importance of general disease burden indices on global patterns of wealth and poverty. One side of this debate argues that tropical climates harbor more infectious diseases and offer inferior agricultural conditions, which together influence the overall level of health in the population [13],[14],[16],[19]–[22]. This is thought to directly harm the acquisition of human capital and labor productivity, and increase mortality rates [23]. The corresponding low life expectancies are known to also influence more subtle household allocations of resources, such as reproductive behavior, child-rearing, and long-term private investment.
On the other hand, some have argued that the effect of geography on development has only been through its historical influence on the formation of government and economic institutions [24]–[26]. Under this scenario, geographic constraints—notably, health conditions—have limited the movement of people and foreign investment that would have created the institutions necessary for long-term economic growth. Property rights, for example, did not enjoy constitutional protection in Central Africa because disease conditions prevented foreign settlers from establishing themselves successfully [24]–[26]. Instead of pro-trade institutions, extractive institutions were formed, and then preserved through reinforcing mechanisms over the course of modern history. In this literature too there is implicit agreement that the geography of human health has had significant impacts on economic development [24]–[27]. However, these effects are interpreted as due to the historical consequences of European colonial expansion, and are not considered intrinsically relevant to economic productivity today. Here, we query the validity of these analyses, which assumed that the underlying disease burden influenced the survival of European colonizers but not that of contemporaneous indigenous populations. The ultimate question is whether health effects are actively important today or are only a relic of history.
The distinction of whether the physical environment has systematically impacted economic productivity directly or only indirectly is important for both practical and theoretical reasons. If health is a fundamental ingredient of economic growth, then health care and nutrition would be essential components of macroeconomic strategies for poor countries, and would also be justified targets of foreign economic aid. However, if appropriate economic institutions are the sole significant barriers to economic development, then such aid may have no long-term economic benefits and would only be justified on humanitarian grounds [28].
There are enormous implications for how we understand broad-scale economic processes if they are systematically coupled to biogeographic and ecological phenomena. The literature on the ecology of disease transmission and evolution suggests intrinsically different behavior of infectious and parasitic disease than is typically assumed by economic models, and raises the importance of initial conditions on long-term outcomes [29]–[31]. An important example of the role of ecological processes on shaping human disease burdens is represented in the growing literature on biodiversity and health [32],[33]. Because VPBDs are dependent on other host species, competing parasites, and predators, their abundance may be sensitive to assemblages of other organisms in the ecosystem. Generally, high species densities increase the number of species that prey on disease vectors and free-living parasites. Lyme disease and malaria are but a few examples of diseases that have been documented to increase with the loss of other species in their food webs [34]–[36]. However, there is also evidence that diversity of plants, mammals, and birds are broadly correlated with diversity of human diseases [37]. This hypothesis is further supported by the fact that biodiversity and human disease burdens are also correlated along a latitudinal gradient.
The possibility that these economic-ecological systems are coupled creates challenges for measuring causal pathways and points to the importance of scientific knowledge for informing statistical analysis. Here, we rely on the latitudinal gradient in income as a unifying framework to pursue a question of significance to the ecology, public health, and economic development literature: what are the relative effects of the burden of VBPDs and per capita income on each other? In pursuit of this question, we develop a statistical model that addresses an independently important question in disease ecology: what is the general impact of species diversity on the burden of VBPDs? To measure these relationships, we estimate simultaneous equations of per capita income and the burden of VBPDs, controlling for a range of factors. We find that the latitudinal gradient in income is explained by both the quality of institutions and the burden of VBPDs. The burden of VBPDs is, in turn, determined by underlying ecological conditions. In particular, it is predicted to rise as biodiversity falls.
The primary challenge for understanding relationships between the ecology of human health and global patterns of economic development through statistical analysis of country-level indicators is the problem of endogeneity [38]: economic activity is hypothesized to be both a cause and a consequence of health. Simple ordinary least squares regression analysis would therefore produce biased estimates.
Endogeneity problems are addressed in econometrics through structural equation methods that rely on instrumental variables (IVs) in multi-stage regressions (for details on IVs see Methods) [39]. IVs must be “relevant” and “excludable”—i.e., correlated with an endogenous explanatory variable of interest but not independently correlated with the dependent variable. There have been a number of studies that have attempted to measure the economic impacts of disease through IV methods [16],[23],[24],[26],[40],[41]. All such studies are limited by a general tradeoff between using broad-based health indicators (such as life expectancy or disability-adjusted life years [DALYs]), which are likely to have the most significant economic impacts, and identifying plausible instruments that are not independently correlated with income. While narrower health indicators, such as specific infectious diseases, are easier to instrument for, their effects on aggregate outcomes are more difficult to measure. As a result, conclusions from this literature have been challenged based on questions of the legitimacy of the instruments [42],[43].
In light of these issues, we focus on the per capita burden of VBPDs as our health indicator; this has several advantages. First, VBPDs have been especially implicated in impacting economic growth. While many directly transmitted diseases, such as measles and influenza, are known to have had significant impacts on global mortality rates, their systematic relationship to economic growth over long time scales is less direct. Their high rates of transmission and short infectious periods are associated with rapid acquisition of host immunity, which often lasts a lifetime. Many directly transmitted diseases are also known as “crowd diseases” and tend to be associated with modern economically driven urbanization, and are less dependent on external environmental conditions. In contrast, VBPDs, such as malaria, leishmaniasis, schistosomiasis, ascariasis, and hookworm, are more often associated with longer infectious periods, diminished immunity, and serial reinfection. They spend much of their life cycle outside of the human host in other animal hosts or free-living stages, and are thus especially dependent on external environmental conditions [44],[45]. While etiologically varied, their common ecological properties provide a basis for instrumentation.
We accordingly use a structural equation modeling approach that estimates two simultaneous equations for income and the disease burden, using relevant geographic and ecological variables as IVs [46]. A schematic of the analysis is presented in Figure 2, which corresponds to the following structural equations:(1)(2)where M represents the natural log of per capita income, and the subscript i corresponds to the country; D represents the natural log of per capita DALYs lost to the following VBPDs: malaria, trypanosomiasis, Chagas disease, schistosomiasis, leishmaniasis, lymphatic filariasis, onchocerciasis, dengue, Japanese encephalitis, ascariasis, trichuriasis, and hookworm [1]; and I is a composite index of six World Bank Governance Indicators (WGI): voice and accountability, political stability and absence of violence, government effectiveness, regulatory quality, rule of law, and corruption [47]. The variable, L, represents distance in latitude from the equator; T is a dummy variable for whether the country is located in the tropics; K is a dummy variable for whether the country is landlocked; E is the natural log of the per capita value of oil, natural gas, and coal production; B is a biodiversity index based on the species richness of plants, birds, and mammals; S is a dummy variable for whether the country is an island; and and are error terms. All variables are for the year 2002 unless otherwise noted. The model structure is discussed in detail in the Methods section, which also presents analysis of a wide range of alternative model specifications. More details on the variables can be found in Table S1 (Text S1).
Table 1 presents the results of our analysis, which tells a coherent story of the relationship between the geography of VBPDs and income (R2 = 0.84). The coefficient estimate of the impact of VBPDs on income, γ1, is −0.40, and is significant at the 1% level. This suggests that the average tropical country, with a logged per capita burden of VBPDs of 1.99, would more than double their per capita income if their disease burden were reduced to that of an average temperate country of 0.19. The effect of VBPD burden on income is also found to be statistically significant in all other supplementary analyses (Methods). Other statistically significant explanatory variables for income are the quality of institutions (γ2 = 0.38), the value of primary energy production (γ5 = 0.12), and landlocked status (γ4 = −0.54). These results broadly echo general conclusions from the literature [13],[48]. The fitted values of the model are presented along with the observed values in Figure 3 (left panel).
The model for the VBPD burden also appears to be well-specified, with an R2 of 0.75 and statistical significance at the 1% level for most of the explanatory variables. Consistent with the literature, the VBPD burden falls with income (β1 = −0.16), absolute latitude (β2 = −2.99), island status (β5 = −0.63), and rises discretely in the tropics (β3 = 0.96). The coefficient estimate for biodiversity (β4 = −0.29) is significant at the 1% level and suggests that if a country with a relatively high biodiversity index of 663 (such as Indonesia), were to lose 15% of its biodiversity, the burden of VBPDs would be expected to rise by about 30%. Figure 3 (right panel) presents the VBPD burdens along with the fitted values. Figure 4 (left panel) presents the biodiversity index along the latitudinal gradient, and Figure 4 (right panel) depicts the partial correlation of biodiversity and the burden of VBPDs.
As far back as Darwin and Wallace's theory of evolution, which was inspired by Malthus' An Essay on the Principle of Population, natural scientists have systematically borrowed theoretical approaches from economics. In the modern era, economic tools such as game theory, optimization theory, and time series analysis, have significantly contributed to our understanding of a range of biological systems, from the evolution of pathogen virulence and animal behavior, to the analysis of population dynamics and ecosystem structure [49]–[55]. However, with a few exceptions [56],[57], integration in the reverse direction (from biology to economics) has lagged behind, leaving many open questions on broad-based biological determinants of economic growth.
The economic conditions of the extremely poor are, indeed, largely due to biological processes, which are manifest in health status [58],[59]. Infectious and parasitic diseases effectively “steal” host resources for their own survival and transmission [60],[61]. These within-host processes at the individual level scale up to global patterns of poverty and disease, and are evident along a latitudinal gradient. What drives these patterns?
There are significant differences between the respective approaches of economics and the natural sciences to understanding the importance of geographic and latitudinal variation. Correlated with latitude is a seemingly endless list of biophysical and socioeconomic phenomena, from soil quality and biodiversity to per capita income and religious diversity. Understanding the latitudinal gradient in biodiversity, for example, is one of many unifying questions in the search for underlying principles of biological organization. Scientists have thus addressed the problem with a correspondingly wide range of approaches and scales of analysis, from population genetics and kinetic theory to population, community, and ecosystem ecology [6]–[10],[62]. The result has been a number of competing paradigms as well as some important consensuses.
The latitudinal gradient in income, in contrast, has not been widely used to explore underlying principles in economics, and does not generally serve as a basis for integration with the natural and physical sciences. One of the most influential explanations in the economics literature is that it is merely an historical artifact, due to the process of colonial expansion from Europe [24]–[27].
Methodologically, one challenge to understanding the relationship between geography, health, and economic development is a lack of scientifically based IVs. For example, [24] used settler mortality rates as an IV for institutions, relying on the assumption that they influenced the formation of institutions but are independent of indigenous health conditions. This finding contradicts basic knowledge in microbiology and epidemiology. Vector-borne diseases, such as malaria, continue to be among the dominant causes of morbidity and mortality of tropical populations, just as they were of colonial settlers; partial immunity is acquired among those (foreign or indigenous) who are able to survive repeated infections [63],[64].
The analysis presented here is based on an opposing hypothesis: VBPDs, while influenced by socioeconomic factors, are also determined by independent ecological processes, thus explaining their geographic signature. Disease conditions have, in turn, persistently influenced economic productivity. Our statistical model is derived from these conceptual differences and accordingly estimates income and the burden of VBPDs simultaneously. We find that the burden of VBPDs has had a substantial and statistically significant impact on per capita income after controlling for other factors. This result stands for a wide range of model specifications.
Among the ecological variables that are found to influence the burden of VBPDs, biodiversity is notable. There is an emerging literature on the relationship between biodiversity and human health, which emphasizes that VBPDs are part of broader ecosystems, and their prevalences are dependent on densities of natural predators, competitors, and other host species [32],[33]. However, understanding broader aggregate relationships have been confounded by three important considerations: (1) general biodiversity indices and disease burdens are positively correlated along a latitudinal gradient [30],[37]; (2) biodiversity and poverty are highly correlated [65]; and (3) the relationship between ecosystem structure and the disease burden may be highly variable over time and space, depending on the specific diseases and specific ecological assemblages [32]. Because of these different factors, a general theory of the effect of biodiversity on VBPDs does not exist. After accounting for the effects of income, geography, and other relevant confounders, we find that biodiversity is predicted to lower burdens of VBPDs. Given the inherent underlying complexity, a fuller understanding requires more detailed studies of these relationships across disease types and ecozones.
The policy implications of these results are straightforward: (1) health conditions have influenced the ability of economies to grow over the long-term, as indicated in differences in contemporary levels of per capita income, and (2) well-functioning, diverse, ecosystems can serve public health interests. The health benefits of biodiversity therefore constitute an ecosystem service that can be quantified in terms of income generated. The theoretical implications may be equally important: economic development is coupled to ecological processes. Such integrated approaches between economics and the natural sciences are therefore necessary for explaining economic heterogeneity around the world and for identifying policies that can lead to sustainable global health and economic development.
Table 1 presents the results of two simultaneous equations estimated from a two-step IV method. For a better understanding of the data and methods, here we first heuristically present a simple example of our statistical model, which is used as a foundation from which we systematically build in control variables. The primary goal of this study is to measure the simultaneous effects of the burden of VBPDs and the distribution of income on each other. In the process of controlling for confounders we address a secondary objective, which is to measure the effect of biodiversity on disease. For heuristic purposes, we begin with a regression model of per capita income as the dependent variable and the burden of VBPDs as an explanatory variable. This approach is guided by a couple of basic statistical considerations, such as avoiding omitted variable bias and simultaneity bias.
Omitted variable bias occurs if the burden of VBPDs is correlated with other variables that are not included in the regression model but are themselves correlated with per capita income. It can be addressed by including the appropriate independent variables into the analysis, the choice of which is guided by theory and previous empirical work. In our preliminary analysis, we control for latitude, which is the most conspicuous variable that is correlated with VBPDs and also may be related to economic activity through other indirect mechanisms.
Simultaneity bias occurs when the explanatory variable is itself a function of the dependent variable. This is a serious issue in our study because poverty is known to be an underlying cause of disease. The standard approach to overcoming simultaneity bias in the econometrics literature is through the use of IVs in a structural equation model [66]. The basic requirements for the IVs are (1) they are correlated with the endogenous explanatory variable (“relevance”) and; (2) they are uncorrelated with the error term (“excludability”) (see Assumptions and Limitations in Text S1 for more discussion of IV methods).
Identifying IVs for the burden of VBPD presents an opportunity for disease ecology to inform our understanding of the role of health on economic development. Two IVs for VBPDs that we test in this preliminary analysis are island status and biodiversity. Island status is a natural choice for an IV because: (1) ecological theory tells us that islands should generally have lower disease burdens due to lower rates of immigration/transmission and higher rates of extinction/eradication [35],[67]; and (2) island status is not independently important for economic growth in ways unaccounted for in the model. The characteristics of islands that could have economic relevance is their size and access to ports. Because we do not have complete data for many small islands, the island countries that we include cover a wide range of sizes, locations, and histories (discussed in more detail in Assumptions and Limitations in Text S1). We account for port access with a dummy variable for landlocked countries in subsequent models. These properties of the IVs are discussed in more detail in the section, Assumptions and Limitations of Instrumental Variables in Text S1.
Biodiversity, however, is a potentially more controversial choice for an IV because the literature on the relationship between biodiversity and health is ambiguous. On the one hand, biologically diverse ecosystems are thought to regulate populations of parasites and vectors through predation, competition, and dilution, putting downward pressure on human disease [32],[33],[35]. On the other hand, species richness has been shown to be correlated with diversity of human pathogens, potentially increasing the burden of disease [37]. The first-stage regression is used to generate fitted values of VBPDs based on the IVs and all other exogenous variables. The first stage regression in this example is:(3)where represents the natural log of the per capita burden of VBPDs for country i; B is an index of the species richness of plants, mammals, and birds (see Table S1 for details); L is the absolute value of the latitude; and is an error term.
Column a in Table 2 presents the parameter estimates of equation (3). Column b presents results where islands are also included as IVs. Both island status (p = 0.00) and biodiversity (p = 0.00) are negative and highly statistically significant correlates of the burden of VBPDs. This is further confirmed by a simple F-test (in the case of both IVs, we test their joint significance) (p = 0.00), such that they easily satisfy the “relevance” criterion [68]. Note that the parameter estimates for biodiversity (−0.34) and islands (−0.71) in these simple first-stage regressions are very similar to the parameter estimates for the full model presented in Table 1 (−0.29 and −0.63, respectively). Figure 5 (left panel) presents the partial correlation of biodiversity and income that corresponds to the results presented in Column b of Table 2.
The second-stage regression is an estimation of the income equation. To overcome simultaneity bias, we substitute the disease independent variable with fitted values of disease from the first-stage regression:(4)where Mi represents the natural log of per capita income of country i, and is the fitted value of disease. Note that the IVs for disease (biodiversity and islands) must be excluded from this second-stage regression (otherwise the model is not “identified”). The results of the second-stage regression are presented in Table 3, and the regression line between disease and income that corresponds to Table 3 (column b) is presented in Figure 5 (right panel).
Testing the excludability criterion is not possible in models with only one IV. However, because the second specification has more IVs than endogenous explanatory variables (it is “over-identified”), we test the over-identifying restriction (Hansen's J). We find no indication that the IVs are correlated with the error term (p = 0.23) [69] (for more details see the Assumptions and Limitations of Instrumental Variables in Text S1). Despite the simplicity of equation (4), the regression has a relatively high goodness of fit (R2 = 0.52), and is highly consistent with the results from the complete analysis presented in Table 1. Specifically, VBPDs are correlated with lower income, and biodiversity is correlated with lower burdens of VBPDs. Our goal now is to test the robustness of these results through a more rigorous analysis that includes a fuller range of statistical considerations.
While equation (3) is an appropriate first-stage estimation of disease for the purposes of estimating a second-stage regression of income, it is not complete for our purposes. Because we hypothesize that income and disease influence each other, the most appropriate statistical approach is to simultaneously estimate equations for both variables. Consider the following second-stage equations of interest:(5)(6)Equations (5 and 6) represent the simplest possible set of simultaneous equations of income and disease that account for latitude, are “just-identified” (i.e., one IV per endogenous explanatory variable), and can therefore be estimated empirically. They each consist of one IV, which is, by definition, an exogenous explanatory variable in one equation that is excluded from the other equation (for details, see Assumptions and Limitations in Text S1). Landlocked status, K, is a common control variable in economics because a lack of ports is a major barrier to trade. However, being landlocked is an irrelevant factor for disease transmission and it is thus qualified as an IV for income; biodiversity, B, is the IV for disease. The fitted values, and , are generated from first-stage regressions: and .
Equations (5 and 6) are estimated via two-step generalized method of moments [66],[69] with Stata 12. The results are presented in columns 1a and 1b of Tables 4 and 5, respectively. A first-stage F-test indicates that landlocked status is a relevant instrument in this simple specification (p = 0.00).
Equations (5 and 6) represent a system of equations that are sufficient to estimate the effects of the disease burden and income on each other. As in the simpler regression results presented in Tables 2 and 3, the burden of disease predicts lower income, and biodiversity predicts lower burden of disease. In order to test the robustness of these results, we introduce a fuller range of control variables in a stepwise fashion. There are two criteria that we used in selecting these variables: (1) they have been found in the literature to be determinants of the dependent variable; and (2) they are expected to be exogenous to this system (in particular, they are not determined by income or disease; for more details, see Assumptions and Limitations in Text S1).
As mentioned above, one of the primary hypotheses of interest is that the latitudinal gradient in income is partly due to disease ecology. The most prominent competing hypothesis is that it is instead due only to economic institutions. We therefore control for the quality of institutions via a composite index of World Bank Governance Indicators (WGI), similar to other studies (Table S1; Text S1). Because institutions, like disease, are thought to be influenced by income, we also instrument for institutions. Previous studies have used settler mortality rates as IVs for institutions, based on the premise that these mortality rates directly influenced colonial expansion, but are not independently correlated with income today [24],[26],[70]. However, we do not use settler mortality for two reasons: (1) we consider it a direct indicator of disease conditions, which we hypothesize to influence income today (these studies did not separately control for general disease burdens); and (2) there is no data on settler mortality rates for most of the countries in our dataset (only for countries that were colonized). Instead, invoking the same premise as these earlier studies, we allow the IVs for disease to also serve as IVs for institutions. First-stage regression results indicate that the IVs for disease are also statistically significant predictors of institutions (p = 0.05; Table S3). Though under-identification tests indicate that the instruments are relatively weak, our inferences are unaffected whether or not institutions is included as a control variable, and whether or not it is instrumented for (these different variations are not presented here).
For income, we consider two more potential IVs: ethnolinguistic fractionalization, F, and primary energy production, E (for details, see Table S1). Ethnolinguistic fractionalization is a natural consideration because it is considered to be a barrier to trade, a potential cause of civil strife, and is accordingly a common IV in global economic studies [70]. However, over-identification restriction tests indicate that ethnolinguistic fractionalization is strongly correlated with the error term and therefore does not meet the criteria for an IV (Table 4, column 6b); this is highly consistent with recent studies by [71],[72] that the disease burden may itself influence human “assortative sociality” and thereby drive patterns of human diversity. On the other hand, the value of primary energy production (oil, natural gas, and coal) is a useful control variable because it is an exogenous source of revenue for economies. For the disease equation, we add a dummy variable for tropical countries, T, because there is overwhelming evidence that many VBPDs thrive in tropical conditions due to metabolic and ecologic reasons [73]. We do not, however, include tropics as a control variable in the income equation because preliminary analyses indicated that tropics are not statistically significant predictors of income, after controlling for other variables (i.e., latitude, disease, and institutions) (p = 0.90), but is collinear with institutions. Thus tropical conditions also serves as an IV for disease.
Tables 4 and 5 present the results of eight different specifications of the simultaneous equations estimated by two-step generalized method of moments in Stata 12 (details of the variables are in Table S1). Each of these specifications has been tested for identification (i.e., the strength of the IVs), spatial autocorrelation, and over-identifying restrictions wherever possible. The IV Moran's I test measures spatial-autocorrelation in the residuals. Statistically significant spatial-autocorrelation was not found in any of the estimates of the income equation (p-values ranged from 0.24 to 0.80), but were found in four of the eight estimates of the disease equation (p-values ranged from 0.07 to 0.54). Such spatial autocorrelation in the residuals tends to vanish when additional variables (i.e., that are geographically determined) are controlled for [74]. However, the addition of more IVs increases the possibility of violating the excludability criterion, indicated by the over-identifying restriction test. The last three model specifications suffer from this problem (p-values for over-identifying restriction test are less than 0.1 in columns 6b, 7b, and 8b, indicating that the IVs are correlated with the error term). Despite these considerations, the parameters are very consistent across all models. The best overall specification is presented in columns 5a and 5b, which has R2s of 0.84 and 0.76, is well-identified with strong instruments and no statistically significant spatial autocorrelation. This system is represented by the following reduced-form equations that correspond to structural equations (1 and 2):(7)(8)The first stage regressions for the estimation of the income equation (7) are:(9)(10)Table S3 presents the outcomes of these first stage regressions. The identification criteria are easily satisfied. Island status and biodiversity are both significant negative predictors of the disease burden in both simple and more complex models. The first stage regression for the estimation of the disease equation (8) is:(11)which is presented in Table S4. The identification criteria are easily satisfied here as well. The landlocked and energy variables are especially effective predictors of income. The estimated effect of biodiversity on disease, and of disease on income, were statistically significant for all model specifications.
|
10.1371/journal.ppat.1000065 | Targeting of Pseudorabies Virus Structural Proteins to Axons Requires Association of the Viral Us9 Protein with Lipid Rafts | The pseudorabies virus (PRV) Us9 protein plays a central role in targeting viral capsids and glycoproteins to axons of dissociated sympathetic neurons. As a result, Us9 null mutants are defective in anterograde transmission of infection in vivo. However, it is unclear how Us9 promotes axonal sorting of so many viral proteins. It is known that the glycoproteins gB, gC, gD and gE are associated with lipid raft microdomains on the surface of infected swine kidney cells and monocytes, and are directed into the axon in a Us9-dependent manner. In this report, we determined that Us9 is associated with lipid rafts, and that this association is critical to Us9-mediated sorting of viral structural proteins. We used infected non-polarized and polarized PC12 cells, a rat pheochromocytoma cell line that acquires many of the characteristics of sympathetic neurons in the presence of nerve growth factor (NGF). In these cells, Us9 is highly enriched in detergent-resistant membranes (DRMs). Moreover, reducing the affinity of Us9 for lipid rafts inhibited anterograde transmission of infection from sympathetic neurons to epithelial cells in vitro. We conclude that association of Us9 with lipid rafts is key for efficient targeting of structural proteins to axons and, as a consequence, for directional spread of PRV from pre-synaptic to post-synaptic neurons and cells of the mammalian nervous system.
| Alpha herpesviruses are common mammalian pathogens (e.g. herpes simplex and chickenpox virus infect humans). These viruses enter and spread in and out of the mammalian nervous system, a defining hallmark of their lifecycle and potential pathogenesis. Neurons are polarized cells, and the movement of certain cellular proteins is highly restricted to the cell body, to dendrites, or to axons. Indeed, the axon harbors only a small subset of neuronal proteins. Thus, it is remarkable that these viruses efficiently gather and move their structural components from the cell body to the axon after infection of neurons. Previously, we identified a small viral membrane protein (known as Us9) that mediates the efficient targeting of virus particles to axons of infected neurons. We now report that Us9 must localize to “lipid raft” domains, specialized regions within cellular membranes to promote axonal targeting. Reducing the affinity of Us9 for lipid rafts dramatically reduces sorting of structural proteins to axons. This is the first report, to our knowledge, to implicate lipid rafts in targeting alpha herpesvirus structural proteins to axons, an essential step for spread of infection in the mammalian nervous system.
| Neurons are highly polarized cells with distinct biochemical and functional properties particular to the cell body (soma), dendrites, and the axon. Soluble, cytosolic proteins likely move into axons and dendrites by bulk flow, while membrane proteins are highly restricted to the somatodendritic membrane, the axonal membrane, or within vesicles [1]. Sorting of membrane proteins to the somatodendritic or axonal compartments of neurons is similar to the sorting mechanism of membrane proteins either to the basolateral or to the apical membranes of epithelial cells [2]. Both cell types share a common component for compartmentalization of their membrane milieu: lipid raft microdomains. Rafts often are defined biochemically as detergent-resistant membranes (DRMs) or detergent-insoluble glycolipid complexes (DIGs) [3]–[5]. These protein-lipid complexes are highly enriched in cholesterol and sphingolipids, and float freely within the liquid-disordered state of the lipid bilayer. Furthermore, they are small and dynamic, often ordering themselves into larger, more stable rafts through protein-protein and protein-lipid interactions [6],[7].
Lipid rafts/DRMs play a key role in the axonal sorting of membrane proteins during neuronal maturation, including axonal growth and guidance [8]. Ledesma et al. showed that two axonally-targeted membrane proteins, Thy-1 and influenza virus hemagglutinin (HA), interact with sphingolipid-cholesterol rafts in rat hippocampal neurons [9]. These microdomains normally were resistant to detergent treatment at 4°C. However, reducing the levels of cholesterol and sphingolipids resulted in the detergent-solubility of both proteins, as well as aberrant sorting into axons [9]. The increased synthesis of sphingomyelin during neuronal maturation was critical for the formation of protein-lipid complexes, an essential step for the targeting of randomly distributed GM1 ganglioside in immature neurons to the axon of fully mature neurons [4].
Alpha herpesviruses (e.g. herpes simplex virus (HSV) and pseudorabies virus (PRV)) replicate and traffic within polarized neurons, a strategy conducive to their lifestyle in the host peripheral nervous system (PNS). Infection begins with virion entry at mucosal surfaces and spread of infection between cells of the mucosal epithelium. The PNS is infected through axon termini innervating this region, and subsequent trafficking of capsids to the cell body. It is here that a reactivatable, latent infection is established that persists for the life of the host [10]. A well known, but poorly understood observation, is that upon reactivation from the latent infection, α-herpesviruses rarely enter the central nervous system despite having what seems to be two rather similar choices: cross one synapse and infect the central nervous system (rare) or traffic back down the axon to the initial peripheral site of infection (very common). Inherent in this choice is the fact that viral proteins must be targeted to axons, a highly specialized neuronal compartment restricted to only a subset of neuronal proteins. The primary problem is to identify the mechanisms that gather and sort the many viral structural proteins to this compartment.
The Us9 gene product, a small, tail-anchored type II membrane protein, has been of interest in understanding the mechanisms of anterograde transport of alpha herpesviruses in the mammalian nervous system [11]–[14]. PRV Us9 directs viral membrane proteins and capsids to axons of infected neurons [15],[16], an absolute requirement for anterograde neuron-to-cell and neuron-to-neuron spread of the virus [11],[17]. Our current data demonstrate that in the absence of PRV Us9, axons do not contain vesicles with viral glycoproteins or mature virus particles [15]. However, the mechanism by which this occurs has not been elucidated.
Favoreel et al. reported that PRV glycoproteins gB, gC, gD, and gE were associated with lipid rafts on the surface of infected swine kidney cells and monocytes [18]. PRV Us9 is essential for the trafficking of these glycoproteins into the axon of dissociated superior cervical ganglia (SCG) neurons [16]. These observations, in conjunction with reports that raft association is critical for axonal targeting of certain neuronal membrane proteins [4],[9],[19] led us to investigate whether PRV Us9 was associated with detergent-insoluble lipid rafts.
All PRV strains were propagated on porcine kidney (PK15) cells at a low multiplicity of infection (MOI) for 48 hours and then collected by scraping cells into the conditioned medium as described previously [15]. The wild-type PRV strain Becker and its derivatives, PRV 99, PRV 160, and PRV 162 were described previously [11], [20]–[22]. PRV 99 is deleted for the sequence encoding both gE and gI. PRV 160 (Us9-null) contains a nonsense stop mutation at position 4 in the Us9 open reading frame. PRV 162 encodes a mutant Us9 protein in which the nucleotide sequence encoding amino acids 46 to 55 have been removed (i.e. the acidic cluster region).
To construct PRV 322, the wild-type transmembrane domain of PRV Us9 was replaced with that of the transferrin receptor (TfR) transmembrane domain [23],[24]. This was performed by the SOEing PCR method [25] using pML47 as the template. This plasmid contains the SpHI/MluI region of the Becker BamHI 7 fragment cloned into pBluescript SK(+).The upstream fragment was amplified using the upstream-forward primer 5′ GCATGCTCTCGCCGGTGT 3′ and the downstream-reverse primer 5′ aaatccaatcaagaaaaagacgatcacag aatagtcccatagcagatacttccactaca GCGGCGGCGTCTCCGGCG 3′. The downstream fragment was amplified using the upstream-forward primer 5′tattgctgtgatcgtctttttcttgattggatttatgattggctacttgggctattgtCGGCACGTGTAGCGAGCGGGT3′ and the downstream-reverse primer 5′ ACGCGTAGCACCACTCGG 3′. Upper case letters represent PRV sequences whereas lower case letters represent sequences introduced to synthesize the TfR TMD domain. Both the upstream and downstream fragments were gel-purified, and ligated together using the upstream-forward primer 5′ GCATGCTCTCGCCGGTGT 3′ and the downstream-reverse primer 5′ ACGCGTAGCACCACTCGG 3′. The SOEing product was cloned directly into the pCR-Blunt II-TOPO vector (TOPO Cloning Kit, Invitrogen, Carlsbad, CA) and sequenced to verify that no extraneous mutations were introduced during PCR. This plasmid was designated pML55. We then digested pML55 with the restriction enzymes BsrGI and MluI to release a ∼1.0 kb fragment that contained the Us9 gene with the TfR transmembrane domain. This fragment was used to replace the BsrGI/MluI fragment of pML68, a pT7Blue vector backbone containing a SalI/SpHI fragment spanning a large portion of the unique-short region (a splice product of the SalI/MluI fragment of pPH 2 and the MluI/SpHI fragment from pGS166). It includes a portion of the gD gene, gI, gE, Us9, and virtually the entire Us2 gene. This construct (designated pML73) was digested with EcoRI/HindIII and cotransfected with purified PRV 165 genomic DNA [26] into PK15 cells. Plaques not expressing EGFP were identified using a Nikon Eclipse TE300 inverted epifluorescence microscope and then subjected to three rounds of plaque purification. This recombinant virus was named PRV 319. Initial characterization of PRV 319 revealed that Us9-TfR unexpectedly formed a dimer inside infected PK15 and PC12 cells (unlike the monomeric form of wild-type Us9). We discovered that two cysteine residues at the N- and C-terminus of the TfR transmembrane domain were responsible for disulfide-bonded dimer formation and protein acylation [27]. We deleted these codons from pML55 using the Stratagene Quickchange II site-directed mutagenesis kit (Stratagene, La Jolla, CA). This construct, denoted pML86, was digested with the restriction enzymes BsrGI and MluI and the 1 kb fragment was cloned into the pML68 shuttle vector. This construct (designated pML88) was digested with EcoRI/HindIII and cotransfected with purified PRV 165 genomic DNA [26] into PK15 cells. Non-green plaques were identified and purified as described above. This recombinant virus was designated PRV 322. The presence of the Us9-TfR chimeric protein, as well as the presence of gE and Us2 (genes directly upstream and downstream of Us9, respectively) were verified by western blot of infected cell lysates.
Antibodies used in this report include: polyclonal rabbit antiserum recognizing Us9 ([28]; used 1∶500 by WB and 1∶200 by IF), the gE cytoplasmic domain ([29]; used 1∶1000 by WB), and PRV antigens (Rb134 [21],[22]; used 1∶10,000 by IF), polyclonal goat antiserum recognizing gB, gC, Us2, and UL34 ([22],[30]; all used 1∶1000 by WB), and mouse monoclonal antibody recognizing transferrin receptor (Zymed, San Francisco, CA). The rabbit polyclonal gH antiserum was a kind gift from Thomas Mettenleiter [31]. All secondary Alexa fluorophores (used at 1∶500) were purchased from Molecular Probes.
A detailed protocol for culturing and differentiating PC12 cells prior to infection with PRV has been published previously [32]. Briefly, we coated the surface of a 150 mm dish with rat tail collagen (type 1) at a concentration of 5 µg/cm2 in 0.02 N ascetic acid for 1 hour. The dish was washed gently three times with sterile water, and 20 ml of complete growth medium (85% RPMI, 10% horse serum, 5% fetal calf serum) was added to the plate. A 100 mm dish with undifferentiated PC12 cells (∼80% confluency) was split 1∶5, and triturated extensively with a sterile 5 ml plunger and 22-gauge needle to dissociate cell clumps. After trituration, 1 ml of cell suspension was added to the 150 mm dish, and cells were allowed to attach overnight in a 37°C incubator. To produce a culture of differentiated PC12 cells, growth medium was replaced with differentiation medium (RPMI, 1% horse serum, nerve growth factor (100 ng/ml)) when cells became ∼30% confluent. Differentiation medium was replaced every other day for 12 days, at which time an extensive network of neurites was visible. Undifferentiated PC12 cells were propagated in growth medium to ∼70% confluency prior to infection with PRV.
Live imaging of viral capsids was performed using a Leica SP5 with an HCX Plan Apochromat 63×1.3 NA glycerin objective. Approximately 15 optical sections were acquired in 0.5-µm steps through the plane of the neurites and cell body. Each frame of the movie is a 2D projection of one stack of images. Prior to imaging, 1 M HEPES was added to the RPMI medium to a final concentration of 25 mM. PC12 cells were cultured on MatTek Corp. glass-bottomed dishes (http://www.glass-bottom-dishes.com/). The dish was warmed to 37°C employing a DH40i Micro-incubation System (Warner Instrument Corp.) run at constant voltage (∼5.5 volts). A 488 nm laser line was used for GFP excitation (10% intensity), with emissions collected from 495 to 553 nm. Images were acquired employing a 1.5 airy unit detector pinhole and scanning at 700 Hertz. All figures were assembled in Adobe Photoshop 7.0.1. Movies were created using ImageJ 1.32j software (National Institutes of Health).
Flotation of lipid rafts by Optiprep™sucrose gradient is well documented in the literature [33]–[35], though subtle differences exist between protocols. We followed those described previously for PRV-infected SK cells [18] and uninfected PC12 cells [36]. Undifferentiated and differentiated PC12 cells were cultured in a 150 mm dish as described above (∼107 cells). Cells were infected at a high multiplicity of infection (MOI = 10) with PRV Becker, PRV 99, PRV 162, or PRV 322. At 12 hours post-infection (hpi), cells were collected in a 50 ml conical tube, and washed twice with cold RPMI medium by brief centrifugation at 3,000 rpm. Cells were lysed with 1 ml of lysis buffer consisting of 1% TX-100 in TNE buffer (25 mM Tris HCl [pH 6.8], 150 mM NaCl, 5 mM EDTA), protease inhibitor cocktail (Roche Diagnostics GmbH, Mannheim, Germany), and 5 mM iodoacetamide. The lysate was homogenized by being passed 15 times through an 18-gauge needle, and then allowed to rock for 30 minutes at 4°C. At the end of the rocking period, the sample was again homogenized briefly, and then mixed with 2 ml of ice-cold 60% Optiprep™ density gradient medium (Sigma-Aldrich, St. Louis, MO). The entire 3 ml mixture was placed at the bottom of a Beckman SW41 ultracentrifuge tube (Beckman, Munich, Germany) and subsequently overlaid with 5 ml of ice-cold 30% Optiprep in TNE and 4 ml of ice-cold 5% Optiprep in TNE. Samples were centrifuged at 34,200 rpm (200,000×g) at 4°C for 20 hours. Twelve fractions were collected from the top to the bottom of the tube (1 ml each), and mixed 1∶1 with 2× Laemmli sample buffer. Samples 3–10 were electrophoresed on a 12% SDS-PAGE gel.
PK15 cells were grown to confluency in 60 mm dishes (∼3×106 cells/dish), and infected with the indicated viruses at an MOI of 10. Following a one-hour adsorption period at 37°C, the cells were rinsed once with PBS and incubated at room temperature for three minutes with 3 ml of citrate buffer (40 mM sodium citrate, 10 mM KCl, 135 mM NaCl, pH 3.0) to inactivate unabsorbed virus. The cells were rinsed three times with PBS, 2 ml of fresh medium were added, and plates were returned to the incubator. At various times post-infection, cells were scraped into the medium and frozen in two aliquots. Aliquots were freeze-thawed three times (−80°C/+37°C) and titered in duplicate on PK15 cell monolayers.
Virions were purified as described previously [28]. Briefly, three confluent 150-mm-diameter dishes of PK15 cells were infected with virus at a MOI of 10. At 16 hpi, medium was collected and centrifuged at 3,000 rpm to remove cellular debris. The clarified supernatant was layered on a 7 ml 30% sucrose-PBS cushion (w/v) in two Beckman SW28 centrifuge tubes. The tubes were then centrifuged in an SW28 rotor at 23,000 rpm for 3 hours. The sucrose cushion was removed, and the virion pellet was resuspended in 1 ml of PBS by 10 one-second pulses in a bath sonicator and gentle pipetting. Virions were then centrifuged through a 1 ml 30% sucrose cushion at 28,000 rpm for 90 minutes in an SW55ti rotor. The pelleted virions were resuspended in 100 µl PBS, and subsequently mixed with Laemmli sample buffer for analysis by SDS-PAGE.
To analyze steady-state viral protein expression, PK15 cells (∼3.0×106) were either mock infected or infected at a high MOI with Becker or PRV 322. At 6 hpi, medium was removed and infected cells were scraped into 330 µl of PBS and 170 µl of 3× Laemmli sample buffer. Samples were homogenized with an insulin syringe, boiled for 2 minutes, and subjected to sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE). Preparation of lipid raft fractions and purified virions for SDS-PAGE is described above. All gels were transferred to an Immobilon-P membrane using a semidry transfer apparatus following the manufacturer's instructions (Labconco, Kansas City, MO). Following transfer, membranes were immediately put into blocking solution (2% bovine serum albumin in TBS (50 mM Tris, 200 mM NaCl) (w/v), 0.1% Tween-20 (v/v)) and incubated for 15 minutes at RT. Primary antibody was diluted in 2% BSA-TBS-Tween solution, and allowed to rock for 30 minutes with the membrane. GM1 was detected using biotinylated cholera toxin B subunit (1 µg/ml, Sigma-Aldrich, St. Louis, MO) and a HRP-streptavidin conjugate (1∶1000, Pierce, Rockford, IL). Following the 30 minute incubation, the membrane was washed three times with TBS-Tween, and placed in HRP-conjugated secondary antibody (1∶10,000, KPL, Gaithersburg, Maryland) diluted in 2% BSA-TBS-Tween for 30 minutes. The membrane was then washed as previously described, and proteins were visualized with the ECL Plus Western blotting detection system (GE Healthcare).
PK15 cells were grown on glass coverslips (∼20% confluence) and transfected with mammalian expression vectors encoding GFP (pEGFP-N1), Us9-GFP (pBB14), and Us9-TfR-GFP (pML92) using Lipofectamine 2000 as directed by the manufacturer's instructions (Invitrogen, Carlsbad, CA). At 24 hours post-transfection, cells were washed with PBS and fixed with 4% paraformaldehyde in PBS for 10 min. After fixation, cells were washed three times with PBS, and stained with Hoechst 33342 (Invitrogen, Carlsbad, CA). Samples were then mounted on a glass slide using Aqua poly/mount (Polysciences, Warrington, PA) and allowed to air dry for 24 hours prior to imaging. Direct fluorescence was visualized using an inverted epifluorescence microscope and the appropriate excitation and emission filters. For indirect immunofluorescence experiments, PK15 cells were grown to 30% confluence on glass coverslips and infected with Becker, PRV 160, and PRV 322. At 6 hpi, cells were washed three times with phosphate-buffered saline (PBS), then fixed with 4% paraformaldehyde in PBS for 10 min. After fixation, cells were washed three times with PBS, and permeabilized for 3–5 minutes with 0.5% Triton X-100 in PBS. After permeabilization, Us9 antiserum was diluted 1∶200 in wash buffer (PBS, 3% BSA, 0.5% saponin) and added to cells for 1 hour. Primary antibody was then removed, and the sample washed three times. Next, secondary antibodies were added to the sample and incubated for 1 hour. Secondary antibody was then removed and the sample was washed an additional three times. Samples were mounted on a glass slide using Aqua poly/mount and allowed to air dry for 24 hours prior to imaging. Optical sections were acquired using a Leica SP5 confocal microscope with a 63×/1.3 NA oil objective.
Detailed protocols for dissecting and culturing PNS neurons from rat embryos have been published previously [32]. Briefly, sympathetic neurons from the superior cervical ganglia (SCG) were dissected from rat embryos at embryonic day 15.5 to 16.5 (Sprague-Dawley rats, Hilltop Labs, Inc., Scottdale, PA) and incubated in 250 µg/ml of trypsin (Worthington Biochemicals) for 10 min. Trypsin inhibitor (1 mg/ml; Sigma Aldrich) was added to neutralize the trypsin for 5 min and then removed and replaced with neuron culture medium (described below). Prior to plating, the ganglia were triturated using a fire-polished Pasteur pipette and then plated in the S compartments of the Teflon ring. The Teflon ring was placed within a 35-mm plastic tissue culture dish coated with 500 µg/ml of poly-DL-ornithine (Sigma Aldrich) diluted in borate buffer and 10 µg/ml of natural mouse laminin (Invitrogen). The neuron culture medium is serum free and consists of Dulbecco's modified Eagle medium (Invitrogen) and Ham's F12 (Invitrogen) in a 1∶1 ratio. The serum-free medium was further supplemented with 10 mg/ml of bovine serum albumin (Sigma Aldrich), 4.6 mg/ml glucose (J. T. Baker), 100 µg/ml of holotransferrin (Sigma Aldrich), 16 µg/ml of putrescine (Sigma Aldrich), 10 µg/ml of insulin (Sigma Aldrich), 2 mM of L-glutamine (Invitrogen), 50 µg/ml or U of penicillin-streptomycin (Invitrogen), 30 nM of selenium (Sigma Aldrich), 20 nM of progesterone (Sigma Aldrich), and 100 ng/ml of nerve growth factor 2.5S (Invitrogen). Two days after plating, the neuronal cultures were treated with 1 µM of antimitotic drug cytosine ß-D-arabinofuranoside (AraC; Sigma Aldrich) to eliminate any nonneuronal cells. The neuron culture medium was replaced every 3 to 4 days, and cultures were maintained in a humidified, CO2-regulated, 37°C incubator. All experimental protocols related to animal use were approved by the Institutional Animal Care and Use Committee of the Princeton University Research Board under protocol number 1691 and are in accordance with the regulations of the American Association for Accreditation of Laboratory Animal Care and those in the Animal Welfare Act (public law 99-198).
Protocols for assembling the trichamber system have been described previously [17],[37]. Tools and reagents, including the Teflon rings (Tyler Research, Alberta, Canada) and the silicone grease-loaded syringe (Dow Corning), were sterilized by autoclaving prior to assembly. Tissue culture dishes (35 mm) were coated with 500 µg/ml of poly-DL-ornithine (Sigma Aldrich) followed by 10 µg/ml of natural mouse laminin (Invitrogen), and then washed and dried; the bottom surface of each dish was etched with a pin rake, creating a series of 16 evenly spaced grooves. We used a silicone grease-loaded syringe attached to an 18-gauge truncated hypodermic needle to apply a thin, continuous strip of silicone grease over the entire bottom surface of the Teflon ring. Next, a 50-µl drop of neuron medium containing 1% methocellulose (serum free) was placed in the center of each tissue culture dish covering the etched grooves. This step prevents the seal from being entirely devoid of moisture, which is needed for axon penetration and growth between the grooves. Finally, the silicone grease-coated ring was gently seated on the tissue culture dish or the surface of the 35-mm dish such that the etched grooves spanned all three compartments, forming a watertight seal between compartments. Neuron medium was then placed in all three compartments immediately after the chamber was assembled. Once the SCG neurons were dissected and dissociated, approximately one half of a single ganglion was plated into the S chamber. Neuron cultures were then maintained according to the protocols for culturing neurons reported above.
Neurons were cultured for approximately 2 weeks in the trichamber system (on 35 mm tissue culture dishes) with frequent medium changes. After 2 weeks, axon penetration into the M and N compartments was assessed visually and only cultures with comparable axon densities were used for experiments. After axons penetrated the N compartment, nonneuronal epithelial cells (PK15 cells) permissive for PRV infection were plated in the N compartment. The neuron medium in the N compartment was supplemented with 1% fetal bovine serum and the cells were allowed to attach and expand for 24 h prior to any experiment. Once the target cells in the N compartment were plated, neuron medium containing 1% methocel was placed in the M compartment. After 30 min, the neuronal cell bodies in the S compartment were infected with virus diluted in neuron medium (approximately 105 PFU). After 1 h, the viral inoculum was removed and replaced with neuron medium. The chambers were then incubated in a humidified 37°C incubator for 24 hours. Both intracellular and extracellular virions in the S and N compartments were carefully harvested by scraping the bottom of the dish with the pointed end of a gel-loading tip. The cells and medium were then pooled and freeze-thawed, and titers were determined.
Trichambers were assembled on UV-sterilized Aclar strips. Neurons were cultured and infected as described above. At 16 hpi, all compartments were washed twice with PBS containing 3% BSA (PBS/BSA), chambers gently lifted, and silicone grease scraped off the Aclar strips. Samples were then fixed with 4% paraformaldehyde in PBS for 10 minutes. Fixative was washed away with three PBS/BSA rinses, after which the samples were permeabilized using a solution of 0.5% saponin and 3% BSA in PBS (PBS/BSA/SAP). Incubations with primary and secondary antibodies were performed for one hour in PBS/BSA/SAP. Following two rinses with PBS/BSA/SAP and one rinse with distilled water, the samples were mounted on glass slides using Aqua poly/mount (Polysciences). Images were collected on a Perkin-Elmer RS3 spinning disk confocal microscope using a 40×1.3 NA oil objective. Z-stacks obtained in 1-micron steps were collapsed and analyzed in Perkin-Elmer ImageView software.
The characterization of lipid rafts has largely been based on their resistance to detergent solubilization at 4°C [3]. This approach continues to be a useful tool to assess the affinity a protein has for lipid rafts when formed at physiologic temperatures [38],[39]. Analyses performed on changes in DRM partitioning (due to a physiological stimulus) have provided the foundation for extensive work on the role of lipid rafts in signal transduction [40]–[42], membrane trafficking [43],[44], and pathogenesis [45]. Accordingly, we investigated the affinity of Us9 for DRMs, as well as several viral glycoproteins whose axonal localization is dependent on Us9 [16].
It is difficult to culture a sufficient number of primary rat neurons to perform large biochemical analyses. Therefore, we used PC12 cells, a widely used rat pheochromocytoma cell line that responds to nerve growth factor (NGF) and acquires many of the characteristics of sympathetic neurons [46]. Differentiated PC12 undergo polarized protein sorting, and cell bodies stain for nonphosphorylated neurofilament H (a somatodendritic marker) while axons stain exclusively for phosphorylated neurofilament H (an axonal marker) [32]. This is consistent with mature, sympathetic SCG neurons [47]. It has been reported that PC12 cell are susceptible and permissive to PRV infection [48], and that a PRV GFP-VP22 fusion protein moves inside neurites with fast axonal kinetics [49]. However, it was unclear whether the Us9-null phenotype in SCG neurons, i.e. a complete block to axonal sorting of viral structural proteins [15],[16], could be recapitulated in this neuron-like cell line (a critical experiment to ensure that PC12 cells could be used to study Us9 biology). We recently reported that in the absence of Us9, GFP-tagged capsids were unable to sort into axons of live SCG neurons [15]. Therefore, we utilized a similar live-cell imaging approach to examine the axonal sorting of GFP-tagged capsids in differentiated PC12 cells. Cells were infected with PRV GS443, a recombinant PRV strain that expresses GFP fused to VP26, a capsid protein [50]. After 12 h, capsid puncta were readily observed trafficking in the anterograde direction within neurites of PC12 cells (n = 20) (Figure 1A, Video S1). Importantly, when differentiated PC12 cells were infected with PRV 368, a GFP-tagged capsid mutant deleted for Us9, no green puncta were observed moving in the anterograde direction (Figure 1B and 1C, Video S2). These findings were consistent with our Us9 studies in dissociated SCG neurons [15]. Interestingly, we also observed the retrograde trafficking of capsid puncta from cells infected with PRV 368 to uninfected, neighboring cells (in the absence of any anterograde sorting of virus particles in the same field of view) (Figure 1D–1F, Videos S3 and S4). This had been described previously in transneuronal spread studies on Us9 mutants in the rat visual system [11],[26], but had not been observed in tissue culture cells. It is noteworthy that we did not visualize “random” egress of GFP-tagged capsids from infected cell bodies. Capsids either sorted into axons (in the presence of Us9), or to sites of synaptic contact with other axons (transneuronal, retrograde transport). The import of this observation is unclear at present, but may suggest that alpha herpesviruses undergo directed egress from neuronal cell bodies. Overall, our findings suggest that differentiated PC12 cells recapitulate the Us9 sorting phenotypes previously observed in primary sympathetic neurons, and are an efficacious cell line to study Us9 biology (specifically that of Us9 and lipid rafts).
Therefore, we compared the raft profiles of viral membrane proteins after infection of non-polarized and polarized PC12 cells. Undifferentiated cells (∼107) were infected with wild-type Becker for 12 hours, solubilized with 1% TX-100, and subjected to a well-described “raft flotation” assay [33]–[35]. Membrane proteins that were solubilized by TX-100 remained at the bottom of the Opti-Prep gradient (40%), whereas proteins in DRMs floated to the 5%–30% interface. We used the prototypic raft and non-raft markers, GM1 ganglioside and transferrin receptor (TfR), as positive and negative controls [51]. Us9 was highly enriched in the raft fraction as compared to the soluble population (Figure 2). To test whether the flotation of Us9 was cholesterol dependent, we treated cells with methyl-cyclodextrin (MCD) prior to solubilization with detergent. MCD depletes cholesterol from cellular membranes, and therefore disrupts the structure of lipid raft microdomains [52],[53]. A 45-minute exposure of 20 mM MCD to Becker-infected cells dramatically decreased the amount of Us9 floating with the raft fraction. We found the viral glycoprotein gB to have a strong affinity for DRMs while gE and gC where not enriched in either the raft or soluble fractions. These findings were consistent with similar experiments performed in non-polarized swine kidney (SK) cells [18]. Surprisingly, PRV gH was completely solubilized by TX-100 treatment, as was transferrin receptor (the negative control). Overall, Us9 and gB have a strong affinity for DRMs in undifferentiated PC12 cells. Both gE and gC were present in the raft and soluble fractions in a ∼1∶1 ratio, whereas gH was completely soluble.
To determine the effect of cell polarization on DRM partitioning, we cultured PC12 cells in low serum conditions in the presence of nerve growth factor for 12 days. This treatment allows for extensive neurite outgrowth from the cell bodies, as well as separation of somatodendritic and axonal marker proteins [32]. We infected cells with PRV Becker for 12 hours, and subjected lysates to raft flotation analysis as performed previously (Figure 3). Us9 and gB were again strongly associated with the raft fraction of the gradient, and floated with GM1. Both gE and gC were highly enriched in the DRM fraction of PC12 cells upon differentiation with NGF. This finding is consistent with the notion that polarization of neurons strongly impacts the affinity and targeting of certain membrane proteins to lipid rafts/DRMs [4]. gH was completely solubilized by TX-100 even in differentiated PC12 cells, and remained in the soluble fraction with TfR. These data suggest that Us9 and gB have a high affinity for DRMs in undifferentiated and differentiated PC12 cells, whereas gC and gE increase their association with DRMs as cells polarize and mature.
The efficient targeting of viral structural components to the axon of infected cells is dependent on both the Us9 and gE gene products [15],[16],[54]. Deletion of either gene results in the reduction of viral capsids and enveloped proteins in the axon, and subsequent reduction of anterograde spread of infection in vitro and in vivo [11],[17],[55]. Upon discovering that both Us9 and gE were present in the DRMs of infected PC12 cells, we tested whether deleting the gE/gI complex affected the ability of Us9 to target to rafts, thereby impacting its ability to function properly in anterograde transport. PC12 cells were infected with PRV 99, a mutant deleted for the gE and gI genes, and subjected to raft flotation analysis. Deletion of gE/gI had no affect on the ability of Us9 to associate with DRMs (Figure 4A), nor did the absence of gB (data not shown). These data again support the notion that Us9 has an intrinsic affinity for DRMs and is not influenced by cell polarity or the presence of two major viral DRM components. It is noteworthy that we found no aberrant targeting of gB, gC, or gE to DRMs in a Us9-null mutant (data not shown).
It is well documented that certain proteins become raft associated upon their phosphorylation during signal transduction [56]–[58]. Us9 contains a conserved acid cluster (AC) region with two serines that are phosphorylated [26], as well as a di-tyrosine motif critical for anterograde, transynaptic spread [26]. To examine whether inhibiting Us9 phosphorylation precluded Us9 association with DRMs, we infected PC12 cells with PRV 162, a mutant that expresses an altered Us9 protein lacking the acidic cluster region. DRMs were prepared from infected cells as performed previously (Figure 4B). Us9 was still enriched in the DRM fraction of the gradient despite the absence of Us9 phosphorylation (note the narrowness of the Us9 band compared to wild-type Us9 in panel A). Taken together, these data suggest that Us9 is highly enriched in DRMs, and its affinity for this lipid microdomain is not dependent on its phosphorylation state.
Both the influenza virus neuraminidase (NA) and haemagglutinin (HA) proteins (both type II membrane proteins) are highly enriched in DRMs/lipid rafts, and this association is critical for the apical sorting of these proteins in polarized MDCK cells [59]–[61]. Importantly, the transmembrane domain (TMD) of these proteins provided the determinants for apical sorting and raft association [59], [62]–[64]. This was demonstrated by swapping the TMD of transferrin receptor (a type II, non-raft associated membrane protein) for the TMD of neuraminidase [63]. Transferrin receptor was normally sorted to the basolateral membrane in polarized MDCK cells, and was efficiently solubilized by 1% TX-100. By contrast, transferrin receptor with the NA TMD was targeted to the apical cell surface and was largely insoluble to treatment with TX-100 [59]. In a reciprocal experiment, the neuraminidase TMD was replaced with the transferrin receptor TMD [62]. This chimeric protein was greatly reduced in lipid raft association, and a virus expressing this protein showed a defect of particle release from the apical cell surface.
We employed a similar approach with PRV Us9 to test whether its transmembrane domain provided the determinants for raft sorting. Both Us9 and transferrin receptor are type II membrane proteins, and have 26 amino acids within their TMD. We constructed a PRV mutant that expressed a chimeric protein with the wild-type Us9 cytoplasmic domain, a transferrin receptor TMD, and wild-type 3 amino acid ectodomain (Figure 5A). This mutant, known as PRV 322, replicates with wild-type kinetics in porcine kidney (PK15) cells. Furthermore, the Us9-TfR protein is abundantly expressed in infected cell lysates and migrates more slowly by SDS-PAGE than wild-type Us9 (Figure 5C). Expression of the upstream and downstream genes, gE and Us2 respectively, were indistinguishable from Becker and suggested that recombination of Us9-TfR into the viral genome did not have polar effects on neighboring genes. Us9-TfR is efficiently incorporated into virions (Figure 5D), along with gE and Us2, but not UL34 which is a component of primary but not mature virus particles [65].
The localization and intracellular trafficking pattern of Us9 has been studied extensively in porcine kidney (PK15) cells [20],[26],[28]. To assess whether Us9-TfR remained a functional Us9 derivative (i.e. did not have an egregious trafficking defect), we examined the steady-state localization of Us9-TfR in transfected and infected PK15 cells. We hypothesized that both Us9 and Us9-TfR would have similar localization patterns as both contain the Us9 acidic domain, the region necessary for localization to a perinuclear cellular compartment [20]. Furthermore, sorting signals for the TfR protein reside in the cytoplasmic tail, not within the transmembrane domain [66]. Thus, the TfR TMD should be functionally inert in the context of Us9 trafficking.
The localization of Us9 fused to GFP mimics the localization of wild-type Us9 inside infected cells [20],[28]. We fused Us9-TfR to GFP to visualize its steady-state level in cells in the absence of infection. Both Us9-GFP and Us9-TfR-GFP were predominantly located to a perinuclear region in the cytoplasm of PK15 cells (Figure 6A). Confocal images of cells infected with Becker and PRV 322, fixed and stained with Us9 antiserum, also showed co-localization of Us9 and Us9-TfR to a perinuclear compartment, though the Us9-TfR signal was slightly more diffuse compared to the Us9 signal (Figure 6B). Overall, the trafficking of Us9-TfR inside transfected and infected cells was very similar to wild-type Us9.
Next we tested whether the replacement of the wild-type Us9 TMD with that of transferrin receptor affected the affinity of Us9 for DRMs. We infected undifferentiated and differentiated PC12 cells for 12 hours, solubilized cells with 1% TX-100, and performed flotation analysis as done previously. Instead of Us9 being heavily enriched in the raft fraction as observed in Becker infected cells (see Figures 2 and 3), Us9-TfR was predominantly found in the soluble fraction, especially in differentiated PC12 cells (Figure 7A and 7B). To assess whether this impacted Us9 function in primary neurons, we took advantage of a trichamber neuronal culturing system [17],[37]. Dissociated SCG neurons are plated in the soma (S) chamber and allowed to mature for two weeks (Figure 7C). During this period, axons are directed between a series of grooves across the methocellulose (M) chamber to the neurite (N) chamber. A monolayer of indicator PK15 cells are then plated on top of the neurites in the N chamber. Cell bodies in the S chamber are infected, virus particles sort into axons in a Us9-dependent manner, and subsequently infect the PK15 cells that amplify the infection. The initial infection is confined to the S chamber via silicone vacuum grease and a methocellulose barrier. Therefore, infection spreads to the N chamber solely through axons that emanate from neuronal cell bodies and extend to PK15 cells [17]. We compared the anterograde transport and spread capabilities of PRV Becker (wild-type), PRV 160 (Us9-null), PRV 322 (Us9-TfR), and a co-infection of Becker and PRV 322 (Figure 7C, lower panel). Though all of the infections produced a comparable number of infectious virus in the S chamber, spread to second order PK15 cells in the N chamber was dramatically different. PRV Becker spread efficiently from neurons to PK15 cells, producing a median titer of 1.2×107 PFU in the N chamber after 24 hours post-infection (Figure 7C). By contrast, the Us9-null mutant (PRV 160) did not spread to PK15 cells and no detectable infectious virus was produced in most dishes. However, in one dish, we detected a low number of infectious particles (1.5×103). We interpret a low yield of amplified virus as a single neuron-to-cell spread event (the burst size of an infected PK15 cell is roughly 1000 PFU). Nevertheless, the neuron-to-cell spread capability of PRV 160 is extremely low compared to wild-type PRV Becker. PRV 322 (Us9-TfR) was completely defective in anterograde spread and was indistinguishable from the Us9-null mutant (no infectious virus detected in the N-compartment). This phenotype strongly correlated with the inability of Us9-TfR to target to lipid rafts/DRMs. When neurons were co-infected with both Becker and PRV 322, titers were virtually identical to those seen with Becker alone, indicating that PRV 322 does not have a trans-dominant effect on anterograde spread of infection.
To assess whether the anterograde, neuron-to-cell spread defect for PRV 322 was at the level of axonal sorting of viral structural proteins (as previously shown for other Us9 mutants [15],[16]), we imaged infected neurons in the trichamber system using a PRV-specific antibody (made against acetone-fixed virus particles) that recognizes both virus glycoproteins and virus capsid proteins [21],[22]. PRV antigen was readily detected in the cell bodies of neurons in the S compartment infected with Becker, PRV 160, and PRV 322 (Figure 8, first column). Viral glycoprotein and capsid proteins were also abundant in the axons of Becker infected neurons within the N compartment (Figure 8, second column). By contrast, no viral structural proteins were observed by immunofluorescence in the axons of a Us9-null mutant (PRV 160) or the Us9-TfR strain (PRV 322), though an extensive network of axons was observed within the field of views by transmitted brightfield illumination (Figure 8, transmitted). These data suggest that the neuron-to-cell spread defect observed for PRV 322 (Figure 7) is the result of its inability to sort structural proteins into the axon of infected neurons.
Overall, these data are consistent with work done on the raft association of the influenza virus neuraminidase protein: substitution of a TMD domain that has a high affinity for lipid rafts, for one with a low affinity, dramatically alters DRM targeting and subsequent protein function.
We have demonstrated that Us9 is enriched in detergent-resistant membranes of non-polarized and polarized PC12 cells. This enrichment is cholesterol dependent, and is essential for Us9-mediated anterograde spread of infection in primary SCG neurons. Us9 is responsible for the axonal sorting of viral capsids [15], as well as the viral glycoproteins gB, gC, and gE [16]. These viral membrane proteins are associated with lipid rafts on the surface of PRV infected swine kidney cells and monocytes, and monospecific antibody-induced patching of one of these proteins led to the copatching of the others [18]. These patches were enriched in the raft marker GM1, but not transferrin receptor [18]. These findings are consistent with the raft partitioning of these viral glycoproteins in polarized PC12 cells (Figure 3). In addition, Us9-dependent targeting of viral membrane proteins requires maturation of cultured SCG neurons (Tomishima and Enquist, unpublished observations). Lipid raft formation during neuronal polarization is likely a key step in this sorting process [4]. Taken together, these data support a role for lipid rafts in the axonal sorting of alpha herpesvirus proteins and structures in the mammalian nervous system.
PRV gE, which is in a heterodimeric complex with the viral glycoprotein gI [22], is also necessary for the efficient axonal sorting of PRV structural components [54]. We propose a model in which Us9, gE/gI, and lipid rafts direct the sorting of vesicles into the axon of infected neurons (Figure 9). Us9 and gE/gI likely associate with lipid rafts in the trans-Golgi network (TGN), the putative site of viral assembly [20],[67],[68]. The presence of Us9 and gE/gI in lipid rafts (those decorating the surface of cellular vesicles) would recruit axonal sorting machinery to a small number of viral assembly complexes in the TGN, i.e. vesicles with viral membrane proteins only, those containing mature virus particles, or L-particles (illustrated as “three vesicle populations” in Figure 9). A limited number of vesicles containing virion components would then be targeted to the axon. Though a Us9-gE/gI complex in a lipid raft is required for efficient axonal transport, Us9 is clearly the more critical component, and the presence of gE/gI seems to enhance this process [17].
It is noteworthy that Us9/gE/gI are not required for infectious particle formation in the cell body of SCG neurons [16],[54], nor for retrograde transport in the mammalian nervous system [11],[69]. Furthermore, we have not observed the accumulation of viral capsids or membrane proteins in the cell body of a Us9-null, gE/gI-null, or Bartha strain lacking all three genes (unpublished observations). Our findings are consistent with a model in which the virus assembly and axonal sorting compartment within the TGN are identical (i.e. both processes use the same material for assembly). A small number of assembly complexes would bind a sorting adaptor protein(s) and go to axons; the majority of assembly complexes would egress the cell body locally, perhaps at sites where axons contact the infected cell body (Figure 9).
What determines whether a vesicle laden with viral structural proteins is directed to the axon as opposed to being released from the cell body? We propose that phosphorylation of Us9, subsequent to its recruitment into lipid rafts, may be the pivotal step. The PRV Us9 acidic domain region is heavily phosphorylated during infection [26], and is essential for anterograde, transneuronal spread in vivo [26]. Phosphorylation occurs predominantly on two serine residues within the 10-amino-acid acidic domain, and mutating these serines to alanines dramatically decreases anterograde spread in the rat visual system [26]. This phosphorylation event would occur after Us9 enters the raft, as the acidic cluster domain is not required for raft association of Us9 (Figure 4B). We are currently investigating whether 1) the association of Us9 with lipid rafts coincides with its phosphorylation inside infected cells and 2) the phosphorylation state of the Us9-TfR chimera is reduced since it is no longer enriched in lipid rafts.
Perhaps not surprisingly, the Us9 acidic domain region is highly conserved among Us9 homologs of other neurotropic herpesviruses, including the human pathogens herpes simplex virus (HSV) and varicella zoster virus (VZV), as well as the animal pathogens equine herpesvirus (EHV) and bovine herpesvirus (BHV), suggesting an important role for this domain in the anterograde spread of virus in the mammalian nervous system [26], [70]–[72]. We predict that phosphorylation of the Us9 acidic domain (within the context of a lipid raft) is necessary for binding an axonal sorting adaptor, which would then mediate anterograde transport inside the axon [16],[73].
Our model addresses how viral glycoproteins/vesicles are sorted into the axon of infected neurons, but does not suggest a mechanism for their function in cell-to-cell spread of infection in cultured epithelial cells [74]–[76]. It may be that these two processes are fundamentally different since 1) gE mutants with a small-plaque phenotype on MDBK cells have wild-type anterograde spread kinetics in the rat visual system [77], 2) deletion of PRV Us9 has no effect on cell-to-cell spread of infection in epithelial cells, but a dramatic impact on anterograde sorting [11], and 3) gB mutants with a small-plaque phenotype on ST cells [78] have wild-type anterograde neuron-to-cell spread kinetics in our trichamber system (Curanovic and Enquist, unpublished findings).
It was intriguing to discover that PRV gH was not associated with detergent-resistant membranes in non-polarized and polarized PC12 cells, and was completely solubilized with 1% TX-100 (as was the non-raft marker transferrin receptor). The virus fusion machinery is composed of gB trimers, as well as gH/gL heterodimeric complexes (reviewed in [79]). PRV gH is essential for entry into uninfected cells, cell-to-cell spread of infection in tissue culture [80], and transneuronal spread of infection in mice [81]. Does gH enter the axon in a Us9 or gE-dependent manner despite its apparent exclusion from DRMs? We are currently investigating this question. Cross-linking experiments performed on purified HSV virions found that hetero-oligomers of gB, gC, and gD were closely associated with one another in the virion envelope (within 11.4Å). The gH and gL proteins could also be cross-linked within the envelope as one might predict. Interestingly, gL was never cross-linked to gB, leading the authors to suggest that organization of these proteins in the membrane “precludes associations of gH/gL with gB” [82]. One explanation for this finding is that gB is present in a lipid raft microdomain, whereas gH is not (a small proportion of HSV gH has been shown to be in the DRM fraction of infected COS cells [83]). It is also feasible that PRV gH may indeed be raft-associated, but solubilization with cold detergent is too stringent. Triton X-100 and CHAPS are reported to be the most reliable detergents for analyzing raft association [84]. However, some membrane proteins solubilized by Triton X-100 do associate with lipid rafts by antibody copatching [18],[51],[85],[86]. At this time it is unclear if gH has a weak affinity for rafts, or is indeed a “true” non-raft protein as is transferrin receptor.
Our findings also highlight the importance of the Us9 TMD domain in raft targeting. Several studies have addressed the importance of the transmembrane segment in partitioning viral and cellular membrane proteins into lipid rafts: influenza virus hemagglutinin [60],[64] and neuraminidase [59],[63], the LMP-1 oncoprotein of Epstein-Barr virus (EBV) [87], and the human immunoreceptor FcγRIIA [88]. It is clear from these studies that amino acids within the TMD (even single amino acids) have dramatic effects on raft partitioning, sorting, or signaling events. A comprehensive analysis of the Us9 TMD may reveal residues important for protein-lipid/protein-protein interactions that are key in promoting axonal sorting of mature virus particles.
Several alpha herpesvirus proteins have been shown to associate with DRMs during virus replication. In addition to PRV membrane proteins ([18]; this study) the virion host shutoff (vhs) protein of HSV-1 was shown to be enriched in organellar membrane fractions which contain virus assembly intermediates [83]. HSV gB is proposed to mobilize lipid rafts during entry, perhaps to mediate cell signaling [81]. The UL11 protein of HSV-2, a myristoyl and palmitoyl tegument protein, associated with DRMs of infected Caco-2 cells [89]. The UL56 protein of HSV-2, a tail-anchored type II membrane similar to PRV Us9, was present in detergent-insoluble lipid rafts; it is predicted to be involved in vesicular trafficking in HSV-2 infected cells [90]. These findings demonstrate that lipid rafts play an important role in the replication cycle of alpha herpesviruses, and underscore the importance of lipid rafts in virus biology [91]–[94].
Two competing models have been presented for anterograde, axonal transport of alpha herpesviruses in neurons (reviewed in [95]). Viral capsids are either transported down the axon independently from viral membrane proteins (and assemble prior to egress), or they are sorted and transported together as a mature virus particle within a vesicle. Recent studies have shown that efficient capsid transport is dependent on at least two viral membrane proteins, Us9 and gE. Deleting either of these genes from HSV or PRV results in a marked decrease of viral capsids from entering axons of infected neurons [12],[15],[54],[96]. It is difficult to conceive how viral membrane proteins could impact capsid sorting unless the two were tightly coupled during axonal entry and transport (i.e. if capsids entered axons separate from viral membrane proteins, deletion of gE and Us9 would have no effect on capsid sorting). These recent findings are consistent with a model where viral capsids and membrane proteins traffic together in axons as mature virus particles (within a vesicle) (Figure 9).
In conclusion, we have shown that PRV Us9 is highly enriched in DRMs of non-polarized and polarized PC12 cells, and this enrichment is critical to axonal targeting and subsequently in neuron-to-cell spread. This is the first report to implicate lipid rafts in the axonal sorting of alpha herpesvirus structural proteins in mammalian neurons. Our future plans include isolating lipid rafts from polarized PC12 cells infected with wild-type PRV, and identifying the cellular and viral proteins present within these lipid microdomains. |
10.1371/journal.pbio.2002690 | Common genes associated with antidepressant response in mouse and man identify key role of glucocorticoid receptor sensitivity | Response to antidepressant treatment in major depressive disorder (MDD) cannot be predicted currently, leading to uncertainty in medication selection, increasing costs, and prolonged suffering for many patients. Despite tremendous efforts in identifying response-associated genes in large genome-wide association studies, the results have been fairly modest, underlining the need to establish conceptually novel strategies. For the identification of transcriptome signatures that can distinguish between treatment responders and nonresponders, we herein submit a novel animal experimental approach focusing on extreme phenotypes. We utilized the large variance in response to antidepressant treatment occurring in DBA/2J mice, enabling sample stratification into subpopulations of good and poor treatment responders to delineate response-associated signature transcript profiles in peripheral blood samples. As a proof of concept, we translated our murine data to the transcriptome data of a clinically relevant human cohort. A cluster of 259 differentially regulated genes was identified when peripheral transcriptome profiles of good and poor treatment responders were compared in the murine model. Differences in expression profiles from baseline to week 12 of the human orthologues selected on the basis of the murine transcript signature allowed prediction of response status with an accuracy of 76% in the patient population. Finally, we show that glucocorticoid receptor (GR)-regulated genes are significantly enriched in this cluster of antidepressant-response genes. Our findings point to the involvement of GR sensitivity as a potential key mechanism shaping response to antidepressant treatment and support the hypothesis that antidepressants could stimulate resilience-promoting molecular mechanisms. Our data highlight the suitability of an appropriate animal experimental approach for the discovery of treatment response-associated pathways across species.
| Major depression is the second leading cause of disability worldwide. However, only one-third of patients with depression benefit from the first antidepressant compound they are prescribed. It is a fundamental problem that the outcomes of individual antidepressant treatments are still highly unpredictable. In clinical studies, discovery of biomarkers for antidepressant response is hampered by confounding factors such as the heterogeneity of the disease phenotype and additional environmental factors, e.g., previous life events and different schedules of psychopharmacological treatment, which reduce the power to detect true response biomarkers. To overcome some of these limitations, we have established a conceptually novel approach that allows the selection of extreme phenotypes in an antidepressant-responsive mouse strain. In the first step, we identify signatures in the transcriptome of peripheral blood associated with responses following stratification into good and poor treatment responders. As proof of concept, we translate the murine data to a population of depressed patients. We show that differences in expression profiles from baseline to week 12 of the human orthologues predict response status in patients. We finally provide evidence that sensitivity of the glucocorticoid receptor could be a potential key mechanism shaping response to antidepressant treatment.
| A “one size fits all” approach is not effective or efficient in the treatment of major depressive disorder (MDD). Although it would be ideal to tailor available treatments to individual patients [1], patient-level antidepressant treatment outcomes are still highly unpredictable [2]. Identification of biomarkers predictive of individual treatment response or molecular biosignatures associated with response would dramatically improve the quality of care for MDD [3]. These biomarkers could also be expected to significantly reduce both treatment and loss-of-productivity costs. The latter become increasingly important because MDD has been shown to be the second leading cause of disability worldwide [4]. Finally, biomarkers could allow patient stratification and enable the selection of pathophysiologically distinct patient subgroups to allow optimized treatment choices based on biology. Such biomarkers could also inform the development of new interventions specifically targeting disease mechanisms in these subgroups.
Conceivably, useful biomarkers for treatment response in depression could be developed through blood-based biomarkers, including genetic approaches, although psychophysiological and neuroimaging approaches are also promising [5]. However, despite considerable efforts, including large-scale hypothesis-free, genome-wide approaches during the past years [6, 7], no biological or genetic predictors of sufficient clinical utility have been identified for routine clinical use. Thus, the most effective treatment for each patient is currently identified through a trial and error process [2].
Among the potential barriers to the development of clinically useful biomarkers in depression, the following 3 have been identified as being most important. First, current symptom-based diagnoses likely group pathophysiologically distinct patients [8], leading to considerable heterogeneity among patients diagnosed with MDD [9, 10]. Second, there are a number of confounding environmental factors such as childhood maltreatment, previous life events, disease episodes, and different psychopharmacological treatment schedules that often remain unidentified and potentially reduce the power to detect true response biomarkers. Third, genetic background, age, and sex are all factors that significantly impact transcription profiles and other laboratory measurements, as well as treatment outcome [11].
In addition to the aforementioned problems, major psychiatric disorders, including MDD, are primarily viewed as brain disorders, so the question of whether peripheral measures can be informative for treatment response to centrally acting compounds such as antidepressants continues to be matter of debate [12]. During recent years, evidence has emerged that disease- and treatment-related changes may be reflected outside the central nervous system [13, 14], revealing a potential role for appropriate animal models to support biomarker discovery in MDD. To the best of our knowledge, neither an appropriate animal experimental approach nor a translational approach systematically addressing the potential of biosignatures predicting or tracking antidepressant treatment response has been published.
To overcome some of the limitations of past approaches, we here present a conceptually novel approach that allows the selection of extreme phenotypes in an antidepressant-responsive mouse strain (DBA/2J [15]) and uses these extreme groups to identify peripheral blood biomarkers associated with behavioral treatment response, which are then tested in a human patient cohort. This strategy exploits the advantages of a murine approach for the purpose of biomarker discovery, i.e., (1) to investigate a highly homogeneous group of animals in which differences in genetic background, age, and sex can be excluded, (2) to perform biomarker discovery under conditions in which interindividual confounding environmental influences, including drug plasma and brain levels, are reduced to a minimum and controlled for, and (3) to allow correlations of peripheral biomarkers with behavior but also with peripheral and central drug concentrations, and to test the overlap of blood and brain expression profiles. We hypothesize that these standardized conditions will facilitate the identification of valid peripheral biomarkers for antidepressant treatment response and allow translation to humans.
Experiments were carried out with male DBA/2J mice (n = 140) from Charles River, France. On the day of arrival, the animals were 6–8 weeks old and from that day on were singly housed in standard cages under a 12L:12D cycle (lights on at 0800 h) and constant temperature (23 ± 2°C) conditions. Food and water were provided ad libitum. Pharmacological treatment of all animals started at an age of 9–11 weeks. Behavioral testing was performed at an age of 11–13 weeks. The experiments were carried out in the animal facility of the Max Planck Institute of Psychiatry in Munich, Germany, and approved by the committee for the Care and Use of Laboratory Animals of the Government of Upper Bavaria, Germany. All experiments were carried out in accordance with the European Communities Council Directive 86/609/EEC.
The sequential steps and experimental procedures are summarized in Fig 2, indicating the number of animals for each experimental group. A large number of animals were treated twice a day with either paroxetine (n = 90), a commonly used selective serotonin reuptake inhibitor (SSRI) antidepressant or a vehicle (n = 50). On treatment day 15, the animals received their last drug administration at 6 AM and were subjected to a FST 4 hours later. Directly after the FST, the animals were anesthetized with isoflurane and decapitated.
Animals were anesthetized with isoflurane and killed immediately following the FST. Trunk blood was collected individually in 1.5-mL tubes. Brains were rapidly dissected and frozen at −80°C.
Due to the complex character of the study, limitations in available specimens, stringent quality control (QC), and exclusion of outlier data, we could not always achieve fully identical sample and group compositions throughout all data analysis levels. This also explains the sporadic appearance of nonconcordant group sizes, which we consider a minor but unavoidable drawback.
Brain and plasma paroxetine concentrations were measured after extraction by high liquid chromatography and quantifications. Paroxetine plasma concentrations were considered as a covariate in the analysis of the microarray data. For details of the respective protocols, see [18].
For determination of brain tissue concentrations of paroxetine, tissue from the cerebellum was dissected and rapidly frozen on dry ice. The remaining trunk blood of each animal was collected in labeled 1.5-mL EDTA-coated microcentrifuge tubes (Kabe Labortechnik, Nümbrecht, Germany). All blood samples were kept on ice until centrifugation at 8,000 rpm at 4°C for 15 min. After centrifugation, the blood plasma was transferred to new, labeled 1.5-mL microcentrifuge tubes. All plasma samples were stored frozen at −20°C until the determination of corticosterone by radioimmunoassay (MP Biomedicals, Santa Ana, CA; sensitivity, 6.25 ng/mL).
The data presented are shown as means + standard error of the mean, analyzed by the commercially available software SPSS 16.0. For comparing 2 independent groups, data were analyzed with 2-tailed, independent samples Student t test in case of normal distribution of the data; otherwise, nonparametric comparisons were applied (Mann–Whitney U test). For variables with more than 2 groups, 1-way ANOVA was performed followed by Bonferroni post hoc testing. Correlations were analyzed with a 2-tailed, bivariate Pearson’s correlation analysis. As nominal level of significance, p < 0.05 was accepted. Values outside the 95% confidence interval (CI) were defined as statistical outliers and excluded from the analyses.
Part of the blood was processed according to the PAXgene blood miRNA Kit manufacturer’s instructions. Briefly, 350 μL of freshly collected trunk blood was immediately transferred into 1.5-mL tubes filled with 966 μL PAXgene solution (RNA stabilizer reagent), gently inverted 10 times, incubated at room temperature (RT) for 2–24 hours, and then stored at −20°C before ribonucleic acid (RNA) isolation. Volume ratio of RNA stabilizer reagent to blood samples was kept at 2.76, according to the manufacturer’s protocol.
We assessed whether the observed gene expression profiles of good treatment responders and poor treatment responders were related to changes in blood cell proportions in the mice using CIBERSORT [19]. The input reference matrix of expression signature profiles of mouse tissue was obtained using ImmuCC [20]. These statistical tools infer proportions of 25 types of immune blood cell types.
To assess the relevance of the gene expression transcripts for antidepressant response differences in humans, we tested their predictive ability to classify response status in a human sample. The sample (n = 86) consisted of a subset of MDD patients treated with antidepressant drug treatment over 12 weeks from 2 samples recruited at Emory University School of Medicine (N = 74 from [21] and N = 12 from [22]). In both studies, patients followed a similar protocol and were randomized to either antidepressant drug treatment or cognitive behavior therapy (CBT), with the difference that patients were randomized to CBT, duloxetine, or escitalopram in PReDiCT [21] and to CBT or escitalopram in [22]. Only the subset of patients in the antidepressant treatment group with sufficient RNA quality at both time points was included in this study. Please see S2 Table for a brief synopsis of demographic and clinical parameters on the patients from clinical studies. Depression severity was assessed at baseline and week 12 using the Hamilton Depression Rating Scale (17 items, HDRS-17). In both samples, blood was drawn at baseline and after 12 weeks of treatment into Tempus RNA tubes (Applied Biosystems).
RNA was isolated from peripheral blood in a 96-well format using the magnetic bead-based technology MagMAX for Stabilized Blood Tubes RNA Isolation Kit, compatible with Tempu Blood RNA Tubes (Ambion/Life Technologies, Carlsbad, CA; cat# 4451893). RNA was quantified using the Nanophotometer, and quality checks were performed on the Agilent Bioanalyzer (Agilent Technologies, Santa Clara, CA). Only samples with RIN ≥ 6 with clear 18S and 28S peaks on the Bioanalyzer were used for amplification; the average RIN was 6.3 (SD of 0.668). RNA was further processed for generation of biotin-labeled amplified RNA using the Amplification Kit (Ambion/Life Technologies, Carlsbad, CA; cat# 4393543). cRNA was hybridized to Illumina Sentrix Arrays HT-12 v4.0 arrays using the Illumina TotalPrep-96 RNA (Life technologies, Carlsbad, CA) and incubated overnight for 16 hours at 55°C. Arrays were washed, stained with Cy3 labeled streptavidin, dried, and scanned on the Illumina BeadScan confocal laser scanner (Illumina, San Diego, CA). QC was performed using the bead-array package in R for 86 samples and 47,282 probes. Probes with p-detection values of <0.01 in at least 10% of the samples in the whole data set were removed. Remaining probes were normalized and transformed using the vsn package in R. Not all samples were hybridized on the same batch, and thus we corrected for chip number using COMBAT. A total of 17,725 transcripts and 86 samples remained after QC.
For the full drug-treated sample, 63 patients were classified as responders and 23 as nonresponders, according to percent changes in HDRS-17 scores from baseline to week 12 (≥50% or <50% change, respectively).
Mouse gene expression transcripts (n = 259) resulting from the microarray analysis and described in S1 Table were mapped to their human orthologue genes present in the Illumina HT-12 arrays (n = 241). Because some genes are represented by more than one probe, 288 probes were included in final analyses. Prediction models were built as soft margin support vector machines for classification using the e1071 packages in R with the parametrization gamma = 0.001; cost = 10. Further analyses included only mouse transcripts at FDR of 5% (n = 85). These were also mapped to their human orthologue genes (n = 77); 66 genes passed QC in the human study, which were represented by 92 probes. The sample was equally divided into training and test data sets for each of the analyses (probes at q < 0.1 and q < 0.05). Gene expression repeated measures from the patients at baseline and week 12 were available; we computed the absolute difference between the expression levels of the transcripts between those time points and tested whether these differences were able to predict response to antidepressant treatment in the test data set. We permuted the response-status labels 10,000 times in the training data set and predicted the response status in our test data. In addition, we compared the obtained prediction accuracy of our selected classification features against 1,000 classification models derived from randomly sampled features. Random feature sets also consisted of absolute difference in expression between baseline and week 12 of treatment and were size matched to the selected feature set. Those data were the input for soft margin support vector machine training and testing as indicated above.
We assessed whether the observed gene expression changes in responders versus nonresponders were related to changes in cell proportions in the human samples using the Cell-type Computational Differential Estimation CellCODE R package [23]. Separate components for neutrophils, T cells, stimulated T cells, NK cells, dendrite cells, stimulated dendrite cells, monocytes, B cells, and plasma cells were extracted using markers from the IRIS reference data set provided by CellCODE.
Two available tools have been used for pathway analyses: DAVID (https://david.ncifcrf.gov/) and Pathway-Express [24].
Both tools were used with a list of gene symbols previously shown to be significantly regulated (q-value < 0.1) with differential paroxetine response and interrogated with respect to a custom background that contained all microarray probes that have been used for computing inferential statistics. The background contained probes that passed our detection and variance filters.
To determine the function overlap of differential paroxetine response with dex-regulated genes, we used data from a microarray experiment in male C57BL/6N mice at an age of 12 weeks (mean body weight 26.8 ± 0.1 g), in which animals were treated with 0.1 mg/kg dexamethasone i.p. or vehicle (N = 10 and 10) between 0900 and 1100 and sacrificed 4 hours later [25]. Trunk blood was collected into microcentrifuge tubes containing PAXGene RNA stabilizer solution and frozen at −20°C. RNA was then extracted using the PAXgene blood miRNA kit (PreAnalytiX), amplified using the Illumina Total Prep 96-Amplification kit (Life Technology), and then hybridized on Illumina MouseRef-8 v2.0 BeadChips.
Analyses were performed using custom scripts in R. First, a common content for both microarray data sets was generated based on Illumina “Probe Ids.” Within that common content, differentially expressed microarray probes were identified for both contrasts using an FDR threshold of q < 0.1. For the differential paroxetine response, 179 probes passed that threshold. Then, the number of array probes overlapping with dex regulation by chance was determined using 100,000 random sampled gene sets of size N = 179. For each trial, the overlap to the fixed dex-regulated gene list (N = 1,882) was determined and all the results were finally compared to the overlap of paroxetine response genes with dex-regulated genes; this was done by counting the number of sampled sets that showed higher overlap (>134) than the differential gene list.
In addition, a 2 × 2 contingency table was computed for dex regulation and paroxetine response and these numbers were further used to perform a hypergeometric test.
Calculation of statistical significance for a possible directionality of gene regulation was performed using a binomial test.
In order to detect the minimum effective dosage of paroxetine for the DBA/2J strain, 2 paroxetine concentrations (1 mg/kg body weight or 5 mg/kg body weight, twice daily) were tested in a pilot study. The lower paroxetine concentration (n = 29) failed to produce a significant behavioral treatment effect in the FST. The only parameter that was altered with the 1 mg/kg dosage was body weight (T39 = −2.490, p < 0.05). Behavioral data, neuroendocrine measurements, and body weight are shown in S1 Fig.
A dosage of 5 mg/kg evoked a significant antidepressant-like response in the FST (Fig 3). The following data were all collected from animals treated with 5 mg/kg paroxetine, which we considered to be the minimum effective dosage for the DBA/2J strain.
There was no significant difference in plasma paroxetine concentrations between the good and poor treatment responder (p = 0.19). For paroxetine brain concentrations, a significant difference between good and poor responders could be detected (p < 0.05) (S3 Fig). Paroxetine brain and plasma concentrations were closely correlated (r = 0.94; p < 0.0001) (S3 Fig). Despite the lack of statistical association, we included plasma paroxetine concentrations as a covariate in further analyses on the transcriptome profiles in peripheral blood samples. Brain paroxetine concentrations were used as covariates in analyses of PFC samples.
To identify signature gene expression profiles characteristic of the animals’ responder status, gene expression data sets of vehicle-treated animals, good responders, and poor responders were created by whole-genome gene expression microarray analysis on blood samples and analyzed (n[vehicle] = 12, n[good] = 12, n[poor] = 12). We evaluated both treatment effect and response status with respect to antidepressant treatment and with respect to paroxetine plasma concentrations. We also investigated whether paroxetine brain or plasma concentrations might have an effect on gene expression levels. Linear and quadratic regression analyses did not reveal any microarray probe that showed significant correlations with the related plasma paroxetine levels when controlling for multiple testing. No significant influence of paroxetine concentrations on gene expression profiles was observed. Nevertheless, identified technical batch effects in the data and measured paroxetine drug concentrations in blood were used as covariates in an ANOVA-based statistical model.
Although no robust gene regulation was apparent when the treatment group (independent of response) was compared to the control group, there was a pronounced effect within the treatment group. We were able to detect a set of 259 transcripts that showed a significant difference in expression due to antidepressant response status at a false discovery controlled significance level of 10% (q < 0.1) (Fig 4; S1 Table), of which 85 had q < 0.05 (S1 Table).
We then aimed to see whether the observed gene regulation patterns in peripheral blood might overlap with effects observed in the PFC from the same animals. To test this, we first performed a cluster analysis on the difference in expression between the responder groups in the set of differentially regulated genes in blood. We then compared the results for these transcripts to the difference in expression between these 2 groups measured in PFC brain tissue in the same animals. The results are summarized in a heat map in Fig 4 and indicate that, within the selected gene set, there is no major common gene regulation pattern associated with response status between both tissues.
No significant differences in immune cell subtypes between the different response groups were detected using CIBERSORT [19] and ImmuCC [20] (see S3 Table).
No significant change in immune cell subtypes using CellCODE [23] was associated with the response groups in the human sample (see S4 Table). Therefore, none of the estimated cell proportions were included in further analyses.
In the next step, we determined whether this transcriptional profile identified in the mouse model would also be relevant in the human data set. Therefore, we tested whether changes in the mRNA expression of the human orthologues of transcripts at FDR of 10% and at FDR of 5%, separately, are associated with response to antidepressant treatment. Differences in expression profiles from baseline to week 12 when using human orthologues of transcripts at FDR of 10% allowed prediction of response status (at least 50% improvement in HDRS-17 from baseline to week 12 for responders) with an accuracy of 76%, using all patients treated with antidepressant. The prediction persisted after we permuted the response-status labels 10,000 times (pperm = 0.0328). When a more stringent FDR of 5% cutoff was applied to the mouse transcripts, the corresponding human orthologues predicted response status with an accuracy of 81% in the human sample. The prediction persisted after we permuted the response-status labels 10,000 times (pperm = 0.0018).
After showing that expression levels of the antidepressant response genes identified from mice are also informative for classification in a human sample, we further analysed the quality of the mouse-based feature selection in the human data set. For this, we compared the classification accuracy of our identified antidepressant-response features to classification accuracy given by randomly chosen and size-matched sets of gene expression probes in the human sample (Fig 5). In analogy to the previous classification approach, we used differences in gene expression from baseline to week 12 in 1,000 random sets of gene probes. Only 25 random gene probe sets showed higher or equal prediction accuracy than our feature panel selected from the animal model. This suggests that the information derived from the mouse experiments allowed the selection of transcripts for which the classification accuracy is better than for random gene expression background (pperm = 0.026).
For functional annotations of the microarray results, we performed pathway analyses and included an overrepresentation analysis with DAVID, and we conducted a second analysis using Pathway-Express. The latter accounts for pathway topology and biological effect size. In both approaches, no significant results passing our threshold criteria were found. The top overrepresented categories in DAVID were entities associated with general gene transcription and did not reach significance levels. Although Pathway-Express showed formally significant results for a few specific Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways, we excluded them because less than 2% of the pathway genes were regulated.
We next integrated our results with another microarray data set that we had previously generated. Those data originated from mouse blood samples taken from animals that had been treated with the glucocorticoid receptor (GR) agonist dexamethasone (dex [25]). To test whether GR activation responsive genes are overrepresented in our antidepressant response gene set, we used a permutation approach and computed the overlap of dex-regulated genes with the paroxetine response genes and compared it to matched random gene sets sampled from the paroxetine array results (Fig 6). Based on 2,852 array probes that constituted a common content for both independent data sets, 179 array probes of the 259 response associated probes could be used for this analysis. The overlap between the probes significantly regulated between the responder group and the ones regulated following dex administration was 134 out of 179. Within 100,000 trials of drawing random gene sets of 179 probes, there were only 70 instances in which a higher overlap occurred. This reflects a permutation-based FDR of 7e-4 for enrichment of dex-regulated array probes in paroxetine response probes. A hypergeometric test yielded a p-value of 5.6e-4, further supporting an enrichment of GR-responsive transcripts among response-associated genes.
Standard enrichment analysis does not take into account the direction of gene regulation, and we were interested to see whether paroxetine response and dex regulation showed a directional overlap. Of the 134 array probes that are significantly regulated by dex and are, at the same time, between the paroxetine response groups, only 38 had a mismatch in the direction of the putative regulation. The majority of the regulated genes (N = 96) are regulated in the same direction in both conditions, and based on a binomial distribution, such a result could not be observed if both outcomes (same and opposite regulation) had the same probability (p = 7.2e-07). Thus, we can conclude that there is a common direction of gene regulation for dex treatment and paroxetine response.
The goal of the present study was to gain insight into the biology of variations in response to antidepressant treatment and to describe molecular signatures associated with response, ultimately aiming at the identification of predictors of treatment outcome. Based on a conceptually novel translational approach, starting with stratification into extreme phenotypes in the mouse, we were able to identify common—i.e., conserved across species—informative transcript sets associated with antidepressant treatment outcome. Intriguingly, we finally show that GR-regulated genes are significantly enriched in this cluster of antidepressant-response genes, pointing to the involvement of GR sensitivity as a potential key mechanism in shaping transcriptional changes and clinical response to antidepressant treatment.
There are 2 obvious gaps of knowledge in depression treatment, namely (1) the lack of biosignatures predicting antidepressant response and (2) the lack of knowledge of the molecular mechanisms mediating the response to antidepressant pharmacotherapy. The latter is of particular importance for the eagerly awaited discovery of conceptually novel antidepressant treatment strategies, which can only be rationally realized with a deeper understanding of the molecular mechanisms underlying clinical response [26].
In recent years, the unbiased, i.e., genome-wide, screening to identify genetic factors that could assist in the prediction of an individual’s drug response has been a major focus in depression research. Despite tremendous efforts, however, the results are fairly modest in identifying predictive genes in large genome-wide association studies [27–29] and even in a meta-analysis [6]. Instead, Tansey et al. [30] recently presented data implicating a highly polygenic architecture involving many common variants scattered across the genome, none of which have very large effects but cumulatively contribute to a substantial proportion of variation in antidepressant response. So far, only a few small studies provided first evidence that biochemical information (e.g., metabolomics) could add to the panel of markers predicting response to a particular antidepressant in patients [31], suggesting that alternative strategies need to be explored.
However, studies to investigate the neurobiology of antidepressant treatment response have been hampered by the fact that no appropriate animal model addressing this issue had yet been described. Therefore, we embarked upon the development of an animal experimental approach modeling the heterogeneity in response to antidepressant treatment as closely as possible. In contrast to studies in patients, this model approach both enables an in-depth analysis of the neurobiological mechanisms shaping individual antidepressant response in the central nervous system and searches for peripheral biosignatures associated with treatment response. There are different approaches to model depression-like phenotypes (i.e., symptoms of depression) in the mouse. While induction of depression-like symptoms following exposure to different types of stress, e.g., chronic social defeat or chronic mild stress is one possible approach, the use of mouse strains with high innate anxiety- and depression-like behavior is also commonly accepted. The selection of the DBA/2J mouse strain, with its well-described high innate anxiety and responsiveness to antidepressant treatment [17], enabled us to perform the pharmacological treatment under basal conditions, i.e., without the need to subject the animals to an additional stress procedure that might have influenced the transcriptome data. A combination of stress exposure and antidepressant treatment within our approach would not allow us to identify the individual contribution of these 2 factors to the phenotype. Nonetheless, a comparison of stress-related and antidepressant response–related molecular events could enable the identification of shared molecular pathways.
Oral treatment with the SSRI paroxetine significantly reduced—as expected—depression-like behavior. Remarkably, in addition to the overall antidepressant-like effect on promoting active coping strategies in the FTS, we detected a high variability in the behavioral outcome. Although the neurobiological mechanisms underlying antidepressant-induced behavioral changes in the FTS still are not fully understood [32], we here used the FST as the laboratory animal equivalent of treatment response because it is the most commonly used test to screen for antidepressant efficacy in rodents [16]. Comparable approaches for stratification and extreme case sampling in animal models have been successfully introduced in the field of stress research [33], and during recent years, they have enabled the identification of a number of key mechanisms shaping individual susceptibility to stress [34, 35]. We considered plasma paroxetine concentration as a covariate on our microarray analyses, but we were not able identify a significant influence on the gene expression profile associated with treatment response.
The selection of a rodent approach for biomarker discovery in psychiatric disorders has the advantage of minimizing potentially confounding variables, which, in clinical depression studies, so far have impeded biomarker discovery [12]. Due to the standardized experimental conditions, factors such as sex, age, and additional environmental factors, including pharmacological pretreatment, the time of day at which the blood sample is taken, physical exercise, food, and many others [36], can be strictly controlled for, thus enabling the detection of true response biomarkers in a hypothesis-free approach. In a second step, these murine biomarkers can then be validated in the human population. Given the complexity of identifying true biomarker candidates in psychiatric disorders, the need to strengthen potential candidates by cross-species approaches [37] and to validate those in independent cohorts is considered crucial [38].
Aiming to enable a translational approach, we focused on the identification of transcriptome signatures in the periphery, because only those are relevant for clinical application. Several studies have investigated the use of human peripheral blood cells as surrogate material for different organs and tissues, including the central nervous system [39–41]. However, inconsistent results have been reported as to the overlap between transcriptome profiles in peripheral blood and brain [14].
To address issues of cross-tissue relevance, we compared peripheral transcriptome signatures with expression profiling data of the PFC of the same good- and poor-responding animals. We did not find any major common response status-associated gene regulation pattern between both tissues. We thus hypothesize that in depression treatment, blood cells might act as sentinels of treatment response but are not generally informative about central regulation processes, at least not in the PFC.
In the next step and as a proof of concept, we sought to evaluate the relevance of the murine transcriptional signature associated with antidepressant treatment response in a human data set. Using a powerful within-participant approach investigating longitudinal transcription changes between baseline and week 12 of antidepressant treatment, we tested whether mRNA expression of the human orthologues of these transcripts changes with antidepressant treatment in peripheral blood in a subset of 2 human studies [21, 22]. Differences in expression profiles from baseline to week 12 of the human orthologues selected on the basis of the murine transcript signature allowed prediction of response status (percent change in HDRS-17 from baseline to week 12) with an accuracy of 76% in the human sample. Using a permutation strategy, we also showed that our set of transcripts was more likely to predict treatment outcome correctly than random sets of transcripts. We thus show the suitability of an appropriate animal experimental approach for the discovery of peripheral treatment response biomarkers. While promising, our findings certainly require validation in independent samples of patients with MDD. One aspect that needs more detailed investigation in future studies is the precise time course and stability of response-associated transcript changes, as we here integrated murine transcript data following 2 weeks of antidepressant treatment with patient data over a 12-week treatment course.
The available evidence makes a compelling case implicating dysregulation of the stress hormone system, the so-called hypothalamus-pituitary-adrenocortical (HPA) system, in the pathogenesis of MDD [42, 43]. Moreover, considerable evidence has accumulated suggesting that normalization of the HPA system might be the final step necessary for stable remission of the disease [44], and it was further hypothesized that antidepressants may act through normalization of the HPA system function [45]. A recent study provided evidence that hormone-independent activation of the GR is involved in the therapeutic action of fluoxetine [46], supporting the neurobiological link between GR signalling and antidepressant action.
We could not detect any difference in corticosterone plasma concentrations between good and poor responders to paroxetine treatment directly after the FTS challenge, although assessment of plasma corticosterone concentrations at 1 time point, i.e., 5 min after the FST, does not exclude potential dynamic changes in HPA system response (i.e., changes in the rise of corticosterone or HPA system feedback following initial activation). Evidence from measurements of HPA system activity in depressed patients, however, supports the notion that in vivo challenges such as the combined dexamethasone/corticotropin releasing hormone challenge test (Dex-CRH test) are superior to single baseline measurements of peripheral glucocorticoid concentrations in discriminating between depressed patients and healthy controls as well as treatment responders versus nonresponders. In addition, recent investigations have shown that dex-stimulated gene expression is a sensitive marker of GR-resistance in MDD [13] and that common genetic variants that modulate the initial transcriptional response to GR activation increase the risk for depression [25]. Therefore, we tested for an enrichment of GR-responsive genes in our antidepressant response gene set, a finding that could point to increased GR sensitivity in good- versus poor-responding animals. We demonstrated that (1) GR-regulated genes are significantly enriched in our cluster of antidepressant-response genes and (2) there is a common direction of gene regulation for dex treatment and paroxetine response. Our data are in line with a large body of previous evidence pointing to the normalization of GR resistance as an important feature of the clinical response to antidepressant treatment [43, 47] and support the intriguing hypothesis that antidepressants could stimulate resilience-promoting molecular mechanisms [48].
Biomarkers or biosignatures, respectively, would not only allow monitoring of antidepressant treatment response in clinical practice but they also could assist in the evaluation of drug actions at an early stage in clinical trials of novel agents that are frequently marred by late attrition [49]. In particular, identifying biomarkers of response will be essential for assessing target engagement of novel mechanisms. We submit that our approach opens up the opportunity to generate a unique database for putative biosignatures predicting response to be assessed and validated in larger patients’ samples.
In conclusion, we expect this translational approach to serve as a template for the discovery of improved and tailored treatment modalities for depression in the future.
|
10.1371/journal.pcbi.0030032 | Repressor Dimerization in the Zebrafish Somitogenesis Clock | The oscillations of the somitogenesis clock are linked to the fundamental process of vertebrate embryo segmentation, yet little is known about their generation. In zebrafish, it has been proposed that Her proteins repress the transcription of their own mRNA. However, in its simplest form, this model is incompatible with the fact that morpholino knockdown of Her proteins can impair expression of their mRNA. Simple self-repression models also do not account for the spatiotemporal pattern of gene expression, with waves of gene expression shrinking as they propagate. Here we study computationally the networks generated by the wealth of dimerization possibilities amongst transcriptional repressors in the zebrafish somitogenesis clock. These networks can reproduce knockdown phenotypes, and strongly suggest the existence of a Her1–Her7 heterodimer, so far untested experimentally. The networks are the first reported to reproduce the spatiotemporal pattern of the zebrafish somitogenesis clock; they shed new light on the role of Her13.2, the only known link between the somitogenesis clock and positional information in the paraxial mesoderm. The networks can also account for perturbations of the clock by manipulation of FGF signaling. Achieving an understanding of the interplay between clock oscillations and positional information is a crucial first step in the investigation of the segmentation mechanism.
| Vertebrate embryos acquire a segmented structure along the anteroposterior axis. Segmentation is critical for patterning of other structures (such as nerves, vertebrae, muscles, and blood vessels) and occurs by the rhythmic separation of balls of cells, called somites, from the anterior end of their precursor tissue, called the presomitic mesoderm. These rhythmic events are associated with oscillatory gene expression in the presomitic mesoderm: waves of gene expression originate at the posterior end and spread anteriorly. When a wave reaches the anterior end, a pair of new somites detaches. The set of genes whose expression oscillates is termed the “somitogenesis clock.” Even though the zebrafish somitogenesis clock has been the subject of intensive study, it is not clear how its oscillations are generated. It has been proposed that the mechanism involves a simple negative feedback loop, with proteins of the Her family periodically repressing their own expression. However, this is incompatible with some experimental results and does not explain how the spatiotemporal pattern of gene expression is generated. Here I propose a model—based on physical interactions between Her proteins—that is compatible with experimental results, and that explains how positional information is used to generate the spatiotemporal pattern of gene expression.
| A somitogenesis clock, linked to the vertebrate segmentation process, has been uncovered in mouse, chick, and zebrafish [1]. Genes involved show oscillatory expression in the presomitic mesoderm (PSM) and are mostly related to the Notch pathway in all three species. In zebrafish PSM, oscillatory genes include her1 and her7, transcriptional repressors whose expression is thought to be enhanced by Notch signaling, and deltaC, which encodes a Notch ligand; expression of each of these three genes is necessary for correct oscillatory expression of all three [2,3]. The Her1 protein represses expression from its own promoter in a cell-culture assay [4], and it has been proposed that self-repression of her genes could drive the somitogenesis clock [5–7]. Intriguingly, however, morpholino blocking of her1 or her7 mRNA translation can lead to downregulation of their transcription [2], while a simple negative feedback loop would predict upregulation of transcription if the repressor protein cannot be translated. What is more, the significance of the requirement for Her13.2, a cofactor for Her self-repression [4], remains unexplored in mathematical models. Here we study computationally the properties of networks where Her proteins must dimerize to act as repressors. Networks are assessed both for compatibility with morpholino knockdown phenotypes and for correct reproduction of the collective creation of an oscillatory pattern in the PSM.
Her13.2 is expressed in a posterior–anterior gradient in zebrafish PSM and heterodimerizes with Her1 to enhance Her1 autorepression [4]. Dimerization is a common feature of the basic helix loop helix family [8], and in the following it is assumed that Her13.2 can heterodimerize with Her1 or Her7, that Her1 and Her7 can also homodimerize or heterodimerize with each other, and that the resulting dimers repress expression of her1, her7, and deltaC. her1 and her7 share a common 12-kb promoter [2], which contains nine copies of the consensus hairy binding site (CACGCG [9]). In the model proposed here, depicted in Figure 1, all dimer combinations between Her1, Her7, and Her13.2 compete for binding to these sites, and repress her1 and her7 transcription with strengths specific to each dimer combination. It is assumed that the deltaC promoter functions similarly to that of her1 and her7, in agreement with closely overlapping expression patterns in the PSM [2,5,6,10,11]; this does not take into account some differences in expression patterns observed in deltaC mutants [3]. Delays are taken into account for mRNA transcription and export from the nucleus and for protein translation.
None of the biochemical reaction rates has been determined experimentally. To investigate the behavior of the model, we therefore sampled values from extended, biologically realistic ranges. We screened for suitable oscillations, with at least a 5-fold difference between mRNA peak and basal levels (see Methods for a full definition), and for reproduction of experimental results obtained by perturbation of the oscillatory machinery. Because mRNA and protein copy numbers often reach low absolute levels at the trough of their oscillatory cycle, the effect of internal noise was assayed using stochastic simulations, with the same parameter sets that were identified as providing suitable deterministic oscillations. Possible reactions in the system were the same in the deterministic and stochastic cases, but stochastic simulations considered all individual reaction events explicitly.
The model depicted in Figure 1 readily gives rise to single-cell oscillations with parameters sampled in the ranges detailed above. A parameter set was deemed robust if each parameter could be varied individually by 50% or more without disrupting oscillations. By this criterion, 50% of 300 randomly selected parameter sets were found to be robust.
As a first step to identify parameter sets reproducing the morpholino-induced disruption of the spatiotemporal pattern of oscillation described in [2], parameter sets were screened with single-cell simulations for downregulation of her1 and her7 mRNA when the Her1 or Her7 translation rates were divided by ten (corresponding to 90% morpholino efficiency in blocking protein translation). Such parameter sets occurred at a low frequency (see Figure 2 for a representative example). While general robustness of the model with these parameter sets was comparable with that in the general case (30% of parameter sets were found to be robust), the parameters with the strongest influence on the oscillation period were found to be strikingly different. Parameter sets leading to correct reproduction of the morpholino phenotypes also led to a very high sensitivity of the oscillation period to the degradation rate of the Her1–Her7 dimer (and to a lower extent to various parameters describing its repressive activity): for 30% of identified parameter sets, variation of the Her1–Her7 degradation rate yielded a higher variation in period than variation of any parameter not related to delays of transcription and translation. In the general case, period-sensitivity to individual parameter variation was more even across parameters and highest for the her1 mRNA degradation rate. (For 10% of identified parameter sets, variation of that degradation rate led to the highest period variation amongst parameters other than delays.) This suggests a central role for a Her1–Her7 heterodimer, which can be tested experimentally.
Examination of the parameter values leading to correct reproduction of the her1 and her7 morpholino knockdown phenotypes showed that the repression strength of the Her1–Her7 heterodimer was strongly biased towards lower values than that of other repressive dimers. This suggests that the mechanism by which Her1 and Her7 are required for their own expression is by Her1–Her7 heterodimers acting as a “protective” species: Her1–Her7 heterodimers repress her1 and her7 expression, but compete with other dimer combinations that repress expression more strongly.
Parameter sets reproducing the her13.2 knockdown phenotype (which consists of disrupted oscillations [4]) were identified independently of those reproducing the her1 and her7 knockdown phenotypes and did not show any notable difference in period distribution. Of the parameter sets leading to correct reproduction of her1 and her7 knockdown phenotypes, 10% also led to disruption of the oscillations when her13.2 expression was knocked down (see example in Figure 2).
We have so far addressed oscillations at the level of individual cells. One important feature of the somitogenesis clock is that waves of expression sweep from posterior PSM to anterior PSM and shrink in the process [10]. The spread of an intercellular signal is not necessary for the short-term maintenance of the oscillatory pattern in mouse and chick [21–23], but it has been suggested that cellular oscillators can influence their neighbors [11,24–26]. In the case of the chick and mouse somitogenesis clocks, a gradient in the strength of intercellular coupling can lead to the formation of the correct collective pattern of oscillatory expression [24]. In those species, no molecular link is currently known between the oscillatory machinery and positional information in the PSM. However, Her13.2 provides such a link in the zebrafish clock [4], which makes it possible to investigate the detail of the mechanism.
To assess whether graded expression of her13.2 could prompt a linear chain of oscillators to form the correct collective pattern of oscillatory expression, a screen was carried out in which the dynamic structure of the PSM was reproduced by adding cells at regular intervals at the posterior end of the chain (corresponding to convergent extension and ingression from the tailbud [27]) and removing cells at the anterior end. (Cells were removed continuously, rather than in blocks corresponding to somites, because the process of segmentation, which is controlled by poorly understood molecular mechanisms, is not the subject of this study.) Each cell influenced its two anterior and posterior neighbors by providing them with the ligand Delta for Notch receptor activation, leading to increased her and deltaC transcription (the dynamics of Notch signaling were not modeled explicitly). Expression of her13.2 mRNA was assumed to be high in posterior PSM and to drop sharply in anterior PSM [4]; the regulation of her13.2 by FGF-8 was not modelled explicitly. The rates of her13.2 mRNA decay and of cell addition at the anterior end set the length of the PSM, and therefore the number of oscillations each cell experienced while in the PSM (that number was set to 12, as can be estimated from [3,27,28]).
A number of parameter sets were identified for which her and deltaC waves of expression swept from posterior to anterior, their width becoming restricted as they went along (see Figure 3A–3C and Video S1 for a representative example). The number of stripes observable at any given time depends on parameter sets and spans the range observed experimentally (from one in old embryos at specific phases, to up to three in younger embryos [10]); it was not attempted to reproduce precisely the rate at which stripes decrease in length. Positional information provided by Her13.2, as identified experimentally, is therefore sufficient to arrange oscillatory expression in the PSM in the correct pattern, within the framework of the model proposed here.
To determine whether the role of Her13.2 is to modulate intercellular coupling or to act on individual oscillators, simulations were run with no intercellular coupling (this was performed by abolishing DeltaC translation, or by replacing Notch signaling with fictitious autocrine signaling). For all parameters studied, it was found that intercellular coupling could be removed without destroying the spatiotemporal pattern. For some parameter sets, this loss of Notch needed to be compensated by an increase in the transcription rates of her1 and her7. A shift in oscillatory phase and a slight change in the oscillatory period occur when her13.2 mRNA expression drops (unpublished data); this is sufficient to set up the spatiotemporal pattern. Note that this does not contradict the fact that Notch signaling is necessary for oscillations in vivo. The model studied here has parameter sets for which Notch signaling is required for individual cellular oscillations with the correct pattern (by providing a sufficient level of her expression) as well as for intercellular coupling. In addition, the model has parameter sets for which Notch is required for intercellular coupling only. The first set of parameters corresponds to the in vivo behavior of the oscillator (such a parameter set was used to produce Figures 2–6). In vivo, intercellular coupling through Notch signaling could have the additional role of setting up the very first waves of expression, which are not the object of this study.
Even if intercellular coupling has no role in the establishment of the oscillatory pattern under the simplified conditions studied here, it could have a crucial role in synchronizing cells in vivo [11,29]. Perturbations were simulated by delaying a number of oscillators in the chain by 5 min. For some parameter sets, such perturbations were resorbed within a few rounds of oscillation (see Figure 3D–3E and Video S1 at 50 min for an example). Synchronization was much stronger for cells located in posterior PSM than for cells in anterior PSM. This might be the basis for greater plasticity of posterior PSM as compared with anterior PSM, as observed in chick [30].
To further study the role of intercellular coupling in resistance to noise, the behavior of chains of oscillators in the presence of molecular noise was assessed as described above for individual oscillators. Due to the high computational and memory costs of simulating individual reaction events with delays for a system comprising more than 50 oscillators with 13 variables each, only a very small number of parameter sets could be studied. Nonetheless, parameter sets were identified for which the patterns of oscillation in stochastic simulations were as expected, close to the deterministic form (Figure 4 and Video S2). For such parameter sets, closely similar successive rounds of oscillation showed low variability between stochastic realizations. A stochastic simulation was run with disrupted coupling, with the same parameter set as used in Figure 3. A few cells went noticeably out of synchrony with their neighbors, and local synchrony was generally not as strong as with coupling, but the global spatiotemporal pattern was not disrupted (Figure 4D–4F and Video S3). Thus, even though this parameter set leads to perturbation resorption in posterior PSM (Figure 3D–3E), the molecular clock is sufficiently precise that this capacity is not required to maintain the spatiotemporal pattern with the molecular noise simulated here.
Morpholino knockdown of her1 leads to residual her1 and her7 oscillations in posterior PSM and to defective stripe formation in anterior PSM [2]. Knockdown of her7 leads to expression of her1 throughout the PSM (with no stripes) and to posterior expression of her7 [2]. Detailed comparison of simulation data and experimental data is not possible because the latter is not quantitative, but most general features can be reproduced. Simulated knockdown of her1 or her7 on the parameter set used in Figure 3 showed residual oscillations in posterior PSM, with a high level of basal expression (see Figure 5A–5B and Video S4 for the her1 knockdown, and Video S5 for the her7 knockdown). These residual oscillations were also present in anterior PSM, but no clear stripes of expression were formed, in agreement with experimental data. The simulations showed that her1 and her7, with the biochemical parameters used, have essentially symmetrical roles. (This is because many parameters for her7 were chosen to be the same as that of their her1 counterpart, to make the computational study tractable.) Some asymmetry has been reported based on knockdown phenotypes [2]; for example, her1 knockdown leads to some her1 expression in anterior PSM, while her7 knockdown leads to loss of expression of her7 in anterior PSM. It is possible that some of this asymmetry stems from differences in mRNA in situ hybridization detection thresholds, which are unknown (her7 could have a higher detection threshold than her1).
Simulated knockdown of her1 and her7 together showed upregulated expression in posterior PSM and a salt-and-pepper pattern in anterior PSM (Figure 5C and Video S6). This is consistent with experimental results showing generalized upregulation [5]. There was a slight discrepancy in that salt-and-pepper expression has been reported to occur throughout the PSM rather than specifically in anterior PSM [5]; this remains to be investigated. Strikingly, individual cells oscillated in the anterior PSM, even though the global pattern appeared constant. This brings computational confirmation to the hypothesis that a salt-and-pepper pattern can occur by desynchronization of cells, rather than by blocking of oscillations at different phases of the cycle in different cells [11]. In the double morphant, the clock started oscillating around the transition between posterior and anterior PSM. Because the oscillation period was roughly similar, the number of cycles experienced by cells in the PSM was about 50% lower than normal.
Another way in which the spatiotemporal oscillatory pattern can be disrupted is by grafting FGF-8 coated beads. This leads to minimal disruptions of oscillations when the bead is adjacent to posterior PSM, but to anterior extension of waves in the anterior PSM [31]. Such grafting experiments were simulated by assuming that an ectopic stripe of high her13.2 mRNA expression is induced around the bead (Figure 6 and Video S7). When the bead was adjacent to posterior PSM, oscillations were not affected because the bead was assumed not to further increase her13.2 expression. When the bead reached anterior PSM, posterior oscillations were still unaffected, but anterior oscillations, if imaged at the right phase, could show anterior extension of a wave in anterior PSM.
Direct characterizations of molecular interactions in the zebrafish somitogenesis clock network are scarce. Many dimerization combinations remain untested, and potential binding sites on the her1–her7 and deltaC promoters remain unmapped. It would be a daunting task to test all possible molecular interactions and measure all biochemical reaction rates; the theoretical work presented here makes readily testable predictions and identifies select biochemical parameters of the network that are likely to have great impact on its behavior.
The existence of Her1–Her7 heterodimerization is a key feature to explain her1 and her7 morpholino knockdown phenotypes, with Her1–Her7 dimers having a protective role by competing with other dimers that repress transcription more strongly. The half-life of the Her1–Her7 heterodimer is predicted to have an important influence on the period of oscillation. Interestingly, dimerization of clock proteins is a feature shared with mouse and chick somitogenesis clocks [32]. However, the mechanism in those two clocks seems to be very different in that Lunatic fringe—which potentiates Notch activation by its ligand Delta—oscillates along with other clock genes, and that a positive feedback loop is likely to drive the oscillations [24].
The models studied here were kept simple to make it easier to extract essential features. The cellular aspects could be expanded by taking into account cell cycling throughout the PSM [29], blocking of mRNA transcription and protein translation in the mitotic phase of the cell cycle, possible cell mingling (observed in chicken [33,34]), and the effect of coupling on a 2-D or 3-D set of oscillators (rather than on a linear chain as in this study). It was assumed that the somitogenesis clock is already active in PSM progenitors (this has been suggested for early oscillations in chick [35], but does not seem to have been addressed in zebrafish); it would also be possible to have the clock inactive in progenitors and kick-started when cells join the posterior PSM. The molecular networks could also be expanded by having different enhancers active in posterior and anterior PSM, as shown experimentally [2]. This study shows that the presence of such different enhancers is not a fundamental requirement of the somitogenesis clock, but it might allow finer reproduction of the spatiotemporal pattern of gene expression and of its disruption by morpholino knockdown. Such an extension of the model might also allow for a role of fused somites, a gene essential for somite formation [36], whose activity is required for the propagation of expression waves into anterior PSM [6,10,37].
This study addressed the molecular mechanism of the somitogenesis clock oscillations. The mechanism by which the oscillations are read out to control mesoderm segmentation is not fully understood and is likely to be linked to the complex process of somite polarity establishment, most thoroughly studied in mouse [38]. Interestingly, out-of-phase oscillation of dimerization partners has been proposed as a mechanism to establish somite polarity [39]; however, the proteins considered in the present model do not oscillate out of phase. A model for segmentation based on reaction–diffusion of factors promoting anterior or posterior somite fate has also been proposed [40], but the genes considered in the present model cannot be related to such factors in a straightforward fashion.
A “clock and wavefront” model, first proposed by Cooke and Zeeman [41], is most often invoked to explain segmentation. However, modifications of the original model [30,42] suppose that clock and FGF-8 wavefront are independent, which has previously been shown not to be the case ([31]; see also Figure 4I in [30]). The present study details a molecular mechanism with strong experimental support by which FGF-8 interacts with the clock to regulate the spatiotemporal pattern of oscillation. This will in turn make it possible to investigate how clock and wavefront interact to regulate segmentation.
The step function S used to shape the her13.2 gradient is defined by
Deterministic oscillations were considered suitable if each interpeak distance in the course of the simulation fell between 10 min and 100 min, and if the amplitude of the peaks was sufficiently high, both in relative terms (at least a 5-fold difference between minimal and maximal values) and absolute terms (at least 30 molecules at the peak value). Potentially spurious peaks arising within 10 min after a previous peak were discarded from the analysis. Both her1 and her7 mRNA oscillations were assayed, each was required to meet these criteria, and the number of peaks undergone by each could not differ by more than one (so as to ensure roughly similar oscillation periods). Only her1 mRNA period oscillation was measured (her1 and her7 play symmetrical roles in the models studied here), either as the distance between the last two peaks in a simulation run, so as to allow the system to have likely reached a limit cycle after the zero initial conditions used to start the simulation, or as the average of all distances between consecutive peaks in the simulation.
The algorithm above requires the absence of monotonicity changes between major peaks. As a control, a different algorithm was also used: the autocorrelations of the her1 mRNA copy numbers through time were computed for increasing timeshifts starting from zero, and the period considered to be the first nonzero timeshift that produced a local maximum of the autocorrelation value. For deterministic simulations, the results were very close to that of the algorithm described above.
Deterministic simulations were performed with an adaptive-stepsize, 4th-order Runge–Kutta–Fehlberg algorithm [43] implemented in a custom C++ program (available on request). Numerical accuracy (taking into account both absolute and relative accuracies [43]) was set to 1%. Time points where derivatives are discontinuous (because of delays or because of the introduction or removal of oscillators in a chain of coupled oscillators) were forced to be part of the integration mesh. To speed up computations and ease RAM requirements, a subset of past solution values was stored, and delayed values required by the derivative function were linearly interpolated. To ascertain the accuracy of this method, a subset of results were compared with that obtained with a method providing 4th-order interpolation [44], with storage of all past integration steps; no significant difference was observed.
Stochastic simulations were implemented following the Gibson–Bruck algorithm [45], with a custom C++ program (available on request).
Simulations were carried out on a set of PowerPC G5 iMacs, PowerPC G5 PowerMacs, and Intel Core Duo iMacs (totaling about 16 processors), using GNU gcc 4.0.1 (with optimization setting on fast) and Intel icpc 9.1 as compilers, and on two SGI ALTIX 350 servers comprising a total of 32 Intel Itanium 2 processors and 64 GB of RAM, using gcc 3.2.3 (with optimization setting on fast) or icc 8.0 as compilers (the ALTIX servers being essentially used for stochastic simulations of chains of oscillators). |
10.1371/journal.pntd.0006640 | Clinical, environmental, and behavioral characteristics associated with Cryptosporidium infection among children with moderate-to-severe diarrhea in rural western Kenya, 2008–2012: The Global Enteric Multicenter Study (GEMS) | Cryptosporidium is a leading cause of moderate-to-severe diarrhea (MSD) in young children in Africa. We examined factors associated with Cryptosporidium infection in MSD cases enrolled at the rural western Kenya Global Enteric Multicenter Study (GEMS) site from 2008-2012.
At health facility enrollment, stool samples were tested for enteric pathogens and data on clinical, environmental, and behavioral characteristics collected. Each child’s health status was recorded at 60-day follow-up. Data were analyzed using logistic regression. Of the 1,778 children with MSD enrolled as cases in the GEMS-Kenya case-control study, 11% had Cryptosporidium detected in stool by enzyme immunoassay; in a genotyped subset, 81% were C. hominis. Among MSD cases, being an infant, having mucus in stool, and having prolonged/persistent duration diarrhea were associated with being Cryptosporidium-positive. Both boiling drinking water and using rainwater as the main drinking water source were protective factors for being Cryptosporidium-positive. At follow-up, Cryptosporidium-positive cases had increased odds of being stunted (adjusted odds ratio [aOR] = 1.65, 95% CI: 1.06–2.57), underweight (aOR = 2.08, 95% CI: 1.34–3.22), or wasted (aOR = 2.04, 95% CI: 1.21–3.43), and had significantly larger negative changes in height- and weight-for-age z-scores from enrollment.
Cryptosporidium contributes significantly to diarrheal illness in young children in western Kenya. Advances in point of care detection, prevention/control approaches, effective water treatment technologies, and clinical management options for children with cryptosporidiosis are needed.
| Cryptosporidium is an important cause of childhood diarrhea. Research on cryptosporidiosis in countries where it is endemic remains limited; few studies have comprehensively examined risk factors for children in Kenya and similar settings. We examined characteristics associated with Cryptosporidium in children with moderate-to-severe diarrhea in rural western Kenya. We found there is little to clinically distinguish cryptosporidiosis from other childhood diarrhea in the absence of point of care diagnostics. Infants had the highest odds of Cryptosporidium infection; it has been previously established that Cryptosporidium infections in infancy can have severe consequences. Prolonged/persistent duration diarrhea and growth shortfalls were significantly more pronounced among cases with Cryptosporidium. Undernutrition and stunting in children in low- and middle-income countries have predicted decreased cognitive and school performance, thus long-term consequences could be appreciable. Using rainwater as the primary drinking water source and boiling drinking water were protective against Cryptosporidium infection, thus certain water sources may contribute to transmission. Like other studies in Kenya, we predominantly identified Cryptosporidium hominis, an anthropogenic species. Advances in point of care detection, prevention and control approaches, effective water treatment technologies, and clinical management options are needed to mitigate the potentially severe and long-term consequences of Cryptosporidium infection in children.
| The Global Enteric Multicenter Study (GEMS) was undertaken to assess the burden and etiology of moderate-to-severe diarrhea (MSD) in seven countries, three in South Asia and four in sub-Saharan Africa. In all African sites, Cryptosporidium was the second-highest enteric pathogen attributable to infant MSD; in GEMS Kenya, Cryptosporidium was a major pathogen across all age groups (0–11, 12–23, and 24–59 months) [1]. Cryptosporidium was also identified as one of five pathogens with the highest attributable burden of infant diarrhea in a study of malnutrition and enteric disease (MAL-ED), a cohort study that compared diarrheal and non-diarrheal stools in children under two years old collected at community surveillance visits at 8 sites in South America, Africa, and Asia [2]. Based on GEMS data, it has been estimated that there are nearly three million annual diarrhea episodes attributable to Cryptosporidium in young children in sub-Saharan Africa [3]. Globally, acute Cryptosporidium infections are estimated to cause 48,000 annual deaths in children under five years old [4].
Cryptosporidium infections in young children in low- and middle-income countries have been associated with excess mortality [5], an excess burden of diarrhea later in life [6], and growth faltering, the deficits of which may not be recovered for those children infected during infancy [7]. Cryptosporidium has been associated with decreases in height-for-age z-scores in children, even in the absence of diarrhea symptoms [4]. Cryptosporidium infections have been associated with persistent diarrhea in Kenya [8,9]. Cryptosporidium is highly tolerant to disinfection with chlorine [10].
Nitazoxanide can treat cryptosporidiosis in immunocompetent children 1–11 years old [11]; however, it is not often available in developing countries [12] and is presently not approved for infants [11]. There is currently no vaccine available for Cryptosporidium; however, the evidence of acquired immunity suggests that one could be effective [12].
Although outbreaks of Cryptosporidium in developed countries have been studied in detail, less is known about risk factors for cryptosporidiosis in countries where it is endemic [10]. Reviews of risk factors for Cryptosporidium infection identified malnutrition, contact with domestic animals, non-exclusive breastfeeding in infants, lack of sanitation facilities, and crowded living conditions as possible risk factors for infection in low- and middle-income countries [13,14]. Few studies have examined risk factors for Cryptosporidium infection in Kenyan children [15–17]. In Kenya, risk factors for Cryptosporidium in children include being HIV-positive [17], or having an HIV-positive mother [15].
We describe the prevalence of Cryptosporidium infections in Kenyan children under five years old with MSD, assess clinical, environmental, and behavioral characteristics associated with Cryptosporidium infection, and describe the outcomes and consequences of cryptosporidiosis.
We evaluated data collected in Kenya from cases enrolled in GEMS, a four-year, prospective, age-stratified, health facility-based matched case-control study of MSD among children aged 0–59 months residing within a defined and enumerated population. The rationale, study design, clinical and microbiologic methods, and assumptions of GEMS have been described elsewhere [18,19]. Briefly, GEMS enrolled MSD cases from selected sentinel health facilities in each of three age strata (0-11, 12-23, and 24-59 months old), along with 1–3 matched community controls who had not had diarrhea in the week before enrollment. MSD was defined as having three or more loose stools in the previous 24 hours, with onset within the 7 days prior to enrollment, and having one or more of the following illness severity characteristics: loss of skin turgor, sunken eyes, required intravenous fluid rehydration, dysentery (blood in stool), or required hospitalization [18].
At enrollment, demographic, clinical, epidemiological information, and stool samples were collected. Cryptosporidium oocyst antigens were detected in whole stool specimens by enzyme immunoassay (EIA; TechLab, Inc, VA, USA). Detailed laboratory methods are described elsewhere [19]. DNA was extracted from a subset of stools that were EIA-positive for Cryptosporidium. Restriction fragment length polymorphism analyses and DNA sequencing of polymerase chain reaction (PCR) products were used to identify Cryptosporidium genotypes for these specimens at the U.S. Centers for Disease Control and Prevention (CDC) [20].
To assess each child’s health status, a home visit including focused physical exams and anthropometric measurements was conducted ~60 days (acceptable range 50–90 days) following enrollment. Mortality that occurred at any time between enrollment and this follow-up was recorded.
In Kenya, children were enrolled between January 31, 2008 and January 29, 2011, and again from October 31, 2011 to September 30, 2012. The study was conducted in Siaya County, in the areas of Gem and Asembo, and during the second enrollment period, in the areas of Asembo and Karemo due to the Kenya Medical Research Institute (KEMRI)/CDC health and demographic surveillance system moving activities. This health and demographic surveillance system has been operating in these communities since 2001. The study setting has high rates of child mortality, malaria, HIV, and tuberculosis, and has been described elsewhere [21,22].
Analyses were performed in SAS 9.4 (SAS Institute, Inc., Cary, NC) and R 3.4.0 (R Foundation for Statistical Computing, Vienna, Austria). To assess variables associated with Cryptosporidium positivity, univariable logistic regression models were used to compute odds ratios (ORs) and 95% confidence intervals (CIs). Since Cryptosporidium risk factors may be modified by age and the sample size might limit detection of interactions, we assessed for effect modification by age category for all variables in separate models with a p<0.05 cutoff for significance. To assess whether each risk factor was confounded by socioeconomic status (SES) we ran models with and without SES and considered confounding if effect sizes changed >10%.
Two multivariable models were generated to identify clinical, demographic, environmental, and behavioral characteristics associated with being Cryptosporidium-positive. The first model examined the clinical presentation at enrollment of case children, including age strata and sex (see Table 1) and all variables in Table 2 (except duration of diarrhea, which includes information collected post-enrollment). The second model sought to identify demographic, environmental, and behavioral characteristics that may be risk factors for Cryptosporidium infection (all variables in Table 1, and the following caretaker-reported water, sanitation, and hygiene characteristics collected at enrollment: primary source of drinking water, whether water was always available from the main drinking water source, whether the child was given stored water in the two weeks prior to enrollment, whether the caretaker boiled or filtered drinking water, whether there was a facility for feces disposal, whether the caretaker uses soap when washing hands, and whether the caretaker washes their hands at the following times: before eating, after defecating, before nursing, before cooking, after cleaning a child, and after touching an animal). Breastfeeding was not considered in either model as collinearity with age was identified, and information on breastfeeding is only available for children under two years old in the first three years of GEMS (n = 1,083); questions on breastfeeding changed during the fourth study year. Model selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO) method using the minimum error lambda [27,28]. While LASSO methods were used to identify variables for inclusion, parameter estimates and CIs are derived by standard logistic regression maximum likelihood methods.
Written informed consent was collected from all parents of children who participated in GEMS. The GEMS protocol was approved by the Scientific and Ethical Review Committees of KEMRI (Protocol #1155) and the Institutional Review Board (IRB) of the University of Maryland School of Medicine, Baltimore, MD, USA (UMD Protocol #H-28327). CDC (Atlanta, GA, USA) formally deferred its review to the UMD IRB (CDC Protocol #5038).
Among the 1,778 MSD case children enrolled, Cryptosporidium was identified in 195 cases (11.0%). Cryptosporidium infections were more frequently identified in infants (<12 months old), with a peak in Cryptosporidium infection at 6–11 months old (Table 1 and Fig 1). Compared to case children aged 24–59 months, infants had over triple the odds of being Cryptosporidium-positive (OR = 3.32; 95% CI: 2.08–5.31). Other demographic and household characteristics were similar between Cryptosporidium-positive and Cryptosporidium-negative cases (Table 1). A non-statistically significant relationship between Cryptosporidium status and having agricultural land was confounded by SES. As no other variable was confounded by SES, only unadjusted effect measures are shown in Table 1.
The clinical presentation of Cryptosporidium-positive and Cryptosporidium-negative cases was similar (Table 2). Mucus in the stool was significantly associated with being Cryptosporidium-positive (OR = 1.72, 95% CI: 1.21–2.51). Only age category and mucus in stool remained in the final multivariable clinical model (not presented). The findings were the same when children who were enrolled multiple times as a case were excluded. Having mucus in the stool remained significantly associated with Cryptosporidium infection controlling for age (aOR = 1.50; 95% CI: 1.05–2.20).
Approximately two-thirds (66%) of Cryptosporidium-positive cases with daily information on diarrhea experienced prolonged or persistent diarrhea, compared to approximately half (51%) of Cryptosporidium-negative cases (Table 2). Compared to cases experiencing acute diarrhea, cases experiencing prolonged diarrhea were significantly more likely to be Cryptosporidium-positive (OR = 1.68; 95% CI: 1.18–2.37); cases experiencing persistent diarrhea were also significantly more likely to be Cryptosporidium-positive compared to cases experiencing acute diarrhea (OR = 3.43; 95% CI: 1.97–5.98).
At enrollment, sex was a significant effect modifier of the relationship between Cryptosporidium and stunting/severe stunting. Among girls, Cryptosporidium-positive cases had significantly greater odds of being stunted at baseline than Cryptosporidium-negative cases (OR = 1.82, 95% CI: 1.10–3.01). There were no other statistically significant differences in malnutrition indicators between Cryptosporidium-positive and Cryptosporidium-negative cases at enrollment (Table 3).
At the 60-day follow-up, Cryptosporidium-positive cases had significantly greater odds of being stunted (aOR = 1.65, 95% CI: 1.06–2.57), underweight (aOR = 2.08, 95% CI: 1.34–3.22), or wasted (aOR = 2.04, 95% CI: 1.21–3.43) compared to Cryptosporidium-negative cases, controlling for baseline status for each measure (Table 3).
Cryptosporidium-positive cases had significantly larger negative changes in HAZ and WAZ measures from baseline to follow-up. When considering HAZ by sex, female Cryptosporidium-positive cases had significantly larger negative changes in HAZ compared to female Cryptosporidium-negative cases (Table 3).
HIV status was available for 58.8% of GEMS-Kenya cases. Of the 114 Cryptosporidium-positive cases with available HIV test results, 5 (4.4%) were HIV-positive, compared to 3.0% (28/932) of Cryptosporidium-negative cases (p = 0.39). There was no significant association between being Cryptosporidium-positive and having an HIV-positive biological mother (n = 1,194 tested; OR = 1.17; 95% CI: 0.77–1.77).
Most children under two years old (81.4%) were partially breastfed. Breastfeeding was similar between Cryptosporidium-positive and Cryptosporidium-negative cases (Table 4).
There was no significant difference in the proportion of Cryptosporidium-positive cases and Cryptosporidium-negative cases who were hospitalized at enrollment (13.3% vs. 10.5%, p = 0.24). Among those with 60-day follow-up information, 9 (4.8%) of 187 Cryptosporidium-positive cases and 53 (3.5%) of 1,531 Cryptosporidium-negative cases died by the time of follow-up (p = 0.35). The cause of death, as per verbal autopsy, for the 9 children who died and had Cryptosporidium identified in their stool was as follows: HIV/AIDS related (n = 5), diarrhea/gastroenteritis (n = 1), pneumonia (n = 1), and anemia (n = 1); one child did not have a verbal autopsy completed.
The most common caretaker-reported primary drinking water sources were rainwater (35%), surface water (31%), other improved water sources (23%), and other unimproved water sources (11%); (Table 5). Compared to cases in households that used rainwater as the primary source of drinking water, case children living in households using other improved sources or unimproved sources (other than surface water) had significantly higher odds of Cryptosporidium infection (OR = 1.72; 95% CI: 1.14–2.58 and 2.12; 95% CI: 1.31–3.41, respectively).
Few caretakers of GEMS-Kenya case children reported boiling or using a ceramic filter to treat drinking water; however, those reporting one of these methods had significantly lower odds of Cryptosporidium infection compared to those who didn’t (OR = 0.52, 95% CI: 0.25–0.96), predominantly driven by those who boiled water. Only two households reported filtering.
Reported handwashing behavior was similar among the caretakers of Cryptosporidium-positive and Cryptosporidium-negative cases (Table 5).
Only age category remained in the final multivariable demographic, environmental, and behavioral characteristics model; thus, this model is not presented.
DNA was extracted from a random subset of 64 (40%) of the 160 Cryptosporidium-positive stool specimens from GEMS-Kenya case children enrolled in the first three years. Nested 18S PCR detected Cryptosporidium in 43 (67%) of these specimens. Of the 43 specimens, 35 (81%) were of the species C. hominis and 6 (14%) were C. parvum. C. meleagridis and C. canis were found in one specimen each.
Of the 195 Cryptosporidium-positive cases, 142 (72.8%) also had one or more additional enteric pathogens identified in their stool; enteric co-infections were common throughout the study population (Fig 2). The characteristics of case children with only Cryptosporidium detected in their stool were generally similar to Cryptosporidium-positive case children with multiple enteropathogens (S1 and S2 Tables); the only significant clinical difference was in the child’s mental state at enrollment (S1 Table). Variables chosen for multivariable models were unchanged when excluding children who were enrolled more than once as an MSD case.
This study evaluated the clinical, environmental, and behavioral characteristics associated with Cryptosporidium infection among children under five years old with MSD in rural western Kenya. Overall, 11% of children with MSD had Cryptosporidium identified in their stool; the majority (81%) of genotyped samples were C. hominis. Among MSD cases, being an infant, having mucus in stool, and having prolonged or persistent duration diarrhea were associated with being Cryptosporidium-positive. Boiling drinking water and using rainwater as the main drinking water source appeared to protect against Cryptosporidium infection in MSD cases. Among girls, Cryptosporidium-positive cases were more likely to be stunted at baseline compared to Cryptosporidium-negative cases. Cryptosporidium-positive cases had longer-term consequences in terms of malnutrition, as these children were more likely to stunted, underweight, or wasted at follow-up (controlling for baseline status), and have significantly larger negative changes in height- and weight-for-age z-scores.
Except for having mucus in stool, which could be associated with Cryptosporidium adhering to the small intestine mucosa, possibly causing inflammation [29], the clinical presentation of children with MSD was similar for Cryptosporidium-positive and Cryptosporidium-negative cases, as was observed in another study of Kenyan children with diarrhea [16]. This finding highlights the difficulty in clinically diagnosing Cryptosporidium among children with MSD in this setting and underscores the need for point of care rapid diagnostics for Cryptosporidium.
Infants were over three times more likely to have Cryptosporidium identified in their stool compared with children aged 24–59 months. The peak of infection at 6–11 months in this study is similar to the age pattern of Cryptosporidium infections previously reported in sub-Saharan Africa, though an earlier peak than other studies in Kenya [9,16]. This timeframe may coincide with the introduction of complementary foods or drinking water. The high prevalence of Cryptosporidium infections in young children is concerning as Cryptosporidium infections in early childhood have been associated with numerous poor outcomes, sometimes lasting beyond the initial infection [6,7], as evident in our findings.
Prolonged and persistent duration diarrhea, and growth shortfalls subsequent to enrollment were significantly more pronounced among Cryptosporidium-positive cases compared to other children with MSD. Prolonged and persistent diarrheal episodes occurring in infants have been previously associated with growth shortfalls [30]. The proportion of Cryptosporidium-positive cases who were underweight and wasted increased from baseline to follow-up. This could result from many days of diarrhea experienced by these children. There was also an increase in the proportion of Cryptosporidium-positive cases who were stunted from baseline (29%) to follow-up (39%). Undernutrition and stunting among children in low- and middle-income countries have predicted decreased performance in school and on cognitive tests in previous research [31], thus even longer-term consequences could be appreciable although unexplored in the current study. It is estimated that growth faltering contributes substantially to the overall global burden of disease from Cryptosporidium infections in children [4]. By the time of follow-up, 4.8% of Cryptosporidium-positive and 3.5% of Cryptosporidium-negative cases had died. Although the difference was not statistically significant, this warrants close future attention since other research has shown an association between Cryptosporidium and excess mortality for children who became infected in infancy [5].
Like other studies in Kenya [9,16,32], our findings indicate that person-to-person transmission is likely the predominant route for Cryptosporidium infection in rural western Kenya, since the main host for C. hominis is humans [10]. Infections may thus more commonly result from exposure to human feces than animal feces. The presence of animals in the compound was examined in univariable analyses but was not reported in detail or included in the risk factor model selection, as the ownership of many types of animals was not associated with Cryptosporidium infection (S3 Table), though it may be associated with unmeasured confounders (e.g., higher income). Notably, in another study, C. hominis was associated with more severe clinical symptoms in Kenyan children compared to C. parvum [32], although we have too little data in GEMS Kenya to examine this.
Using rainwater as the main drinking water source was common and was significantly protective against Cryptosporidium infections. Rainwater may be less contaminated with Cryptosporidium, or this finding could be related to the seasonality of Cryptosporidium infections. The proportion of households using rainwater as the main drinking water source varied by month, ranging from 3%-74%. We did not explore Cryptosporidium infections by month/season, as the biweekly enrollment targets for GEMS make interpretation of pathogen-specific seasonal analyses challenging. However, the fact that using rainwater as the primary drinking water source and boiling drinking water were both protective against Cryptosporidium infections indicates that drinking water source choices and certain treatment options may be effective in reducing Cryptosporidium infections and signals that water may play a role in transmission.
A limitation of this work is that a comparison could not be made between Cryptosporidium-positive GEMS cases and GEMS controls, as reliable population weights were not available at the time of analysis. The factors associated with Cryptosporidium compared to other individuals with MSD may be different from those risk factors that would be seen when compared to healthy controls. Using MSD as a condition for inclusion for our study may lead to spurious associations, as it is potentially a common effect of both the exposures and the outcome of interest [33]. Our ability to explore data on those with only Cryptosporidium infections was limited due to the small number of single-pathogen Cryptosporidium infections; however, those who presented with Cryptosporidium alone had similar characteristics to those who presented with multiple-pathogen Cryptosporidium infections. We were not able to assess the association between breastfeeding and Cryptosporidium because of (1) the collinearity between breastfeeding and age, and (2) the small number of children who were either exclusively or not breastfed in certain age groups. We also could not examine anthropometric outcomes by age or by the number of enteric pathogens isolated in stool, due to the small number of Cryptosporidium-positive children in some age groups and the small number of cases who presented with single-pathogen Cryptosporidium infections. We examined HIV status and malnutrition in our analyses, and performed a sensitivity analysis related to enteric co-infections; however, we did not have information on other co-morbidities that may be associated with Cryptosporidium infection.
The burden of diarrhea attributable to Cryptosporidium differed between GEMS and MAL-ED, especially for children 1–2 years old; however, GEMS generally considered more severe cases of diarrhea than MAL-ED. Other differences between the studies have been described elsewhere [34]. While GEMS and MAL-ED found Cryptosporidium to be significantly associated with diarrhea among infants, there were differences between study sites; in other studies Cryptosporidium has been isolated in non-diarrheal stools as often as in diarrheal stools [35,36]. Dissimilarities in study design (e.g., the time from diarrhea onset to stool collection) or laboratory methods may partially explain the differences observed, and host susceptibility and other risk factors are likely to vary across settings [35]. However, Cryptosporidium infections have been associated with growth shortfalls in asymptomatic children without diarrhea, thus identification and treatment of Cryptosporidium should remain a priority for young children in settings where it is endemic [35,36].
The high prevalence of cryptosporidiosis among young children in our study, coupled with other research that shows extended long-term effects of Cryptosporidium infections and diarrhea early in life, underscores the need for preventive measures aimed at households with young children, as well as improved diarrhea case management. Early diagnosis and management of cryptosporidiosis may mitigate subsequent growth deficits and other long-term consequences. Increased availability of nitazoxanide or new treatments, point of care rapid diagnostics for Cryptosporidium, additional insights into the role of appropriate WASH practices and technologies in childhood cryptosporidiosis, and vaccine development could reduce the burden of disease in such settings. Since Cryptosporidium-positive cases experienced more days of diarrhea and subsequent malnutrition than other MSD cases, increased promotion of the use of zinc in the management of diarrhea, and continued feeding of children with diarrhea should be undertaken, per WHO/UNICEF guidelines [37]. As rotavirus vaccine coverage increases, potentially leading to an altered enteric pathogen landscape, continuing to examine the impact and relative importance of Cryptosporidium infection among infants should remain a priority.
|
10.1371/journal.pntd.0007709 | Leprosy in elderly people and the profile of a retrospective cohort in an endemic region of the Brazilian Amazon | Leprosy has a global presence; more than 180 thousand new cases were registered in 2013, 15% of which were found in the Americas. The elderly are a very susceptible demographic in terms of developing illnesses, mainly because of characteristics natural to the senescence of the human organism. This study’s goals were to analyze leprosy in an elderly population from a hyperendemic region of the Brazilian Amazon in a historical series from 2004 to 2013 and to determine the clinical and epidemiological profile of a series of leprosy cases of elderly people in the period spanning from 2009 to 2013.
To achieve these goals, an observational, longitudinal, retrospective and descriptive study was put together to analyze leprosy in elderly people from data acquired from the Notification Aggravations Information System. Furthermore, a profile of the disease from a retrospective cohort based on data collected from medical records was developed.
The number of new cases and the leprosy detection rate decreased across the observed period but remained stable among the elderly. The trend for the next ten years indicates decreases in the number of cases and in the detection rate in the general population and an increase in only the elderly. The overall profile was characterized by a predominance of males (64.32%), the multibacillary clinical form (87.57%), Type 1 reaction episodes (37.50%) and some physical incapacity at diagnosis (49.19%). The risk of reaction was greater in the first six months of multidrug therapy, and the positive result from the skin smear was associated with the greater chance of reactional condition development.
The resulting data demonstrate that leprosy amongst the elderly deserves attention because of the increased susceptibility to disability in this age group, with their higher risk of reaction and their greater level of co-morbidity.
| Leprosy, despite being an ancient disease, still represents a challenge to public health systems today. There are still just a few studies about it, particularly among the elderly. It is known that they constitute a very heterogeneous group in terms of immune response to infections, alterations to the peripheral nervous system and predisposition to situations of vulnerability and functional dependency. The Amazon region is a hyperendemic region for leprosy and has been trying to address, along with the rest of Brazil, a rapid increase in the population’s life expectancy. This article surveys medical records from elderly people diagnosed with leprosy in a five-year period at the metropolitan region of Belém, state of Pará (Brazil), identifying a predominance of the multibacillary forms of the disease, a high prevalence of leprosy reactions mainly during treatment with multidrug therapy, and the presence of some physical incapacity in most of the people evaluated. It is expected that this study will contribute to knowledge about the clinical and epidemiological characteristics of leprosy among the elderly and stimulate the making of new studies on the theme.
| Poverty is closely related to the occurrence of Neglected Tropical Diseases (NTDs), and leprosy has high endemicity mainly in Africa, the Americas, Southeast Asia and the Eastern Pacific, where social and economic inequalities reflect adverse conditions of life and health for the population [1]. Brazil is number two in the world for leprosy, registering more than 30 thousand new cases a year [2]. In North Brazil, the state of Pará, a region with a low Development Human Index (DHI) in the country, is considered hyperendemic with 3,917 notified cases in 2014 alone [3].
Leprosy can affect people of all ages, including the elderly. The growth of the elderly demographic is a remarkable reality. In 2000, there were already more people aged 60 and above than children aged below 15 in the world, and, according to projections, in 2050, these numbers will hold[4]. When diagnosed and treated late, leprosy leads to physical disability [5] that, when combined with the aging process and other comorbidities, can cause the loss of personal autonomy to the elderly person [6]. Additionally, leprosy reaction episodes, the phenomena potentially responsible for the functional loss of the peripheral nerves [7], also results in disability that contributes to greater vulnerability and dependency in the elderly [8].
With aging, alterations in the peripheral nervous system occur, such as a reduction in fiber myelination, decreasing nervous system conduction speed [9] and compromising the pressure and tactile senses. These alterations make the elderly person more susceptible to skin lesions that may render the leprosy diagnosis more difficult and interfere with the evaluation of these patients [10, 11].
It is necessary for health professionals, relatives and caretakers to be attentive to leprosy symptoms in elderly people, mainly in endemic areas, such that it is necessary to perform a detailed clinical investigation to make sure that diagnosis and treatment occur as early as possible. Therefore, the objective of this study was to analyze leprosy in the elderly population from the state of Pará in a historical series spanning from 2004 to 2013 and to determine the clinical and epidemiological profile of a series of leprosy cases among the elderly in a highly endemic area inside the Amazon region in the period from 2009 to 2013.
Since this was a retrospective cohort study with elderly people, the Human Research Ethics Committee from the State University of Pará accepted to proceed to data compilation and analysis with no previous informed consent obtained from the participants (CAAE protocol number 41597615.5.0000.5174 CEP-CCBS/UEPA). With the objective of assuring the confidentiality of the collected data from the medical records, the people responsible for the health units surveyed signed a Consent Term for Database Usage. All clinical and epidemiological data were anonymized.
This study consisted of two parts, the first being an observational, longitudinal, retrospective study by means of the complete analysis of leprosy cases diagnosed in elderly people in the state of Pará, so as to observe the trend of the disease over the timeframe during which it was observed. The second part is a clinical, epidemiological approach evaluating a retrospective cohort of leprosy cases in the elderly, so as to complement the profile observed in the evaluation of the longitudinal study.
The sample of the longitudinal study was composed of new leprosy cases in the state of Pará in the historical series spanning 2004 to 2013 noted in the Notification Aggravations Information System (SINAN) and available at the Computing Department of the Unified Heath System website. These data were used to analyze leprosy trends and determine the detection rate in new cases of the disease per 10000 inhabitants among the general and elderly populations. Calculations of the detection rate were based on the values of the total resident population in the state of Pará taken from annual estimations of the Brazilian population elaborated by the Brazilian Institute of Geography and Statistics (IBGE).
The epidemiological clinical characterization was completed with data on a retrospective cohort among elderly patients diagnosed with leprosy that started and concluded treatment at the Sanitary Dermatology Specialized Reference Unit ‘Dr. Marcello Candia’, the Tropical Medicine Center at Federal University of Pará, the Basic Health Unit of Guamá and the Health School Center of Marco in the period from 2009 to 2013. Although the sample does not represent all of Pará State, we assume its importance in light of the study locations’ status as recognized leprosy centers in Pará State in addition to receiving patients from various locations and being themselves located in endemic areas. Though small, the sample proves enlightening as it relies on a multi-professional team equipped to treat leprosy, with each patient evaluated efficiently and each case handled adaptively.
For the longitudinal study the sample was of the new leprosy cases in the state of Pará noted in the SINAN in the historical series spanning 2004 to 2013. The epidemiological clinical characterization was of the medical records in the period from 2009 to 2013. All patients 60 years-old or older were considered elderly, in accordance with the National Geriatric Policy of the Brazilian Ministry of Health.
The types of data collected included demographics (age group and gender), clinical (operational classification, clinical form, therapeutic scheme used, disability grade at the time of diagnosis, associated comorbidities and presence of leprosy reactions) and laboratorial (smear skin index at the time of diagnosis and serology results for ELISA anti-phenolic glycolipid-1—anti-PGL-1 at the time of diagnosis). The clinical forms obeyed the Madri Classification [12] and the higher disability grade at the time of diagnosis followed the Disability Grade Classification from the World Health Organization (WHO).
The descriptions of the medical records were used to consider the reactional episodes made by a health professional and referring to clinical signs and symptoms typical to leprosy reactions or just to the type of reaction and were classified as mixed reaction (types 1 and 2) in all the patients who presented, simultaneously or not, with reaction episodes of types 1 or 2 [13]. In the absence of indication on the medical record about the type of reaction, the described reactions were classified according to the clinical criteria found in the Directives Project from the Hansenology Brazilian Society and the Dermatology Brazilian Society [14]. Patients with unclassified reactions were the ones whose medical records lacked any report of reactional signs and symptoms or classification regarding the type of reaction were registered only as “in reaction” or “in reactional state” in the medical records, such as those who presented with only clinical signs and symptoms followed by treatment without the possibility of classifying the situation because of the absence of specific characteristics of a certain type of reaction. Only reactions that occurred during the multidrug therapy and up until 24 months after medical discharge were considered, excluding reactions present at the time of diagnosis.
For longitudinal study, a the sample was composed of 50,094 new leprosy cases in the state of Pará noted in the SINAN, corresponding to all notification in the historical series spanning 2004 to 2013. The calculation methods used for the presented detection rates were as follows? 1) Leprosy detection rate in the general population = Number of confirmed new leprosy cases in residents / Total resident population in the given period X 100,000; 2) Leprosy detection rate in children under 15 years of age = Number of confirmed new leprosy cases in residentes under 15 years of age / Total resident population in the given period X 100,000; 3) Leprosy detection rate in the elderly = Number of confirmed new leprosy cases in resident elderly people / Total resident population in the given period X 100,000. The epidemiological clinical characterization was completed with to 185 eligible medical records in the period from 2009 to 2013.
This study analyzed the leprosy trends for the next 10 years with variables that included the detection coefficients per 10 thousand inhabitants and the number of new cases among the general population and new cases among the elderly population in a ten-year period. To obtain these values, polynomial regression models for temporal series were used with modeling of third order polynomial regression and curve adjustment models.
To analyze the data from the retrospective cohort study, measurements of central tendency (arithmetic median) and variability (standard deviation) were calculated. To verify intergroup differences, Chi-square tests or G Tests were used. Survival curves were generated using the Kaplan-Meier test to evaluate the occurrence of the first reactional episode according to the administered therapeutic scheme. The Odds Ratio (OR) was calculated between the final disclosure (leprosy reactions during and after the treatment) and the laboratorial exam results, with consideration of the 95% confidence interval (IC). The Spearman correlation non-parametric test was used to verify the degree of association between the smear skin index and the number of reactional episodes via the Pearson correlation coefficient (r).
The data were analyzed with BioEstat 5.3 software considering a significance level of 5% (p-value ≤ 0.05).
The number of new leprosy cases registered among the general population of the state of Pará in the historical series spanning the years from 2004 to 2013 was 50,094, and 5,447 of those cases included elderly individuals. There was a reduction in the number of new cases in the elderly population from 2004 to 2009 and a peak in the year 2012. There was an increasing trend between 2010 and 2012. The detection rate in the general population of the state of Pará was highly variable, diminishing from 9.61 to 4.89 per 10 thousand inhabitants. Among the elderly, the detection rate dropped throughout the years, with decreasing variation, and had a high of 0.87 and a low of 0.64 per 10 thousand inhabitants (Fig in S1 Fig).
The analysis of survival after the occurrence of leprosy reactions starting from the beginning of multidrug therapy treatment according to the operational classification and the number of doses administered to the elder individuals diagnosed with leprosy showed that the first reactional episode occurred mainly in the first six months of treatment, demonstrating that the risk for reaction was higher in the initial months of treatment and decreased progressively with time (Fig in S2 Fig).
Initially, a survey was made of 256 medical records from elderly patients diagnosed with leprosy in the aforementioned period. However, 62 patients who experienced interruptions in treatment at the surveyed units because of death, abandonment or transfer were excluded, which included 9 patients who had incomplete or inadequate medical records. The 185 remaining cases were followed from the start of treatment up until at least 24 months after the medical discharge. From the 185 elders surveyed, 64.32% were male and 69.73% were in the 60- to 69-year-old age group (67.50 years, on average). The predominant operational classification was multibacillary, and 62.70% of the elderly presented with a dimorphic clinical form. Among the therapeutic schemes used, 69.73% of the elderly went through 12 doses of multibacillary multidrug therapy (MDT/MB) and only 9.73% went through 6 doses of paucibacillary multidrug therapy (MDT/MB) (Table in S1 Table).
The occurrence of leprosy reactions in the elderly was 64.86% (Table in S2 Table). From those reactions, 37.50% presented with type 1 reactions, with a predominance of the Borderline and Lepromatous clinical forms, and 65.83% of the patients were treated with prednisone during the episodes. Among the elderly that developed reactions, only 115 presented with reactional signs and symptoms described in the medical records, and 34.78% manifested new erythematous plaques, infiltrated and/or edematous, and 21.74% manifested signs and symptoms of neuritis in type 1 reactions (Table in S3 Table).
Among the 147 elderly that went through the smear skin examination at the time of diagnosis, 65% of the patients who developed reactions presented with a positive skin examination index at the time of diagnosis. The odds ratio demonstrated that individuals with a positive skin examination index at the time of diagnosis had a 6.07 times higher chance of developing a reaction compared to individuals with negative results (p < 0.0001). The chances of developing a reaction was 2.93 times higher in those patients with a positive ELISA anti-PGL-1 serology result performed at the time of diagnosis (p = 0.1605) (Table in S4 Table).
Among the elderly, 49.19% already presented with some physical incapacity at the time of diagnosis (Table in S5 Table). The presence of disability grade 1 or 2 at the time of diagnosis was more prevalent in the Multibacillary, Borderline and Lepromatous clinical forms. Considering the comorbidities present, systemic arterial hypertension (28.65%) and diabetes mellitus (13.51%) were the most prevalent (Table in S6 Table).
The leprosy detection rate per 10 thousand inhabitants in the general population of the state of Pará showed a decreasing trend throughout the years. The elderly population of the state also showed a reduction in the number of cases, but the detection rate suffered a less significant drop and an approximately constant conformation. Therefore, the present study demonstrated a trend for a decrease in new cases and in the detection rate among the general population for the next ten years, and, in contrast, a trend for an increase when only the elderly population was evaluated. Such a situation among the elderly population can be explained by the higher life expectancy achieved in later decades, which resulted in a higher number of new cases diagnosed in this group.
Concerning the gender of the elders in the series of cases, there was a predominance of males that corroborates the data from Monteiro et al. [15], in which 60.3% of patients were males older than 60 years, and from Nobre et al. [16], in which there were 15.11% more males than females. However, considering that the population studied was composed of elders and that there was a feminization of the aging process because of the higher life expectancy of women [17], a predominance of females would be expected in this study. There was not a predominance of females, probably due to the long incubation period of leprosy, which can last up to seven years [11]. Maybe the men were infected previously and manifested the signs and symptoms only at an old age. It is also possible that a late diagnosis occurred, because the time between the appearance of signs and symptoms and the diagnosis can vary from a month to seventeen years [18].
The predominance of the aforementioned 60- to 69-year-old age group can be explained by the fact that it is the major age group among the elderly in the state of Pará and in Brazil, and, because these younger elders typically have more social contact, they are more susceptible to contracting leprosy. According to data from the Brazilian Institute of Geography and Statistics [19], in the last ten years, the frequency of leprosy in individuals older than 60 years old in this age group was bigger than on all the other age groups.
There was a predominance of the multibacillary forms over the paucibacillary forms of the disease, with a preponderance of the Borderline clinical form, which was in agreement with the study from Vieira et al. [20] that found a higher prevalence of the Borderline form in the leprosy profile of all ages, summing to 42.21% of the cases found.
According to Miranzi, Pereira and Nunes [21], the occurrence of multibacillary cases has a directly proportional relation to increased age. This relation could be due to the long incubation period of the disease combined with late diagnosis.
Concerning the reactional episodes, the type 1 reaction was more frequent in the evaluated patients, such as in the studies of Pinto et al. [22] and Chabra et al. [23] who studied individuals from different age groups. In this reactional type, the active participation of T lymphocytes occurs, with tissue production of Th1 cytokines (IL-2 and IFN-α) and pro-inflammatory cytokines such as the TNF-α [24]. In the elderly, in turn, there was an increase in the number of memory T lymphocytes in relation to the naive T lymphocytes due to chronic exposure to infectious agent antigens throughout life, implicating greater cytokine production and contributing to the pro-inflammatory state in the elderly [25,26], which may explain the higher occurrence of type 1 reaction in this group of individuals.
The prednisone was the most used medicine to treat leprosy reactions, especially type 1 reactions, as expected, given that corticosteroids are recognized as the drug of choice in this reaction for its suppressive effect on the inflammatory process, diminishing the INF-ɣ and TNF-α pro-inflammatory cytokines, and for their importance in the recovery of neural functions in the post-reactional period [27].
Regarding treatment of type 2 reactions, there was greater use of thalidomide associated with prednisone both in the isolated reactions and the mixed ones, a result similar to the ones found by Teixeira, Silveira and França [28] and Nazario et al. [29] in research with patients from various age groups. In Brazil, thalidomide is the drug of choice for treating type 2 reactions because of its immunosuppressive effect, allowing most patients to reach full resolution of the skin lesions within seven days [30,13]. However, moderated and aggravated type 2 reactional episodes can occur with peripheral neuritis such that the associated use of systemic corticosteroids may be necessary [31]. Effects associated with the use of corticoids in the elderly relate primarily to comorbidities that accompany aging, such as hypertension, muscular atrophy, and osteoporosis, for example. Effects related to the use of corticoids in the elderly may be diminished by the use of prescriptions only in severe episodes, such as leprous reactions, besides gradual reduction of the dosage [32, 33].
The analysis of survival based on the occurrence of leprosy reactions in relation to time demonstrated that the first reactional episode occurred mainly in the first six months, both in paucibacillary and multibacillary individuals. In the initial months of multidrug therapy treatment, the risk of occurrence was higher and diminished progressively throughout the months. In general, the reactional episodes appeared in the first six months of multidrug therapy in virtue of the rapid destruction of the bacilli by the medicine, which increases the risk of reactions considerably [34,35].
This study presented evidence of the importance of smear skin index elevation as a risk factor for the development of the reaction from the observation of a positive association between the smear skin index at the time of diagnosis and the number of reactional episodes during and after multidrug therapy. Additionally, the present study verified the higher chance of developing reactions when compared to individuals with negative results on this same test. Such findings are in agreement with the results found in the studies of Antunes et al. [36] and Brito et al. [35] and support the causal association between the bacillary load and the development of reactional states in the scientific literature.
Positive ELISA anti-PGL-1 serology at the time of diagnosis did not represent a predictive factor relevant to the occurrence of reactions during and after MDT treatment in the present research. There is a possibility, however, that the studied variables did not have a positive association due to the low number of patients who had the test, given that the serology needed to investigate leprosy is not obligatory for a diagnosis in Brazil.
In relation to the physical capacity evaluation, 95.68% of the elders were evaluated at the time of diagnosis in a way that the health units surveyed followed the recommendation of the Health Ministry, which indicate that a physical disability evaluation must be performed in at least 90% of the leprosy patients at the time of diagnosis and at the time of discharge to be considered active and of good service quality [37].
The greater prevalence of grade 1 physical incapacity among the patients with the presence of a physical disability is a piece of data that must be considered with caution because the evaluation of sensibility in the elderly can be compromised by the neurological alterations resulting from the senescence process in such a way that the altered sensibility as an outcome of leprosy can also be associated with aging itself. There may be reduction in the sense acuity in elders due to morphological alterations, size, density and location of the nociceptors in a way that, as aging progresses, more distance or touch pressure is needed for touch be perceived [38].
The more frequent comorbidities in the present study were arterial hypertension and diabetes mellitus, corroborating the data from Perry [39], whose research about the life quality of people with leprosy from all age groups found that diabetes mellitus and systemic arterial hypertension were the most frequent comorbidities among these patients and can, combined with leprosy, contribute to the installation and aggravation of physical disability and interfere in the social and economic lives of the patients.
Because the elderly constitute a population with a tendency to have health problems, in such a way that it is estimated that 80% of them suffer from at least one chronic disease, this increase in the number of chronic diseases is directly related to the higher functional incapability [40]. It would be valid that the health services that care for elders with leprosy incorporated their routine evaluation scales of functional capacity in this population. In this way, it would be possible to more completely evaluate the impact of the disease on the quality of life of these individuals, facilitating the institution of early rehabilitation.
The limitations of the present study are related to its design as it is about a retrospective cohort performed from the review of medical records, and the quality and veracity of the data entries made by the medical professionals are factors that interfere with the trustworthiness of the analyzed data. In the face of research scarcity about leprosy in the elderly, it is suggested that new prospective studies be made to contribute to greater knowledge about this theme and to the creation of strategies for early diagnosis and disability prevention, seeking to decrease costs in the health system, loss of family relationships and compromises to the autonomy of the elderly.
The temporal analysis of leprosy among the elderly in the state of Pará demonstrated increasing trends for new cases and for the detection rate in the general population and a trend for an elevation in these values in the elderly population for the next ten years.
Regarding the epidemiological and clinical profile, it was verified that there was a predominance of males in the 60 to 69 year-old age group and a predominance of the multibacillary operational classification. Leprosy reactions were highly prevalent, and the first reactional episode occurred most frequently in the first six months of multidrug therapy. Patients with positive smear skin at the time of diagnosis presented higher chances of developing leprosy reactions. However, positive ELISA anti-PGL-1 serology in the diagnosis was not a predictive factor relevant to the occurrence of the reactions.
Prednisone was the most used medicine in the treatment of the reactional episodes. A high proportion of the elders already presented with some physical incapacity at the time of diagnosis. Systemic arterial hypertension and diabetes mellitus were the predominant comorbidities. Therefore, the leprosy amongst the elderly deserves attention because of the increased susceptibility to disability in this age group, with their higher risk of reaction and their greater level of co-morbidity.
|
10.1371/journal.pmed.1002601 | The San Diego 2007 wildfires and Medi-Cal emergency department presentations, inpatient hospitalizations, and outpatient visits: An observational study of smoke exposure periods and a bidirectional case-crossover analysis | The frequency and intensity of wildfires is anticipated to increase as climate change creates longer, warmer, and drier seasons. Particulate matter (PM) from wildfire smoke has been linked to adverse respiratory and possibly cardiovascular outcomes. Children, older adults, and persons with underlying respiratory and cardiovascular conditions are thought to be particularly vulnerable. This study examines the healthcare utilization of Medi-Cal recipients during the fall 2007 San Diego wildfires, which exposed millions of persons to wildfire smoke.
Respiratory and cardiovascular International Classification of Diseases (ICD)-9 codes were identified from Medi-Cal fee-for-service claims for emergency department presentations, inpatient hospitalizations, and outpatient visits. For a respiratory index and a cardiovascular index of key diagnoses and individual diagnoses, we calculated rate ratios (RRs) for the study population and different age groups for 3 consecutive 5-day exposure periods (P1 [October 22–26], P2 [October 27–31], and P3 [November 1–5]) versus pre-fire comparison periods matched on day of week (5-day periods starting 3, 4, 5, 6, 8, and 9 weeks before each exposed period). We used a bidirectional symmetric case-crossover design to examine emergency department presentations with any respiratory diagnosis and asthma specifically, with exposure based on modeled wildfire-derived fine inhalable particles that are 2.5 micrometers and smaller (PM2.5). We used conditional logistic regression to estimate odds ratios (ORs), adjusting for temperature and relative humidity, to assess same-day and moving averages. We also evaluated the United States Environmental Protection Agency (EPA)’s Air Quality Index (AQI) with this conditional logistic regression method. We identified 21,353 inpatient hospitalizations, 25,922 emergency department presentations, and 297,698 outpatient visits between August 16 and December 15, 2007. During P1, total emergency department presentations were no different than the reference periods (1,071 versus 1,062.2; RR 1.01; 95% confidence interval [CI] 0.95–1.08), those for respiratory diagnoses increased by 34% (288 versus 215.3; RR 1.34; 95% CI 1.18–1.52), and those for asthma increased by 112% (58 versus 27.3; RR 2.12; 95% CI 1.57–2.86). Some visit types continued to be elevated in later time frames, e.g., a 72% increase in outpatient visits for acute bronchitis in P2. Among children aged 0–4, emergency department presentations for respiratory diagnoses increased by 70% in P1, and very young children (0–1) experienced a 243% increase for asthma diagnoses. Associated with a 10 μg/m3 increase in PM2.5 (72-hour moving average), we found 1.08 (95% CI 1.04–1.13) times greater odds of an emergency department presentation for asthma. The AQI level “unhealthy for sensitive groups” was associated with significantly elevated odds of an emergency department presentation for respiratory conditions the day following exposure, compared to the AQI level “good” (OR 1.73; 95% CI 1.18–2.53). Study limitations include the use of patient home address to estimate exposures and demographic differences between Medi-Cal beneficiaries and the general population.
Respiratory diagnoses, especially asthma, were elevated during the wildfires in the vulnerable population of Medi-Cal beneficiaries. Wildfire-related healthcare utilization appeared to persist beyond the initial high-exposure period. Increased adverse health events were apparent even at mildly degraded AQI levels. Significant increases in health events, especially for respiratory conditions and among young children, are expected based on projected climate scenarios of wildfire frequency in California and globally.
| Large wildfires are becoming more frequent and are expected to increase with climate change. Smoke from wildfires can cause health problems, especially for children, older persons, and people who already have respiratory or heart problems.
Researchers had access to data on emergency department visits, hospitalizations, and outpatient visits from California’s Medicaid program, Medi-Cal. This allowed for analysis of the effects of wildfire among a particularly vulnerable population, which included a large proportion of young children. It also provided an opportunity to examine changes in outpatient visits.
Researchers were able to look at health problems during the time when the wildfire smoke was most intense and also at later periods to see if people had health problems that may take more time to develop. They chose to study a very large wildfire that happened in San Diego County in 2007.
During the peak fire period, emergency department visits for respiratory conditions increased by 34% and visits for asthma by 112%. There was no change in visits for heart-related problems.
Some healthcare visit types remained high even after the peak fire period. For example, outpatient visits for acute bronchitis were 72% above the usual rate in the 5-day period following the peak fire period.
Young children had bigger increases in visits during the peak fire period than older age groups. Children aged 0–4 had a 136% increase in emergency department visits for asthma, and very young children aged 0–1 experienced a 243% increase.
Researchers studied how health visits changed on days with more intense smoke using data from smoke models. Emergency department visits for asthma went up 73% on days following an air quality day designated as “unhealthy for sensitive populations,” based on wildfire smoke and using the United States Environmental Protection Agency (EPA)'s Air Quality Index (AQI) air pollution levels as a guide.
We expect increases in respiratory problems during wildfires, possibly even at mildly degraded levels of air quality. People may continue to seek care for some persisting conditions.
Young children appear at highest risk for respiratory problems during a wildfire, which is cause for particular concern because of the potential for long-term harm to children’s lung development.
The risk of future wildfires on the health of Californians will continue to be shaped by global climate change, as well as the anticipated growth of vulnerable subpopulations. Planning to protect the health of vulnerable populations is important.
| Large forest fires have become more frequent in the Western United States since the 1980s [1–3]. Under most future climate scenarios, the frequency and size of wildfires in the southwestern states are expected to increase [4]. Climate models predict up to a 74% increase in area burned in California and a possible doubling of wildfire emissions by the end of the century [5]. Wildfires release large amounts of particulate matter (PM) and other toxic substances into the air, including carbon dioxide, carbon monoxide, and methane [6–7]. In the coterminous US, yearly emissions of fine PM from wildfire smoke are estimated to be between 118,000 and 986,000 metric tons and carbon dioxide emissions between 24 and 134 million metric tons, in addition to other compounds and gases [6]. In 2012, wildfires contributed 20% of the fine particulate emissions in the US [8].
Smoke from fires can be transported to affect populations far downwind [9]. Projected trends in climate change show that, globally, the number of people who will experience adverse health effects from wildfires is increasing [10–12]. The number of persons who are vulnerable is also expanding because more people live near wildlands [13].
Wildfire smoke exposures have been associated with adverse health outcomes, including premature death and increased inpatient hospitalizations and emergency department presentations [14–16]. Smoke from wildfires produces inhalable particles that are 10 micrometers and smaller (PM10) and fine inhalable particles that are 2.5 micrometers and smaller (PM2.5). PM10 and PM2.5 have consistently been linked to respiratory outcomes, particularly asthma exacerbations [15–17] and in some studies, cardiovascular outcomes [17–20]. Relatively few studies of wildfire smoke have examined the health effects on vulnerable populations. However, the nature and intensity of health impacts are expected to depend on characteristics of the receptor population [16,17,21]. Research on vulnerability to ambient air pollution has identified subpopulations with increased susceptibility to the effects of PM; these include persons with chronic diseases [22], as well as older adults, children, and possibly those with lower education, income, and employment status [23]. Although PM of wildfire origin differs from ambient air pollution in composition and exposure patterns, current research suggests that elderly and young populations will also be especially vulnerable to wildfire-derived PM [16,17,24]. Children warrant particular concern because their lungs are still developing, and exposure to ambient air pollution has been shown to permanently impair lung function [25].
Individual socioeconomic position or status (SES) factors such as personal income and education are accompanied by a broad range of factors that influence health, including prevalent comorbid conditions such as respiratory and cardiovascular diseases, as well as access to healthcare, social stress, and environmental quality of the community [26]. Often, these factors are difficult to isolate.
California’s Medicaid program, Medi-Cal, is a public health insurance program covering health services for low-income individuals, including seniors, persons with disabilities, families with children, children in foster care, pregnant women, and childless adults with incomes below 138% of the federal poverty level. These eligibility criteria create a population that tends to be focused on low-income women and children, plus others with varying disabilities. Beginning at age 65, Medicare is available regardless of income, so for this group, Medi-Cal only pays secondarily or for certain services not covered by Medicare.
In this study, we investigated change in healthcare utilization—including differential health responses by age groups and type of health service—related to wildfire smoke exposure from a large complex of fires in San Diego County in 2007 within a vulnerable population, Medi-Cal beneficiaries who resided in San Diego County at the time.
In late October of 2007, a complex of fires burned nearly 1 million acres in San Diego county, resulting in the evacuation of an estimated 515,000 county residents and numerous road, school, and business closures [27]. San Diego county had a population of 3,095,342 according to the 2010 US Census [28], with the population concentrated along the coastal areas.
Medi-Cal beneficiaries numbered 345,257 in San Diego County in July 2007 [29]. Medi-Cal administrative claims data were obtained from the California Department of Health Care Services’ (DHCS) Management Information System/Decision Support System (MIS/DSS) data warehouse for San Diego County for the period of August 1 through December 31, 2007 to accommodate reference dates surrounding the late-October fire period.
We conducted 2 types of analyses. The first was a county-wide analysis of Medi-Cal claims data, which compared rates for emergency department presentations, inpatient hospitalizations, and outpatient visits during the fires with reference periods. The second was a case-crossover analysis that examined exposures by residential zip code and emergency department presentations with respiratory diagnoses.
For the county-wide analysis, we identified October 22–26 as the peak fire-exposure period (P1) based on a previous study that analyzed this fire using data from the BioSense Platform, an integrated national syndromic surveillance system [30]. We defined 2 following periods, P2 (October 27–31) and P3 (November 1–5), for analysis in order to identify any health outcomes that might be sensitive to cumulative or lagged exposure to wildfire smoke.
For the case-crossover analyses of exposure to varying concentrations of PM2.5, the population was limited to those beneficiaries with a valid San Diego County zip code listed for their residential address. Where possible, post office-box–only zip codes were mapped to real-address zip codes in the same subregion, municipality, and neighborhood. Exposures were based on the modeled PM2.5 for these 101 real-address zip codes.
Wildfire PM2.5 concentrations were estimated through the use of coupled models of wildfire smoke emissions and atmospheric dispersion [31]. Spatially and temporally resolved estimates of wildland fire emissions were computed using the geospatial tool Wildland Fire Emissions Information System (WFEIS); model outputs were then introduced into the meteorological atmospheric transport model Hybrid Single-Particle Lagrangian Integrated Trajectories (HYSPLIT) to produce PM concentration estimates computed to a 0.01-degree grid (approximately 1 km2) on an hourly basis. Hourly model outputs were used to estimate daily average wildfire PM2.5 concentrations (μg/m3) by zip code, as described previously [31]. All analyses in this study are based on PM originating from wildfire sources, so all PM in this manuscript refers to wildfire-only PM. We interpolated relative humidity and temperature data from a Remote Automated Weather Station database to county subregional areas for the period of August to November 2007 (environmental data availability period).
Medi-Cal dataset variables included county of residence and home zip code of the patient, date of the medical visit, general type of service provided, where the visit occurred, classification of the provider (i.e., hospital, emergency department, outpatient, excluding claims related to nursing homes, etc.), and diagnosis that was being treated (by International Classification of Diseases [ICD]-9 code, up to 2 diagnoses per claim). Patient demographic variables included sex and age. A unique, de-identified beneficiary code (beneficiary ID) was provided with the dataset; names were not included. Eligible subjects were San Diego County residents who had a qualifying Medi-Cal fee-for-service claim during the study period. Qualifying claims included those for inpatient hospitalizations, emergency department presentations, and outpatient visits (clinic and physician office visits). The DHCS Data and Research Committee and California’s Health and Human Services Agency’s Committee for the Protection of Human Subjects approved the study protocol. We performed data management and analysis using SAS version 9.4 (SAS Institute; https://www.sas.com/en_us/home.html) and Excel for Mac version 14.4.3 (Microsoft, https://www.microsoft.com/en-us/).
The beneficiary ID linked all claims records for each beneficiary. Beneficiaries aged 65 and above were excluded from the study because claims for these beneficiaries were not adequately represented in the Medi-Cal data due to their dual eligibility for Medicare and Medi-Cal.
Episodes of care (“encounters”) were identified from the subset of records with at least one valid diagnosis code. For each beneficiary, inpatient status was assessed for each day from August 1 through December 31, 2007. Inpatient hospitalizations were identified as periods of one or more contiguous days with associated inpatient claims records; the start date of the earliest record was used as the admission date. Emergency department claims records for each beneficiary from the same date were grouped together into a single episode of care. Overnight emergency department presentations were identified, and records from both those dates were grouped into a single episode of care. Physician office and clinic claims records for each beneficiary from the same date were grouped together into a single episode of care, referred to hereafter as outpatient visits. To reduce misclassification of inpatient diagnosis, errors in ascertainment of inpatient status, and errors in date of inpatient admission, the episodes-of-care dataset was limited to episodes with admission during the period of August 16 to December 15, 2007 (encounter data availability period).
Episodes of care were identified as being related to the outcomes of interest based on the primary and secondary diagnoses from any associated claims records, except inpatient hospitalizations, which were limited to claims records from the first 14 days of the hospitalization. Encounters for components of a respiratory index and a cardiovascular index were identified as outcomes for analysis, based on ICD-9 coding in a previous study of a large wildfire event in California (Table 1) [32]. The respiratory index included asthma, acute bronchitis, chronic obstructive pulmonary disease (COPD), bronchitis—not otherwise specified, pneumonia, upper respiratory infections, cystic fibrosis, bronchiectasis, extrinsic allergic alveolitis, respiratory symptoms, and other acute and subacute respiratory conditions caused by exposure to fumes, vapors, or external agents. The cardiovascular index included ischemic heart disease, dysrhythmia, congestive heart failure, cerebrovascular disease including stroke, and peripheral vascular disease. We also examined total visits (all-cause) for each healthcare setting to provide context for results for the outcomes of interest.
During the health data availability period of August 1 to December 31, 2007, there were a total of 5,454,360 Medi-Cal claims for San Diego beneficiaries, derived from 217,067 residents with at least one claim of any type (not limited to the claim types we examined). We excluded 40,216 residents aged 65 and above. After these exclusions, during the fire period of October 22–26, 2007, there were 26,556 San Diego County residents with at least one Medi-Cal claim (15.0% of beneficiaries). The individuals with at least one claim during the health data availability period and fire period are described by age, sex, and race/ethnicity (Table 2).
Among our study population and during the period of August 16 to December 15, 2007, we identified 25,000 emergency department presentations, 17,009 inpatient hospitalizations, and 269,842 outpatient visits. Young children aged 0–4 comprised 14.4% of inpatient hospitalizations, 15.1% of emergency department presentations, and 28.8% of outpatient visits. Very young children (aged 0–1) accounted for 12.8% of inpatient hospitalizations, 10.8% of emergency department presentations, and 15.8% of outpatient visits.
Wildfire-derived PM2.5 concentrations are shown in Table 3. During the most intense initial period of the firestorm P1, the mean of the 24-hour average PM2.5 concentrations of all the zip codes was 89.1 μg/m3. The highest of all the zip codes’ daily averages occurred during this window of time, 803.1 μg/m3. In comparison, the US EPA 24-hour air quality standard for PM2.5 is 35 μg/m3, and concentrations over 250 μg/m3 correspond to AQI level “hazardous.”
Estimated average daily wildfire PM2.5 concentrations by zip code through the course of the fire period are shown in Fig 1. Concentrations spiked sharply on October 22 and continued through the initial 5-day fire period, then declined. The mean PM2.5 concentration on the first day of the 5-day fire period was 160 μg/m3 (AQI “very unhealthy”), which then dropped to 29.9 μg/m3 on the 5th day (AQI “moderate”). The fire boundaries and daily average PM2.5 concentrations by zip code in San Diego County are mapped for the 5-day exposure period (P1) (Fig 2).
In multivariate models adjusted for daily temperature and relative humidity, an increase in the average PM2.5 of 10 μg/m3 for the daily, 48-hour moving, and 72-hour moving averages was associated with a 3%, 5%, and 8% increase, respectively, in the likelihood for asthma emergency department presentations, with similar but attenuated increases for respiratory visits (Table 4). ORs were greater when examining moving averages over several days, suggesting that the models were capturing cumulative and lagged effects. Square terms did not reach significance in any of the models, so linear models were selected. We did not find effect modification by age, including after stratifying by sex.
Unhealthy AQI levels were associated with increased respiratory conditions in emergency department presentations, adjusting for temperature and relative humidity (Table 5). The AQI models fit best with a 1-day lag compared to same-day– or 2-day–lagged models. The AQI levels “unhealthy for sensitive groups” (OR 1.73; 95% CI 1.18–2.53) and “unhealthy” (OR 1.79; 95% CI 1.30–2.23) both were associated with significantly elevated odds of an emergency presentation the day after exposure versus the AQI level “good.” The strongest effect was seen in the same-day model for the highest exposure category, hazardous (OR 2.41; 95% CI 1.39–4.18).
By examining multiple respiratory and cardiovascular endpoints across 3 healthcare settings and 3 exposure periods as well as for different age groups, we have compiled a relatively comprehensive view of health events during this significant wildfire complex. While outcomes such as respiratory conditions were clearly elevated, visits for other outcomes were decreased. These observed results must be viewed in the context of the extensive nature of the fire and the resulting evacuations and other disruptions. These unusual conditions likely altered healthcare-seeking behavior; residents may not have accessed healthcare other than for the most urgent conditions. A review of the relationship between the 2007 wildfires and the emergency department of the University of California, San Diego hospital found a 5.8% decrease in admissions during the fires, although the rate of patients with a chief complaint of shortness of breath increased significantly and the rate of patients who left without being seen nearly doubled [37]. Also, an assessment of the 2003 fires in San Diego noted that emergency department presentations initially declined during the fire period, corresponding to days when authorities recommended that students and employees stay home [38].
Our study examined Medi-Cal beneficiaries, a group representing a vulnerable, although fairly substantial, subset of the general population. We would anticipate their response to the health stressor of wildfire smoke to be similar in nature to the general public but possibly increased in magnitude. Asthma, as in other wildfire studies, appeared to be the most sensitive to wildfire smoke exposure [16]. Our findings support a wildfire smoke association with the infectious respiratory outcomes pneumonia, bronchitis, and upper respiratory infections despite inconsistent results from previous studies [16,39]. Airway injury from wildfire smoke exposure could predispose bacterial pneumonia. Previous wildfire studies generally have found positive associations with COPD [16]. Because COPD is a condition more prevalent in the older population, who were excluded from our analysis, this may have limited our ability to study this condition.
Similar to COPD, cardiovascular outcomes are generally more prevalent in older adults, so the absence of this population from our study is relevant here as well. However, our study is not unusual in its null cardiovascular findings for wildfire smoke exposures, despite the scientific relationship between general particulate air pollution and cardiovascular disease [40]. The reasons for this are unclear. The lower prevalence of cardiovascular events in general in comparison with respiratory conditions—along with the possibility that cardiovascular impacts from wildfire smoke may occur at a smaller magnitude than respiratory impacts—may require a larger study to detect an excess. Another factor may be that only certain diagnoses are elevated, and broadly combining all cardiovascular conditions may obscure an association. Moreover, persons with underlying cardiovascular disease may be seen for respiratory rather than cardiovascular conditions (competing diagnoses) during wildfires. Too few studies have examined specific cardiovascular outcomes to have a clear picture of which are related to wildfire exposure [15], although a recent analysis of an extensive California wildfire season provided strong evidence for increased cardiovascular risk [20].
Using sequential exposure periods during and after the peak smoke exposure allowed examination of changes over longer time frames. Studies typically do not detect any increases beyond 3 to 5 lag days. This design allowed us to show some conditions persisting over longer periods of time. Cumulative exposure may be relevant for conditions such as asthma, bronchitis, or pneumonia, which may gradually develop or worsen over time. Inhaled PM may prompt inflammation and alter immune functions, increasing susceptibility to respiratory infections. Also, patients may not seek care until their symptoms become severe.
Our examination of outpatient visits was an exception to the majority of wildfire research studies in the US, which have largely relied on inpatient hospitalization and emergency department data [15]. We noted that patients continued to seek care in outpatient settings while the initial surge in emergency department presentations was declining.
The AQI is a widely used public health tool, yet few wildfire studies have made associations with the AQI categories. The sensitivity of our study population was revealed in its response to even modestly increased concentrations of PM, as excess adverse health events began to occur at an AQI level designed to represent the first threshold at which susceptible persons are advised to consider limiting their exposure. These results provide evidence for the value of the AQI as a communication tool in conveying health risks of wildfire smoke to the public, especially because the AQI addresses the immediate day, and health events were shown to generally rise with increasing same-day AQI exposure categories.
While children are thought to be more vulnerable to effects of wildfire smoke, the literature has not been conclusive [16]. The mixed results for children may be due to different effects between very young children and older children because null results are often seen in studies that combine all ages or do not include very young children. Wildfire smoke effects among children aged 6 to 18 have been noted in a cohort study of schoolchildren who experienced increased respiratory symptoms [41]. Children’s heightened susceptibility to wildfire smoke may be related to their smaller airway size [42]. In our study, this vulnerability was most evident among the very youngest children, aged 0–1, for whom the increase in emergency department presentations during the initial wildfire period (243% increase in asthma) was the highest of any group we evaluated.
Several studies that have stratified on very young children have shown significant associations between increased respiratory admissions and/or visits and wildfire smoke exposures [32,43,44]. However, the magnitude of the association in our Medi-Cal population appears to be greater than what has been found previously in general populations, although results are not directly comparable because methods differ between studies. A study examining 0- to 4-year-olds found a potential 5% increase in the odds of physician visits for asthma, for a 60 μg/m3 increase in PM10 [41]. Our findings of 236%, 267%, and 131% increases in asthma emergency department presentations, inpatient hospitalizations, and outpatient visits, respectively, suggest a particularly high association among young children (0–4 years). This may be related to underlying vulnerability of the Medi-Cal population. Many factors may contribute to vulnerability, e.g., one study identified increased asthma risks only among children with asthma and obesity [45]. Overall, the very young in our study experienced significantly elevated risks of unusually high magnitude.
The few studies that have examined underlying population vulnerability have tended to use community level analyses that found that various measures of lower SES will confer greater risk from wildfire smoke [15,19,31,46,47]. Although a Canadian study did not, this null finding may be related to Canada’s more comprehensive healthcare system [48]. Several studies only detected wildfire health effects in a subgroup with both health and SES vulnerabilities—the indigenous population in Australia—as parallel analyses with the general population failed to detect an effect [48,49]. An analysis of the same San Diego wildfire using Kaiser Permanente health plan members appeared to have possibly lower increases in emergency room visits than our findings, although the analyses are not directly comparable [50]. Our study population of Medi-Cal beneficiaries would encompass multiple susceptibility factors, which may manifest during disasters in ways beyond those directly related to baseline health, e.g., having fewer resources to evacuate, less effective home air filtration, or less control over work schedules.
A limitation of this analysis is that, because Medi-Cal data was used, the study population is not representative of the general population. At the same time, some of the populations most vulnerable to the health effects of wildfires are well-represented among Medi-Cal beneficiaries. For example, over 50% of the state’s aged 0–4 population is covered by Medi-Cal [51]. Children are generally more vulnerable to air pollution due to their higher ventilation rate and other factors [52]. A further limitation may be our use of only fee-for-service claims. In 2007, 48% of San Diego Medi-Cal beneficiaries were in managed care [29], and we have no information on differences between the fee-for-service and managed-care populations that could affect our findings. Medi-Cal data only included a primary and secondary diagnosis code, so any condition not occurring within the first 2 codes would not be identified. There is always a possibility of misclassification in the diagnosis codes or missing data on utilization; however, this should be limited by using medical claims data that are required to be submitted for payment. In addition, the relatively short time frame of this study should reduce any limitations that are a result of changing Medi-Cal eligibility over time.
Our wildfire smoke models allowed geospatially and temporally resolved outputs of particulate concentrations. However, our analysis was based on patient residential zip code, so exposure misclassification would occur because people change location during the day. Wildfire-related disruptions could also have prevented people from seeking care or have caused diversion to facilities outside the area, which would bias our results toward the null. Still, because of the widespread nature of the smoke across much of the populous area of the county, the use of exposure periods defined by sets of wildfire dates appeared to perform relatively well in capturing a broad population risk.
As the population ages and the prevalence of comorbidities increase, the number of persons who are susceptible to wildfire exposures will also grow. Nationally, the proportion of the population over age 65 is anticipated to grow from 15% to 24% by 2060 [53]. Increasing prevalence of diabetes and obesity in the US [54] will also impact cardiovascular health. Unless these trends are reversed, the growing older population will also be less healthy, leading to a greater segment of the population vulnerable to PM from wildfires.
Our study of Medi-Cal beneficiaries identified a significant increase in adverse respiratory events from wildfire smoke exposure and suggested that health risk may persist beyond several immediate days of high–PM exposure. Our findings contribute to growing evidence that, in addition to acute respiratory events such as asthma exacerbation, exposure to wildfire PM may predict infectious conditions, including upper respiratory infections, bronchitis, and pneumonia, which may take longer to manifest. The substantial risk noted among the youngest children is cause for concern because of the potential for long-term harm to children’s lung development. The vulnerability of our study population was also shown in its sensitive response to deteriorating air quality because excess adverse health events began to occur at mildly degraded levels of air quality.
The risk of future wildfires to the health of Californians will continue to be shaped by global climate change, as well as the characteristics and anticipated growth of vulnerable subpopulations. The recognition that climate change will increase the burden most severely on disadvantaged communities creates the imperative for public health to help prepare and protect these vulnerable populations.
|
10.1371/journal.pcbi.1005348 | THE REAL McCOIL: A method for the concurrent estimation of the complexity of infection and SNP allele frequency for malaria parasites | As many malaria-endemic countries move towards elimination of Plasmodium falciparum, the most virulent human malaria parasite, effective tools for monitoring malaria epidemiology are urgent priorities. P. falciparum population genetic approaches offer promising tools for understanding transmission and spread of the disease, but a high prevalence of multi-clone or polygenomic infections can render estimation of even the most basic parameters, such as allele frequencies, challenging. A previous method, COIL, was developed to estimate complexity of infection (COI) from single nucleotide polymorphism (SNP) data, but relies on monogenomic infections to estimate allele frequencies or requires external allele frequency data which may not available. Estimates limited to monogenomic infections may not be representative, however, and when the average COI is high, they can be difficult or impossible to obtain. Therefore, we developed THE REAL McCOIL, Turning HEterozygous SNP data into Robust Estimates of ALelle frequency, via Markov chain Monte Carlo, and Complexity Of Infection using Likelihood, to incorporate polygenomic samples and simultaneously estimate allele frequency and COI. This approach was tested via simulations then applied to SNP data from cross-sectional surveys performed in three Ugandan sites with varying malaria transmission. We show that THE REAL McCOIL consistently outperforms COIL on simulated data, particularly when most infections are polygenomic. Using field data we show that, unlike with COIL, we can distinguish epidemiologically relevant differences in COI between and within these sites. Surprisingly, for example, we estimated high average COI in a peri-urban subregion with lower transmission intensity, suggesting that many of these cases were imported from surrounding regions with higher transmission intensity. THE REAL McCOIL therefore provides a robust tool for understanding the molecular epidemiology of malaria across transmission settings.
| Monitoring malaria epidemiology is critical for evaluating the impact of interventions and designing strategies for control and elimination. Population genetics has been used to inform malaria epidemiology, but it is limited by the fact that a fundamental metric needed for most analyses—the frequency of alleles in a population—is difficult to estimate from blood samples containing more than one genetically distinct parasite (polygenomic infections). A widely used approach has been to restrict analysis to monogenomic infections, which may represent a biased subset and potentially ignores a large amount of data. Therefore, we developed a new analytical approach that uses data from all infections to simultaneously estimate allele frequency and the number of distinct parasites within each infection. The method, called THE REAL McCOIL, was evaluated using simulations and was then applied to data from cross-sectional surveys performed in three regions of Uganda. Simulations demonstrated accurate performance, and analyses of samples from Uganda using THE REAL McCOIL revealed epidemiologically relevant differences within and between the three regions that previous methods could not. THE REAL McCOIL thus facilitates population genetic analysis when there are polygenomic infections, which are common in many malaria endemic areas.
| Malaria has declined significantly over the past decade, but continues to cause half a million deaths annually [1]. Calls for elimination have shifted research efforts towards developing new approaches for transmission reduction, including the identification of source and sink regions and hotspots that sustain transmission [2–4]. Plasmodium falciparum population genetic tools are increasingly being used to inform these efforts [5–12] and have been proposed as a means to establish the direction of parasite flows and to determine elimination status both by identifying the source of imported infections and by establishing that no local transmission is occurring [13–17]. However, in malaria-endemic regions, infections are frequently characterized by multiple different genotypes (polygenomic infections), which makes interpreting genetic data challenging. As a result, population genetic analyses of malaria parasites have often been limited to monogenomic infections, greatly reducing the utility of available data and potentially introducing biases into results.
Rapid technological developments have led to a proliferation of approaches for characterizing malaria parasite genomes, each with different implications for cost, suitability for field samples across a range of transmission settings, and applicability to different research questions [5,18–23]. Many genotyping approaches are based on a small number of single nucleotide polymorphisms (SNPs). SNP data are cheap and straightforward to obtain from commonly used dried blood spot (DBS) samples, collected in a variety of field settings, and remain the most common approach for genotyping studies. However, a high prevalence of polygenomic infections can render estimation of even the most basic parameters from SNP data, such as population allele frequencies, difficult.
Population allele frequencies are usually estimated from monogenomic infections [6,7,24], because of the challenge of estimating the true proportion of each lineage from heterozygous SNP loci resulting from high complexity of infection (COI, the number of clones in an individual). However, constraining data sets to only monogenomic infections may introduce systematic biases because these infections may not be representative. Such constraint also greatly limits the precision of estimates when the majority of samples are polygenomic. It is common to use the proportion of heterozygous calls in each individual or the fraction of polygenomic infections to compare genetic diversity between populations [6,7,16,25–27]. However, the complexity of infection underlying polygenomic infections can vary dramatically, and the probability of a particular locus being heterozygous will depend on its allele frequency in the population. COIL (estimating COI using likelihood), was recently developed to provide a more quantitative measure of genetic diversity [28], but unless supplied with external allele frequency data, relies on monogenomic infections to estimate allele frequencies and is therefore problematic when a large fraction of infections are polygenomic. While external allele frequency data can be obtained from parasite population genomic data such as the Pf3K project (http://www.malariagen.net/projects/pf3k), these estimates are only available in specific locations, and may exhibit considerable heterogeneity in space and time.
Here we introduce a new Bayesian approach, Turning HEterozygous SNP data into Robust Estimates of ALelle frequency, via Markov chain Monte Carlo, and Complexity Of Infection using Likelihood (THE REAL McCOIL), to additionally incorporate polygenomic samples, using Markov chain Monte Carlo methods to simultaneously estimate allele frequency and COI. We tested two versions of our method on a series of simulations and then applied it to data on 105 SNP loci in 868 samples from cross-sectional surveys in three regions of varying endemicity in Uganda [29–31]. The allele frequencies estimated by our new approach were used to calculate FST [32], a measure of genetic differentiation between sites, and FWS [33], a measure of the within-host genetic diversity. These results demonstrate the utility of THE REAL McCOIL to obtain accurate estimates of COI and allele frequency from SNP data, which can be used to characterize genetic diversity and perform population genetic analyses of parasite populations even in very high transmission settings.
The cross sectional survey was approved by IRBs at the University of California, San Francisco (#11–07138) and SOMREC at Makerere University, Uganda (#2011–203).
We developed a Markov chain Monte Carlo (MCMC) method to simultaneously estimate population allele frequency for each SNP and COI for each individual. Since estimating COI and allele frequencies are highly related to each other, our approach explored the uncertainty of both at the same time, and by doing so, incorporated information from polygenomic infections. Assuming there are n individuals and k loci, the parameters to be estimated include complexity of infection for each individual (M = [m1, m2, …, mn]) and population allele frequency for each locus (P = [p1, p2, …, pk]). We used the data in two ways: a categorical method, in which we considered SNP at locus j of individual i, Bij, to be heterozygous or homozygous (0 [homozygous minor allele], 0.5 [heterozygous], 1 [homozygous major allele]), and a proportional method, in which the proportion of major allele at locus j of individual i, Sij, was calculated from the relative signal intensity for each allele (Sij=A1ijA1ij+A2ij, where A1 and A2 represent the signal intensity of major and minor allele that are obtained from Sequenom or similar types of SNP assays, respectively [34]). The notations are summarized in Table A in S1 File. Similar to COIL, THE REAL McCOIL assumed that different loci are independent, that different samples are independent and polygenomic infections are obtained from multiple independent infections, and that the samples were collected from a single homogeneous population.
We sampled COI of each individual from a zero-truncated Poisson distribution with mean m¯, and population allele frequency of each locus from a uniform distribution U(0, 1). For each individual, we independently sampled allele(s) for each locus from Bernoulli (pj). We determined the relative proportion of different lineages within the host by sampling the proportion of each infection from a uniform distribution U(0, 1). For comparison, we additionally tried sampling from a truncated exponential distribution with the rate λ = 1. After obtaining within-host allele frequency (STij), we drew SOij from a normal distribution with mean = STij and variance σ2=εI, where ε represents the level of measurement error. We sampled the intensity of the signal I for each locus of each individual from the sum of a Poisson distribution with average I¯=8 and a normal distribution with mean = 0 and variance = 0.25. Simulations were designed to represent the type of raw data obtained from Sequenom or similar types of SNP assays, where an intensity value is obtained for each potential allele [34]. If the intensity of signal was smaller than Imin, we assumed the data were missing. We obtained the intensities of two alleles, A1 and A2, by A1=ISOSO2+(1−SO)2 and A2=I1−SOSO2+(1−SO)2, and determined heterozygous calls or homozygous calls by the relative intensity of signals of two alleles, which was characterized by arctan(A1A2), the angle in polar coordinate system. The SNP was called as heterozygous if arctan(A1A2) was within (d1, d2) and homozygous otherwise (Fig B in S1 File). For simulated data with measurement error ε >0, we used (d1, d2) = (5, 85). For real data, (d1, d2) was determined by expert review of each locus as described below.
We compared the performance of the categorical and proportional versions of our method to COIL, assessing the difference in parameter estimates and variation. We simulated violations of the model assumptions, specifically independence among loci, independence among parasite lineages within the same host, and a single, homogeneous population. Dependence among loci was simulated by different proportions of loci (p) that were linked. We simulated relatedness (r) among lineages within the same host by sampling alleles either from an existing lineage within the same host (with probability r) or from the population (with probability (1-r)). We simulated two equally sized subpopulations with either the same or different average COI and with various levels of difference in allele frequencies and treated them as one single population to test the robustness of the assumption that the population was well-mixed. We also simulated missing data and populations with COI up to 20.
Dried blood spot samples were obtained from representative cross-sectional surveys performed in 2012 and 2013 as part of the East African International Centers of Excellence in Malaria Research (ICEMR) program. Surveys were performed in each of three sub-counties in Uganda: Nagongera in Tororo District, Kihihi in Kanungu District, and Walukuba in Jinja District. Details of these surveys, along with entomological and cohort data from the same sites have been published [29,31,36,37]. In brief, 200 households from each sub-county were randomly selected from a census population, and all children and an age-stratified sample of adults were enrolled from each household. All samples taken from individuals with evidence of asexual parasitemia by microscopy were selected for Sequenom SNP genotyping, and an age-stratified subset were also selected for merozoite surface protein 2 (msp2) genotyping. The Sequenom assay consisted of 128 SNPs selected to be polymorphic and at intermediate/high frequency in multiple popluations (https://www.malariagen.net/projects/p-falciparum-community-project). After removing variants with elevated missing rate, we retained 105 SNPs (see S1 Table for SNP data) and three of them are in known drug resistance loci. Samples were genotyped according to the relative intensity of the two alleles, as previously described [21]. Genotyping of msp2 was performed with alleles sized by capillary electrophoresis, as previously described [38]. The number of unique alleles were called by a single, expert reader, with allele counts > 5 grouped into a single category due to difficulties in accurately distinguishing artifacts from true alleles at high complexities of infection.
After excluding samples with more than 25% missing SNP data and loci with more than 20% missing data from the analysis, the numbers of individuals included were 462 (71%) [Nagongera], 48 (51%) [Walukuba], and 74 (59%) [Kihihi], and the numbers of loci were 63 (60%) [Nagongera], 49 (47%) [Walukuba], and 52 (50%) [Kihihi]. After these cutoffs, only the analysis of Nagongera included one drug resistance locus, and others included none. We used a permutation test with N = 10,000 to compare estimated COI between groups because there were many ties. In the analysis, we assumed that error rates e1 and e2 were both 0.05 and εest = 0.02. FWS was calculated by (1−HW/HS), where HW and HS are 2pW(1−pW) and 2pS(1−pS) respectively and pw and ps are within-host allele frequency and population allele frequency respectively [33]. The HW/HS ratio was estimated by performing linear regression between HW and HS with fixed intercept = 0.
We simulated data of 100 SNPs from populations with an average COI of 3, 5 and 7 and sample size of 100, and compared estimates of COI and allele frequencies using COIL and THE REAL McCOIL. When average COI was 3, all three methods estimated COI well, although allele frequency estimates from COIL were less precise than THE REAL McCOIL (mean absolute deviation [MAD] = 0.077 [COIL], 0.019 [THE REAL McCOIL categorical], 0.019 [THE REAL McCOIL proportional], Mann-Whitney test p-value < 2×10−16) (Fig 1). When average COI was 5, however, COIL did not estimate COI or allele frequencies accurately (MAD = 1.45 [COI] and 0.15 [allele frequency]), and when COI was 7, it was unable to estimate allele frequencies due to a lack of monogenomic infections. In contrast to COIL, which consistently underestimated or failed to estimate COI in populations with greater numbers of polygenomic infections, THE REAL McCOIL estimated both COI and allele frequencies well even when COI was high (for categorical and proportional methods, respectively: COI = 5, MAD = 0.61, 0.45 [COI] and 0.024, 0.019 [allele frequency]; COI = 7, MAD = 0.86, 0.79 [COI] and 0.025, 0.015 [allele frequency]). Thus, the ability of THE REAL McCOIL to jointly estimate allele frequencies and COI from all available data resulted in considerably improved performance in estimates of both quantities, especially when the average COI was high.
Furthermore, we compared the performance of the categorical and proportional methods when we included measurement error in simulations of observed within-host allele frequency. The categorical method modeled measurement error by incorporating the probability of calling homozygous loci heterozygous (e1) and vice versa (e2) in the likelihood equation, and the proportional method modeled measurement error by assuming that the difference between true and observed within-host allele frequencies decreased with the intensity of the signals, and was proportional to the error parameter (εest). Fig C (A)(C) in S1 File shows that measurement error resulted in a systematic bias in estimates of COI. However, this bias was relatively minor and fairly robust to misspecification of measurement error, especially when the proportional method was used. In addition, allele frequencies were accurately estimated by both methods (Fig C (B)(E) in S1 File). If parameters for measurement error were not specified, THE REAL McCOIL fit them as part of the MCMC. Fig C (D)(F) in S1 File shows that the probability that the 95% credible interval contained the true COI when error parameters were fitted was higher than those when error parameters were greatly mis-specified.
We next simulated specific violations of the model assumptions to test the robustness of our approach. In particular, we examined the impact of linkage disequilibrium between loci, genetic relatedness of parasites within an individual host, and relatedness between subsets of individuals within the overall population (population substructure). When a proportion of loci (p) were completely linked, COI was slightly overestimated (Fig D in S1 File). When different lineages in the same host were not independent, COI was underestimated and the level of underestimation of COI increased with the level of relatedness (r) (Fig E in S1 File). When we treated two subpopulations as one population, COI was underestimated and the difference between true and estimated COI increased with the difference in the average of COI and the difference in allele frequencies between two subpopulations (Fig F in S1 File). Of these three violations of model assumptions, only a high degree of relatedness between parasites within an individual host resulted in substantial bias in estimates of COI, and none substantially affected estimates of population allele frequencies. Genotyping of real samples often results in missing data; both methods performed well even when 50% of the data were missing (Fig G in S1 File). Furthermore, we tested how the number of loci influences the performance of estimating COI. While the probability that 95% credible interval contained the true COI did not change with the number of loci, the average difference between true and estimated COI decreased (Fig H in S1 File). THE REAL McCOIL provided unbiased estimates even when COI was very high (e.g. 15–20), despite the uncertainty of the estimates increasing with true COI (Fig I in S1 File).
We next applied THE REAL McCOIL to data on 105 SNPs generated from smear positive individuals identified in cross-sectional surveys in three regions of Uganda [36,37] and compared results obtained from THE REAL McCOIL to those using COIL. Both categorical and proportional methods were applied and showed consistent results; for simplicity we therefore present only results from the categorical method.
Nagongera, Kihihi, and Walukuba have been shown to have transmission intensities varying by approximately 100 fold, with entomological inoculation rates recently measured at 310, 32, and 2.8 infectious bites per person year, respectively [29]. Using COIL, the estimated COI was relatively low, with little difference between the 3 sites (median COI = 2 [Nagongera], 2 [Walukuba], and 1.5 [Kihihi]) (Fig 2A). In contrast, results from THE REAL McCOIL show that the COI in Nagongera and Walukuba were similar, and much higher than that in Kihihi (median COI = 5 [Nagongera], 4.5 [Walukuba], and 1 [Kihihi])(Fig 2A, Table B in S1 File and S2 Table). These differences between sites were not captured by COIL because of its dependence on monogenomic infections to obtain estimates of allele frequencies, which were rare in these individuals. We also compared our results to COI estimated using another standard method, msp2 typing, which was performed on a subset of the samples (Fig J in S1 File). Unlike THE REAL McCOIL, however, msp2 typing estimated similar COI in Walukuba and Kihihi (p-value = 0.49) (Fig 2A). msp2 encodes an antigen that elicits strong antibody responses, and this discrepancy may be due to complex population structure arising from immune selection. The difference may also result from the resolution of msp2 typing, which is constrained to COI ≤ 5 [39], or the fact that it is a single marker, rather than a collection of genome-wide markers.
The high COI observed in the lowest transmission site of Walukuba was unexpected but reflected clear differences in the proportion of heterozygous calls, which was similar between Nagongera and Walukuba and lower in Kihihi (Fig K in S1 File). The distributions of age and parasite density were similar between the sites, and thus unlikely to explain these differences (Fig L and Fig M in S1 File). We calculated FWS, an inverse measure of outcrossing [33,40], and found that it was significantly negatively associated with our COI estimates (Fig 3; Pearson’s correlation test between log(COI) and FWS, ρ = −0.93 [Nagongera], −0.94 [Walukuba], and −0.95 [Kihihi], p-values < 2.2×10−16 for all). FWS in Nagongera and Walukuba are similar and lower than that in Kihihi, suggesting that the level of outcrossing is smallest in Kihihi, which is consistent with the pattern of COI.
We also examined the relationship between COI and epidemiological and geographical factors within each site. In Nagongera, COI in young children increased with age until peaking at age 7, and then decreased; sample sizes for the other two sites were too small to estimate trends (Fig N in S1 File). Interestingly, parasite density was negatively correlated with COI after adjusting for age (partial correlation r = −0.15 [Nagongera], −0.27 [Walukuba], −0.23 [Kihihi], p-values = 0.0011 [Nagongera], 0.058 [Walukuba], 0.043 [Kihihi]). This negative association was most pronounced in those aged 3–10 years in Nagongera (Fig O in S1 File), and may reflect the dominance of particular clones in acute, high-density infections. No differences in COI were observed between households with or without Insecticide Treated Nets (ITNs), or between sampling years.
In Kihihi, elevation and COI were negatively associated (r = −0.259, p-value = 0.026), consistent with the previously identified negative associations between elevation and mosquito density, the incidence of malaria, and serological evidence of exposure [41]. Interestingly, the unexpectedly high COI observed in Walukuba was largely driven by samples collected from the West of this sub-county, (Fig 2B; medians = 5 [West] and 3 [East], p-value = 0.027). We have previously noted that mosquito densities in Walukuba are lower in the West, which is closer to urban centers, as compared to the East, which is a fishing village comprised largely of makeshift wooden housing [42]. One potential explanation for this seemingly paradoxical finding—high COI in the lowest transmission part of the lowest transmission site–is that a substantial proportion of these infections were imported from areas of higher transmission, where parasite populations are more diverse and co-transmission of multiple genetically distinct parasites is more likely.
Finally, we compared allele frequencies from each of the three sites to determine whether there was any evidence of population differentiation. We found little genetic differentiation between sites measured based on our estimated allele frequencies (FST ranged from 0.004 to 0.04; Table C in S1 File and S3 Table), although Kihihi, which is somewhat geographically isolated, had slightly higher FST with respect to the other two sites.
Despite the availability of increasingly efficient genotyping technologies for molecular epidemiology, the prevalence of polygenomic infections in malaria-endemic regions hinders the estimation of basic population genetic parameters for Plasmodium falciparum. While COIL can estimate COI using allele frequencies from monogenomic infections or external data, direct estimation of allele frequencies from all samples is a preferable approach, particularly when no relevant frequency data are available and sample size is sufficient to overcome stochastic sampling error. THE REAL McCOIL accomplishes this by incorporating information from polygenomic infections to simultaneously estimate COI and population allele frequencies. We show through detailed simulations that our approach is robust to most model assumptions and can readily handle missing data. In addition, THE REAL McCOIL can utilize raw SNP genotyping data, allowing the method to be robust to errors in allele calling. Analysis of genotyping data from Uganda show that THE REAL McCOIL is able to identify nuances in field data that previous methods could not. In particular, compared with msp2 genotyping or applying COIL to SNP data, we identified much higher average COI overall and epidemiologically relevant variation between and within study sites.
Through a number of simulations, we show that results obtained from THE REAL McCOIL are robust to assumptions that loci are independent and that the parasite population is homogeneous. As would be expected, a high degree of relatedness between parasites within an individual host resulted in substantial downward bias in estimates of COI. This is not trivial, as parasites in some epidemiological settings may be closely related within a host, e.g. due to co-transmission [43]. Fortunately, we found that this bias follows a clear linear pattern and can either be corrected if the level of relatedness is known, estimated directly from the data, or can at least be given reasonable bounds (Text B in S1 File). While estimating the level of relatedness may be challenging, enough information may be present in the data to do so in some cases, as demonstrated by a recent paper which estimated this parameter from sequence-read data [44]. THE REAL McCOIL can also be applied to read-based SNP data, and in theory can be extended to estimate relatedness. While we note that the most obvious model for measurement error in sequence-read data is a binomial distribution (Text C in S1 File), a normal distribution as applied in our current version offers a reasonable approximation and has computational advantages.
Genotyping of one or a few highly polymorphic antigen markers, such as msp1 and msp2, is currently the most common method for determining COI [45,46]. The use of capillary electrophoresis has improved resolution of alleles, but due to the creation of PCR artifacts it is still difficult to accurately measure COI > 5 [38]. Deep sequencing of antigens such as csp is an alternative approach [47,48]. However, with all of these approaches, immune selection on these genes within individuals and in a population can bias estimates of COI in ways which are difficult to predict [49,50]. Since loci under different types of selection can evolve independently in the presence of recombination, the diversity and geographic distribution of loci under immune selection may not be the same as observed among SNP loci. Both recombination rate and immune selection pressure will vary systematically with transmission intensity, resulting in complex associations between different genetic markers. Therefore, multiple genetic lineages defined by SNP panels may be associated with few msp2 alleles, or vice versa, depending on the transmission setting and selective environment. In addition, if lineages within the host are related, using multiple markers across the genome is more likely to detect multiple lineages than using one region of the genome. FWS, based on the difference between within-host and population heterozygosity, is a related metric used to quantify within-host diversity [33]. While FWS is correlated with COI, the metric is conceptually different because it is influenced by both the relative proportions of lineages within the host and population allele frequencies [21,33,40]. estMOI [51] uses phasing information from sequence reads and the number of unique allelic combinations to estimate COI but requires deep sequencing data and can be biased by sequencing error. Some methods that use SNP data to estimate haplotype frequencies also simultaneously estimate COI [52,53]. However, current haplotype-based methods can only consider a limited number of loci (~7) because the number of possible haplotypes quickly expands with the number of loci. We expect that THE REAL McCOIL is better at estimating COI than these methods because it can incorporate a much larger number of SNPs. Moreover, COI estimated from THE REAL McCOIL could be used as a prior in tools estimating haplotype frequencies.
Application of THE REAL McCOIL to genotyping data from Uganda allowed us to calculate allele frequencies and FST, which was not possible to do from the raw data or using COIL due to the high proportion of heterozygous calls. THE REAL McCOIL also provided estimates of COI for all sites, which demonstrated associations with epidemiologic factors not identified using msp2 genotyping. Interestingly, we identified a high COI in the lowest transmission site, potentially indicating importation of parasites from higher transmission areas. Although the possibility remains that recent transmission reduction left complex, chronic infections in its wake, explaining the high COI observed in Walukuba, the simplest explanation is that these infections were imported from high transmission settings nearby. Additionally, our results demonstrated that COI increased with age until age 7, and subsequently decreased, consistent with studies based on msp1 and/or msp2 typing [54–59]. Previous studies reported inconsistent associations between COI and parasite density for children > 2 years old (positive [55,58,60], none [54,61], or negative [62]). We observed a negative association between COI and parasite density in children aged 3–10 in Nagongera. Although higher parasite density may help detect more strains within the host [63–65], the detection of minority strains may be more influenced by relative proportions of the strains [39]. Individuals with high parasite densities may be relatively immunologically naïve and have one or few lineages dominating the infection [66]. Lower parasite densities may be associated with partial immunity and parasite persistence, and consequently the accumulation of parasite lineages [67–71]. Also, parasite lineages are more likely to persist and accumulate in people with low parasite density because they are less likely to have clinical symptoms [70,72] and be treated. The discrepancy between studies can be due to different genetic markers, different transmission setting and immune levels, different contribution of co-transmission vs. superinfections, or some combination of these factors.
In summary, THE REAL McCOIL facilitates population genetic analysis of SNP data from polygenomic infections, which are common in many transmission settings and may predominate even in low transmission settings. Population allele frequency, which was previously difficult to estimate if the majority of samples were polygenomic, can be estimated by THE REAL McCOIL, allowing downstream analysis that requires frequencies, such as estimating FST, FWS, and effective population size (Ne) [32,33,73]. THE REAL McCOIL is not only limited to P. falciparum, but can also be applied to other parasite species with polygenomic infections [74], including Plasmodium vivax [75]. Codes for THE REAL McCOIL are available on GitHub (https://github.com/Greenhouse-Lab/THEREALMcCOIL).
|
10.1371/journal.pcbi.1000952 | HIV Promoter Integration Site Primarily Modulates Transcriptional Burst Size Rather Than Frequency | Mammalian gene expression patterns, and their variability across populations of cells, are regulated by factors specific to each gene in concert with its surrounding cellular and genomic environment. Lentiviruses such as HIV integrate their genomes into semi-random genomic locations in the cells they infect, and the resulting viral gene expression provides a natural system to dissect the contributions of genomic environment to transcriptional regulation. Previously, we showed that expression heterogeneity and its modulation by specific host factors at HIV integration sites are key determinants of infected-cell fate and a possible source of latent infections. Here, we assess the integration context dependence of expression heterogeneity from diverse single integrations of a HIV-promoter/GFP-reporter cassette in Jurkat T-cells. Systematically fitting a stochastic model of gene expression to our data reveals an underlying transcriptional dynamic, by which multiple transcripts are produced during short, infrequent bursts, that quantitatively accounts for the wide, highly skewed protein expression distributions observed in each of our clonal cell populations. Interestingly, we find that the size of transcriptional bursts is the primary systematic covariate over integration sites, varying from a few to tens of transcripts across integration sites, and correlating well with mean expression. In contrast, burst frequencies are scattered about a typical value of several per cell-division time and demonstrate little correlation with the clonal means. This pattern of modulation generates consistently noisy distributions over the sampled integration positions, with large expression variability relative to the mean maintained even for the most productive integrations, and could contribute to specifying heterogeneous, integration-site-dependent viral production patterns in HIV-infected cells. Genomic environment thus emerges as a significant control parameter for gene expression variation that may contribute to structuring mammalian genomes, as well as be exploited for survival by integrating viruses.
| Cellular gene expression is a fundamentally stochastic process due to the intrinsic randomness of the underlying biochemical reactions involved. The resulting stochastically generated expression heterogeneities have important biological consequences and also encode information about the underlying dynamics that generate them. A fundamental goal of transcriptional biology is to understand the quantitative regulation of gene-expression dynamics, which in eukaryotes depends on factors specific to each gene in concert with its surrounding cellular and genomic environment. We investigated the regulatory effects of variable genomic environments by quantitatively measuring expression heterogeneity from diverse single genomic integrations of the HIV promoter in cultured cells. Systematically fitting a model of stochastic gene expression to our measurements reveals transcript production in bursts as the underlying dynamic that accounts for the large heterogeneities observed within single-integration clonal cell populations, with the size of transcriptional bursts as the primary feature that varies over genomic integrations. Our findings implicate genomic environment as an important quantitative control parameter that eukaryotic cells might use to shape gene-expression patterns, and that lentviruses such as HIV, whose genomes are semi-randomly integrated into the genomes of the host cells they infect, may exploit to sample diverse and heterogeneous expression patterns that evade treatment.
| The life cycle dynamics of HIV-1 within a host are shaped by numerous apparently stochastic processes, from the statistics of immune cell infection in humans, to mutation during reverse transcription, semi-random integration of the proviral DNA into the host-cell chromosome, and stochastic viral gene expression thereafter [1]–[7]. We and others have experimentally shown that expression from the HIV-1 promoter is indeed stochastic and shaped by host factors at the viral integration site [4], [8], [9], [10], and we have argued as well that the resultant expression heterogeneities are important in the genesis of viral latency [4], a ubiquitous feature of infection that currently confounds our ability to cure HIV in patients [7], [11], [12], [13]. Gaining a deeper understanding of the factors that influence cell-cell variability in viral gene expression may thus shed light on how to ameliorate the effects of latency, and more generally on the processes that affect the expression of any gene.
The semi-random integration of HIV-1 into the host genome provides a particularly ideal opportunity to dissect the relative contribution of genomic environment as a fundamental element of expression regulation that may contribute importantly to expression dynamics and heterogeneities in eukaryotes. It is now well established that HIV-1 integration is biased towards actively transcribed chromosomal locations [3], [14], and it has been demonstrated that mean expression levels from model HIV-1 viruses correlate with specific epigenetic features at their integrations [9] and of their surrounding genomic regions [3]. Prior studies in other systems focused on how the population average expression of genetic constructs depends on integration context, and have found correlations with the expression levels of surrounding genes and with the local 3-D chromatin structure [15], as well as with DNA methylation, nucleolar association, and DNA diffusional mobility [16]. Importantly, these studies inform us about the features of genomic environment that might affect mean expression levels. However, the effects of genomic environment, or integration site, on stochastic expression and heterogeneity have not yet been explored.
The discrete and stochastic nature of gene expression has been appreciated for some time [17], [18], [19], and it has become increasingly recognized that the resulting expression variability may significantly impact diverse biological functions, shaping the outcomes of cellular decisions, being exploited as a tool for survival in changing environments, and often inducing qualitatively different behaviors than would be predicted from a deterministic understanding [20]–[26]. Theoretical and computational analyses have explored the relative contributions of key processes to heterogeneity in gene expression, including open-complex formation, transcriptional elongation, translation, post-transcriptional and translational processing/modification, as well as chromatin regulation [27]–[32]. Importantly for this study, the latter, an integral element of epigenetic control over gene expression, yields potentially slow and probabilistic dynamics that have been postulated as a significant source of expression heterogeneity in eukaryotes. In parallel with these theoretical studies, experimental approaches have been developed to characterize expression noise arising both ‘intrinsically’ from the biochemical processes directly involved in the expression of any individual gene, as well as ‘extrinsically’ from variability in other cellular processes that more homogeneously affect the expression dynamics of groups of genes simultaneously, such as cell cycle or concentration fluctuations of upstream transcription factors [33]–[37]. Interestingly, genome-scale measurements of expression heterogeneity demonstrate correlations with gene functional class, implying that perhaps noise is a “selected” feature of a gene's expression pattern [38]–[43].
Despite the apparent complexity of cellular transcriptional regulation, for many genes across a broad range of cell types, the patterns of cell-to-cell expression variability within isogenic populations are remarkably well described by simple stochastic models that represent the gene – including the associated genomic environment, chromatin structure, transcriptional regulators, and transcriptional machinery – as existing in a small number of discrete configurations, or states, with expression heterogeneities depending on probabilistic transitions between states and on probabilistic transcript and protein production and degradation [27], [44]. These models are often necessarily abstract, yet they parsimoniously capture many essential features of transcriptional biology. Model fits to clonal single-cell experimental data, primarily in Saccharomyces cerevisiae for eukaryotic studies, has allowed the inference of underlying gene-state and transcriptional dynamics that largely account for observed expression heterogeneities in a number of instances [39], [45]–[49].
A wide range of transcriptional dynamics have been revealed by such analyses, from continuous [38], [48], to ‘pulsatile’ [45], [47], to ‘bursty’ [46], [48]. These diverse dynamics are effectively characterized by the frequency of gene-activation events, their duration, and the number of transcripts produced during each event [29], [30], [48], [50], and contrasting results have emerged concerning the relative contributions of cellular regulation of each of these quantities to specifying the expression pattern of any given gene. For example, several pioneering studies, of single, targeted integrations of inducible, synthetic constructs, in yeast have suggested that the concentration of inducer largely controls the frequency of gene-activation events rather than the number of transcripts produced by each event [45], [47]. A subsequent study in yeast – which considered three targeted integrations of similar constructs into 1) adjacent locations on a single chromosome, 2) homologous locations on sister chromosomes, or 3) non-homologous chromosomal locations – similarly found that transcriptional activation frequency varied between locations [36]. Genome-scale studies of stochastic gene expression in yeast suggest as well that the primary feature of transcriptional dynamics that varies between genes, over a wide range of genes, is the frequency of transcriptional activation events [38], [39]. In contrast, an elegant study in mammalian cells quantified expression heterogeneities – from a single, random integration of a Tet-inducible construct into one locus in the genome – by using fluorescent in-situ hybridization (FISH) to directly visualize single transcripts [46]. The authors concluded that transcripts are produced in bursts, and that the typical number of transcripts produced during a burst (referred to as the ‘transcriptional burst size’), rather than the frequency of bursting, was the primary measure that varied with tetracycline induction level.
While the above studies have begun to characterize the dependence of gene-expression dynamics on a number of cellular inputs, a systematic, quantitative investigation of the contribution of genomic environment over a broad range of genomic integration positions remains to be conducted. Furthermore, the contrasting observations as to whether transcriptional activation frequency, transcriptional burst size, or some other feature of transcriptional dynamics represents the primary variable that cells modulate to control expression patterns in these diverse systems raise key questions of how important features of genetic, epigenetic, and regulatory architecture may differ in yeast and mammalian cells.
Here we explore the fundamental relationship between genomic environment and expression heterogeneity from a diverse set of semi-random single integrations of a model HIV-1-promoter/GFP-reporter construct in cultured Jurkat T-cells. Systematically and rigorously fitting a model of stochastic gene expression allows us to infer the underlying expression dynamics that account for the single-gene expression distributions that we measure from single-integration clonal populations. Our analysis reveals that transcript production in bursts accounts for the wide, highly skewed, expression profiles that we observe, and importantly that transcriptional burst size is the primary feature that varies across viral integrations. These results interestingly suggest that the virus samples a particularly ‘noisy’ range of possible expression profiles across cellular integrations, and open a number of important questions for further study. We propose several qualitative models that may explain this inferred variation of transcriptional dynamics with genomic environment and discuss the implications of our findings for HIV dynamics, and for cellular gene expression in general.
Although HIV-1 requires transactivation by the virally-encoded protein Tat to amplify its expression [51], the HIV-1 promoter still supports basal transcription in the absence of Tat [9], which occurs initially after viral infection but before significant viral protein is produced. The dynamics of this basal expression, and the associated expression heterogeneities that result, may play an important role in affecting the cellular ‘decision’ between lytic viral production and latency [4]. To study heterogeneities in basal expression from the HIV promoter, we infected Jurkat T-cells, at a low multiplicity of infection (MOI), with a model HIV-1 virus that contains the full-length LTR driving expression of a GFP reporter but no viral genes. Cells with single integrations were isolated by fluorescence activated cell sorting (FACS) and expanded into clonal populations. The resulting clonal GFP expression profiles were quantified by flow cytometry and smoothed for comparison to model distributions in the analysis that follows. Thirty-one such clones, with average florescence levels ranging over an order of magnitude, and expression profiles clearly distinguishable from a measured autofluorescence profile, were selected for analysis (Fig. 1A). Integrations whose mean fluorescence was less than twice the autofluorescence mean were not included in our analysis.
The shape features of our experimental distributions (such as mean, variance, skewness, etc.) are diagnostic of the underlying expression dynamics that generate them – and of the regulatory role of various molecular ‘inputs’ such as integration position (as well as promoter structure and concentrations of transcription factors, which were the ‘inputs’ considered in several other elegant studies: [38], [45], [46], [47]). For example, a simple model assuming transcript number fluctuations as the primary source of expression heterogeneity, with only the rate of constant transcript production varying with a given ‘input’, predicts a Poisson-like distribution shape variation whereby distribution variances ( considered as a measure of distribution width and expression heterogeneity) vary proportionately to the mean (, for mean ). Such a variation, illustrated by the lower dashed line in Fig. 1B (‘Poisson’), has been observed over a large set of yeast promoters under multiple experimental conditions [38], [39]. Alternatively, a model in which distribution shape variations are effectively described by a simple scaling of single-cell fluorescent values by an ‘input-controlled’ constant value (, where is the probability of observing fluorescence , for a normalized value of the ‘input-controlled’ parameter ) would predict distribution variances to vary in proportion to the mean squared (, Fig. 1B upper dashed line, ‘Scaling’). Such a shape variation might arise if heterogeneities are instead due primarily to probabilistic transitions between promoter configurations that specify different transcription rates, with only these transcription rates varying (proportionately) with the ‘input’ from one clonal distribution to the next. In contrast to these possibilities, we find that the trend in distribution-shape variation over our set of clonal populations is best described by a relationship where the distribution variance changes proportionately to the mean raised to the 1.7±0.2 power (Fig. 1B, solid regression). This characteristic trend differs significantly from either of the above simple models (P<0.025), suggesting that neither is sufficient, and that integration-site variation may specify a more complex modulation of promoter and transcriptional dynamics in our system.
To visualize additional features of the expression distributions over the set of clones, we translated each to a common mean fluorescence, and correspondingly scaled its fluorescence values about that mean based on the variance regression in Fig. 1B, revealing a ‘typical’ distribution shape that is wide and highly skewed (Fig. 1C). These features are signatures of a bursty underlying transcriptional dynamic [27], [50], as we discuss in further depth below.
A simple stochastic model that captures a number of essential features of transcriptional biology, and that can reproduce a range of single-gene expression profiles, assumes that the promoter may exist in either an activated state (φa) that produces mRNA probabilistically at a fixed rate (κt+), or repressed state (φr) that is unproductive (Fig. 2A). These model states may represent different characteristic configurations of chromatin and/or transcriptional complexes, with transitions between them occurring at rates κa and κr. Together with the active-state transcription rate, these lumped parameters represent contributions from diverse modes of genetic and epigenetic transcriptional regulation that may depend essentially on features of the genomic environment at the viral integration sites. Variants of this model have been used in other studies as well to analyze single-gene expression data [45]–[49] and have also been studied theoretically [27], [52], [53], [54].
In the analysis that follows, we always consider steady-state model distributions, since longitudinal measurements over the course of a week on several clones indicate that distribution shapes are relatively stable over our time scale of interest (see Fig. S3 and Sec. S.VII for further discussion). Furthermore, we determine all rate constants relative to the transcript degradation rate (, estimated to be approximately 0.2 h−1, see Text S1 Sec. S.V), as their relative rather than absolute values determine expression profiles at steady state. In addition, we adopt the working hypothesis that our experimental distribution shapes are determined by the intrinsic processes represented in our model at fixed values of its rates – possible contributions of extrinsic sources of variability have been considered in earlier work on this system [4] and are discussed further in the Supporting Information (Text S1, Sec. S.VIII).
The qualitative expression regimes of the two-state gene model fundamentally depend on the relative values of the gene-state transition rate constants (Fig. 2B), with different dynamics corresponding to different potential underlying transcriptional regulatory mechanisms. ‘Fast’ gene-state dynamics (, perhaps specified by fast binding and unbinding of transcription factors) approximate continuous transcription from a single fixed gene state and can generate relatively narrow Poisson-like expression profiles, which widen for ‘Intermediate’ dynamics (). ‘Slow’ gene-state dynamics (, due perhaps to slower changes in chromatin configuration) may generate multiple transcripts after each transition to a relatively stable active state, and the dynamics can be described as ‘pulsatile’ [45], [47]. Distributions become bimodal in the extreme case.
Another dynamic regime that has received considerable attention can be termed the transcriptional ‘bursting’ regime, in which the gene inactivation rate is fast (), and the transcription rate is sufficiently large ( not small) that transcriptional bursts of average size are produced during short excursions of frequency to a relatively unstable active gene-state (see Text S1, Sec. S.V for further discussion, and Refs. [29], [30], [48], [50], [55]). Distributions in the ‘bursting’ regime are wide and highly skewed, in qualitative agreement with the ‘typical’ HIV-LTR distributions from our measurements (compare Fig. 1C and 2B, solid curves), with both the protein and transcript distribution means and variances approximately demonstrating a relatively simple dependence on transcriptional burst size and frequency: ; (see Text S1, Secs. S.III – S.V). Indeed, by assuming a model solution in the ‘bursting’ regime, one can analytically calculate a unique transcriptional burst size and burst frequency that reproduce the mean and variance of each of our experimental distributions, with good qualitative agreement in distribution shape (see Text S1, Secs. S.III and S.VI for further discussion). Furthermore, variation in transcriptional burst size and frequency, individually or in combination, can account for the range of distribution-shape variation discussed in Fig. 1B, with the best agreement to our experimental observations occurring if burst size and burst frequency typically vary simultaneously, but with the dominant effect coming from burst-size variation (Fig. 2C).
Though the relatively slow time scale of protein degradation in our system () effectively ‘filters’ some of the dynamic information propagated from model transcript to protein distributions, we emphasize that the calculated protein distribution shapes still reflect the underlying transcript distribution shapes and demonstrate distinctive features in each expression regime (Fig. 2B). Below, through careful analysis, we will make use of this observation, building on the qualitative analysis developed here, to quantitatively infer the underlying heterogeneity-generating gene-state and transcriptional dynamics within our system from measured protein expression distributions, and to determine quantitative bounds on our ability to distinguish between different dynamic regimes. Of considerable benefit for this analysis, cytometry-based protein measurements can be acquired rapidly (e.g. compared to microscopy-based transcript-counting measurements), allowing good resolution of the probability distributions that underlie the expression histograms collected over populations of cells, and enabling measurements on sufficient numbers of clones to identify trends in the variation of single-gene expression distributions over integration sites.
While the analysis above provides intuition as to the dynamics and regulation that may underlie our experimental observations of the HIV LTR, it is solely a qualitative assessment based on the assumption of ‘burstiness’ and comparisons to ‘stereotypical’ model distributions. In reality, model distributions vary continuously between regimes, and means and variances provide an incomplete characterization of the actual distribution shapes. Therefore, to better determine the degree to which transcriptional ‘bursting’ best accounts for our experimental distributions, and the degree to which it can be distinguished from other possible dynamic regimes, we used a systematic fitting routine to identify the best-fit combination of transcription rate and gene-state transition rates for each distribution. Transcript degradation, protein production, and protein degradation rates (κt−, κp+, and κp−, Fig. 2A) were fixed at values that were separately measured.
An important indicator of the dynamic regime of our system is the average time that the promoter remains in the active configuration following a gene activation event, relative to the average life time of a transcript (see Fig. 2B), specified by τ = κt−/κr, which we refer to as the ‘active duration’ (τ). We therefore began by identifying best-fit sets of model parameters for each clone over a range of fixed values for τ. We arrived at a robust estimate of the range of parameters for which the model quantitatively accounts for our measured distributions by considering the ratio, Devr, of each best-fit deviation at a given τ, to a bootstrap-estimated 95% upper bound on the deviation expected due to uncertainty in our measurements, which served as a metric for identifying model fits whose quality was statistically comparable. Fits for which the values of Devr differ by less than 1 for a given clone were considered to be effectively indistinguishable, since their differences may be accounted for by uncertainty in our experimental data, and these fits were thus considered to identify a range of parameters for which the model gives a statistically comparable account. The work-flow for our analysis is summarized in Table S1; the definitions that we used to quantify fit deviations, as well as the error model used for our bootstrap error calculation, are discussed briefly in the Materials and Methods, and in more depth, Text S1, Secs. S.I and S.VII, together with Fig. S1.
We find that the optimal agreement between model and experiment always occurs at short active-state durations (sample fits given in Fig. 3A), with deviations increasing for larger values of τ (Fig. 3B), past a distinguishability cut-off (where Devr has increased by 1) that effectively marks the resolution limit of our analysis, which we call τMax for each clone (Fig. 3C). The value of τMax defines a range of active durations (bounded below by τ = 0), for which the quality of model fits for a given clone is comparable, and acts as a measure of how well our analysis can distinguish a ‘bursty’ underlying dynamic from other regime possibilities. Small values of τMax indicate model fits where short-lived gene activation events, which are a hallmark of transcript production in bursts, provide a significantly better account of our experimentally measured distributions than a less noisy dynamic (i.e. one specified by longer active durations). Because we do find that the best model fits always occur at the shortest active durations (where the relative deviation Devr = DevrOpt, Fig. 3B), we conclude that a transcriptional dynamic in the ‘bursting’ regime does indeed always give the best quantitative account of our data, and we further note that larger predicted transcriptional bursts (generally associated with brighter clones) are correlated with better regime resolution (Fig. 3C). Finally, we note that our systematic distribution fitting procedure always resulted in improved fits over those obtained by only considering the first two distribution moments, with the improvement often statistically significant. Nevertheless, small systematic deviations remained, which are discussed further in Fig. S2.
From the optimal fits above we identified best-fit transcriptional burst sizes and frequencies that specify the predicted transcriptional dynamics for each integration clone. Importantly, we find that the transcriptional burst size is the primary feature that varies over the set of genomic environments sampled by our 31 viral integrations, increasing from a few transcripts in very dim clones to tens of transcripts in very bright clones (Fig. 4A, with σ/μ = 3.5 for the distribution of log10(b)). Consistent with the qualitative analysis in Fig. 2C, we find that transcriptional burst size varies approximately sub-linearly with expression-distribution mean (, R2 = 0.66). In contrast, the transcriptional burst frequencies inferred through our analysis are scattered about a characteristic value of one burst per several transcript degradation times, corresponding to several transcriptional bursts per cell-division time (Fig. 4B). In addition, these frequency values vary no more than several-fold (σ/μ = 2.2 for log10(κa)), and they demonstrate little correlation with distribution mean (, R2 = 0.2). These results were maintained, to within the accuracy of our analysis, when the scattering gate used to control for cell-size variability in our experimental distributions was decreased by a factor of 6 from the value that was found to be optimal for our analysis (see Text S1, Sec. I and Fig. S4), indicating a robustness to this source of uncertainty, which had been found to significantly impact results from other cytometry-based analyses of expression variability [38]. Further, we find no significant correlation between the inferred transcriptional burst sizes and burst frequencies that might influence the interpretation of their trends with expression mean (see Fig. S5).
Our findings thus indicate that burst-size variation makes the dominant contribution in controlling single-gene expression profiles and represents the primary feature of transcriptional dynamics whose modulation distinguishes typical bright from dim clones. Importantly, the trends noted in Fig. 4 characterize the modulation of a ‘typical’ LTR integration by the sampled genomic environments. However, we must emphasize that the significant scatter of both the burst sizes and burst frequencies inferred for each individual clone about these ‘typical’ variations, as well as the possibility that a different trend may exist for very dim integrations (which were not considered in this study), suggest a potentially richer behavior that may still be uncovered through further study.
Another recent study has also considered a two state model to analyze expression variability from the HIV LTR [56]. This study similarly suggests that transcript production occurs in bursts and that both burst size and frequency vary with LTR integration position, though the analysis is qualitative, based only on consideration of distribution moments. In contrast to our findings, they emphasize burst frequency modulation as structuring distribution-shape variation over integration positions, as well as in response to pharmacological perturbation, though the later finding is difficult to interpret, as a steady-state model is considered to analyze data that are clearly varying in time. Additional quantitative analysis, including systematic model fitting, would be necessary to characterize the relative contributions of burst-size and burst-frequency modulation in this study, and to determine whether its findings are consistent with our own.
A correlate of our findings – that transcription in short bursts underlies basal expression heterogeneities from the HIV LTR in the absence of Tat – is that the active promoter configuration is short-lived. This implies that the promoter would be observed in the active configuration for only a small fraction of cells in a clonal population at any given time at steady state. The value of this fraction in the two-state model, which we refer to as the ‘active fraction’, f, is related to the activation frequency (κa, whose value is relatively well resolved for each clone by our analysis, Fig. 4B) and active duration (τ, for which our analysis only provides an upper bound τMax, Fig. 3C), as . Our analysis provides a predicted upper bound on this fraction for each clone as (Fig. 5, bars), where any value of f below is consistent with our analysis. Small values of specify clones for which the active fraction is indeed predicted to be small, while larger values indicate clones for which its value is less well resolved. In particular, our analysis predicts that although the brightest and dimmest integration clones considered in our study differ in mean expression by a factor of approximately 30, the brightest clones will nevertheless only be observed with the integrated LTR in the ‘active’ transcript-producing configuration less than 20% of the time.
Transcriptional burst size – defined by the ratio of the transcription rate to promoter-inactivation rate (or the product of transcription rate and the active duration ) – can be modulated by two qualitatively difference ‘Modes’ of regulation. First, integration position could affect the dynamics of promoter inactivation (), reflecting integration-position effects on the stability of the active configuration, possibly due to direct effects of the surrounding chromatin configurations on the energetics of the active configuration and/or the stabilizing effects of regulatory factor recruitment by surrounding regions (Mode 1). Alternately, integration position could affect transcriptional productivity in the active state (), which may also be affected by modulation of chromatin configuration and/or recruitment of regulatory factors by surrounding genomic regions (Mode 2). We have seen in our analysis to this point, that model fits of our cytometry data cannot separate the two constituent parameters that define transcriptional burst size, and therefore they cannot resolve these two possible ‘Modes’ of its regulation (e.g. Fig. 3; a similar parametric indeterminacy has been noted by [46], [48]). Furthermore, the potentially overlapping effects of many molecular regulatory mechanisms on transcriptional dynamics may make it difficult to define experiments that directly distinguish these ‘Modes’, and to decouple their regulatory contributions.
However, our analysis predicts that each ‘Mode’ of control leads to a distinct pattern of active-fraction variation over the set of integration clones (Fig. 5, symbols): for Mode 1 the active-fraction varies proportionately to the clonal expression mean, whereas for Mode 2 the scatter in active fraction predicted over our set of integration clones reflects scatter in the predicted burst frequencies. We thus suggest that future experimental analysis of the active fraction may provide a means of distinguishing between these two key ‘Modes’ of integration-site modulation of gene expression.
Our findings, that expression from the HIV promoter is characterized by transcript production in bursts and that the site of viral integration primarily modulates transcriptional burst size, contribute to an emerging paradigm for transcriptional regulation that emphasizes the importance of stochastic/probabilistic dynamics [20], [27], [50], [57]. In particular, the expression patterns that we observe from single integrations of the HIV promoter cannot be accounted for by transcription from a single, fixed state of promoter activation, which would involve a single transcription rate that specifies a comparatively narrow single-gene expression profile with little variation over a population of cells. Rather, our analysis predicts that the large expression heterogeneities observed in this system (Fig. 1) are shaped by probabilistic transitions between (at least) two distinct configurations (Fig. 2A), with the promoter spending only the minority of time in the transcriptionally active configuration even for the most productive integrations (Fig. 2B, Figs. 3B,C and Fig. 5). Furthermore, our analysis suggests that an essential component of the regulatory effect of genomic environment at the viral integration site is to modulate the dynamics of transitions between states of differing transcriptional activity, in addition to possible effects on the transcriptional activity of each state (Figs. 4 and 5). Of note, it is only by systematically fitting a quantitative model to our measurements that these underlying dynamics were revealed, as quantitative single-cell measurements of protein expression only provide an indirect measure, and it is only by applying our systematic analysis to observations across a diverse sampling of integration-modulated expression patterns that we succeeded in extracting a characteristic effect of integration position on transcriptional dynamics.
Transcript production in bursts is a particularly ‘noisy’ transcriptional dynamic that can generate significant cell-to-cell expression variability, which is reflected in wide and highly skewed single-gene expression distributions across clonal populations (Fig. 1C, 2B). In particular, the ‘typical’ distribution identified in Fig. 1C demonstrates a coefficient of variation (standard deviation/mean, or relative width), corresponding to 60% variability. This value is significantly larger than the values observed for most eukaryotic promoters in several large-scale studies (compare to data in: [38], [39], [58]), and we anticipate that a number of features of the HIV promoter, some of which are common in mammalian promoters, may conspire to account for this ‘noisy’ expression pattern.
Similar to HIV expression shortly after infection, our system lacks the viral transcriptional activator Tat. In the absence of Tat the LTR has been observed to bind repressive factors that maintain a non-conducive chromatin configuration [14], , and the likely greater stability of this ‘inactive’ configuration may limit the fraction of time that a transcriptionally ‘active’ configuration can be maintained. On the other hand, like many mammalian promoters, the HIV LTR contains multiple binding sites for repressing and activating elements (which still bind in the absence of Tat), several of which affect chromatin state and bind competitively and/or cooperatively. For example, the histone-acetyltransferase (HAT) p300 and the activating NF-κB component RelA are thought to bind their respective HIV-1 Sp1 and NF-κB sites cooperatively to activate transcription, and in competition with the histone deacetylase (HDAC) recruiting activity of SP1 and the p50/p50 homo-dimer that bind the same sites respectively to inhibit transcription [10], [62]–[66]. One may hypothesize that this competition could thus lead to an infrequent all-or-none binding of activating factors that directly remodel promoter-bound nucleosomes to establish a transcriptionally conducive chromatin configuration [10]. In addition, the LTR includes a number of other cis-regulatory elements that bind transcriptional activators such as NFAT and AP-1 [67], [68], as well a TATA motif that contributes core transcriptional complexes [69], [70]. These elements may enable more efficient recruitment, assembly, and stabilization of a productive transcription complex, with transcriptional reinitiation potentially yielding multiple transcripts from each gene-activation event (the presence of a TATA box has been linked to increased expression noise in other studies as well, see for example: [41], [47], [71]). In combination, the above features may specify transcript production during short, infrequent bursts, consistent with the results of our analysis.
Intriguingly, a recent mammalian genome-wide mapping of HAT and HDAC association found them simultaneously bound to a large number of active promoters, suggesting that simultaneous regulation by competitive epigenetic regulators may be more common than previously thought [72]. It is therefore possible that transcript production in bursts represents a more general feature of mammalian expression regulation, and it will be interesting to discover how properties of the HIV promoter shape its transcriptional dynamics, and whether similar promoter architectures specify ‘bursty’ dynamics for other genes.
Our findings suggest that transcriptional burst size is a more ‘locally’ determined property, more sensitive to those features of genomic environment that vary significantly between integration sites, whereas transcriptional burst frequency is, by comparison, a more ‘globally’ determined feature, specified by interactions with the cellular environment that may be more promoter-specific but less significantly integration-site dependent. Burst frequency reflects the statistics of assembling the more active promoter configuration from an inactive one, and we might speculate that this transition depends in part on large-scale chromatin reorganization and dynamics that are coordinated globally across the nucleus [73], [74], [75]. Burst size, on the other hand, is a property of the transcriptionally ‘active’ configuration, and we may conjecture that some of the reorganization that accompanies its establishment also may provide opportunities for important ‘local’ features to exert their regulatory influences. For example, chromatin remodeling may expose new binding sites for transcriptional regulators [31], [76], and the initiation of transcriptional activity could contribute to association with ‘nearby’ transcription factories where additional transcriptional regulators are localized, and where interactions with surrounding (and possibly distant) genomic regions may be enhanced [73].
At a more basic level, a feature of transcriptional burst size that could more generally account for a greater sensitivity to genomic environment is its complimentary dependence on transcriptional productivity and the stability of the active promoter configuration. We had noted earlier that this complimentary dependence specifies two distinct ‘Modes’ by which surrounding genomic regions may differentially affect the resulting transcriptional dynamics (see Fig. 5), both of which might be effected by recruitment of transcriptional regulators by surrounding genomic regions, epigenetic features of the surrounding regions, and the transcriptional activity of neighboring genes [15], [75], [77]. If we assume that a ‘typical’ more productive integration increases , τ, and all proportionately (i.e. without assuming a weaker dependence for burst frequency), then the dual dependence of burst size would dictate that it vary as burst frequency squared (), and the scalings and would result, which fall within the 95% confidence interval of our regression analysis in Fig. 4. This possibility is consistent with our suggestion in the previous subsection that the architecture of the HIV LTR may effectively couple the control of gene activation and inactivation, and with the hypothesis that the chromatin regulators that may control these dynamics could also modulate the active-state transcription rate either directly or indirectly. Such a combined ‘Mode’ of modulation would specify an active-fraction variation intermediate between that predicted for the two pure ‘Modes’ of modulation considered in Fig. 5, and might be used to distinguish it experimentally.
Burst-size variation with promoter induction level from a tetracycline-inducible construct at a single genomic position has been noted in another study using mammalian cells [46], though this result contrasts with a number of yeast studies that have identified the frequency of gene-activation events as the primary feature that varies with genetic-construct induction level [45], [47], over a single set of three targeted genomic loci [36], and over a large set of endogenous promoters [38], [39]. It thus remains to be determined whether our observation of burst-size variation represents a mode of transcriptional regulation particular to mammalian cells or to transcriptional control by genomic environment, or whether it is determined by any specific features of the HIV promoter that dictate a unique coupling to mammalian genomic environments that might be shared by other ‘bursty’ promoters and cell types. Future studies investigating greater numbers of genomic integrations, in our and other systems, that correlate expression variability with promoter and surrounding genomic sequences, may provide important answers to such questions.
The observation that integration site primarily modulates transcriptional burst size from the HIV promoter implies that viral integrants sample a ‘noisy’ set of basal expression distributions by semi-randomly integrating in the genome. Specifically, relative distribution widths (i.e. the coefficient of variation) are approximately maintained and comparable between ‘dim’ and ‘bright’ integrations. This contrasts with the naive expectation that dimmer integrations should demonstrate greater relative expression heterogeneity due to larger relative fluctuations typically generated by smaller numbers of molecules, as would be the case if burst frequency were the primary covariate over viral integrations (see Fig. 2C), and as was found to be the case over a large sampling of yeast promoters [38], [39].
The basal expression patterns, and their associated expression noise, that were measured here reflect the range of expression dynamics that may be generated initially from an HIV infection after its semi-random integration into the host genome but prior to significant production of viral proteins [4], [9]. Productive viral replication depends on subsequent production of the HIV protein Tat, which mediates expression transactivation by enhancing both transcript elongation from the LTR as well as the binding of other transcriptional activators [51], [78]–[81]. In an intact virus, this positive feedback would act to amplify the basal expression fluctuations observed here.
We anticipate that certain ranges of parameters representing integration-site dependent basal fluctuations in promoter activity may act to specify distinct infected-cell fates, as illustrated in Figure 6 where the drawn region boundaries are hypothetical and the insets depict representative expression phenotypes that result when Tat is expressed from the HIV LTR in a minimal viral system that we had studied in earlier work [4], [10]. Promoter integrations with smaller basal transcriptional burst sizes, and with frequencies that do not effectively couple one burst to the next, will never produce sufficient Tat for transactivation and may represent unproductive infections (Region I). On the other hand, promoter integrations specifying larger basal burst sizes and sufficient frequencies will quickly and stably transactivate after a small number of initial transcriptional bursts and may represent a productive infection (Region II). In contrast, those integrations with small to intermediate basal burst sizes and frequencies will only infrequently (stochastically) generate sufficient Tat for positive feedback activation. Moreover, the transactivated state may be subsequently destabilized by the infrequent occurrence of consecutive smaller and more widely spaced bursts, to generate a bimodal expression pattern (Region III). We have hypothesized that the dynamics of this phenotype, which include significant delays in switching between non-productive and productive expression phenotypes, may create a sufficient time window for the establishment of latent infections in vivo [4], [10]. Future experimental and computational analysis may provide additional insights into the role of Tat in amplifying basal, integration-modulated, expression fluctuations, as well as their hypothesized role in fate determination of HIV-infected cells.
While other studies have considered the effects of genomic environment on mean expression, we have analyzed its effects on expression heterogeneities. By applying an integrated computational and experimental approach, we have characterized the modulation of underlying transcriptional dynamics by genomic environment in human cells. Since classes of human promoters often share common enhancer and repressor motifs, it is possible that two such promoters at different genomic loci would demonstrate significantly different transcriptional dynamics, as we have observed from different integrations of a single promoter in our system. In this way, genomic architecture would provide an additional axis of expression regulation complementary to that specified by individual promoter sequence architectures, and promoter and genomic architectures might evolve in parallel to optimize their coupled contributions to transcriptional control [72], [82]–[85]. Similarly, integrating viruses such as HIV, whose host-cell specificity determines the range of possible genomic environments that could be selectively sampled, may evolve promoter architectures that best exploit this host-regulatory axis to adopt a set of expression patterns that enhance, or even optimize, viral replication fate.
The HIV-1 based lentiviral plasmid, pCLG, (encoding the HIV-1 LTR and GFP) was packaged and harvested in HEK 293T cells using 10 mg of vector, 5 mg pMDLg/pRRE, 3.5 mg pVSV-G, and 1.5 mg pRSV-Rev, as previously detailed [4], [86]. Harvested lentivirus was concentrated by ultracentrifugation to yield between 107 and 108 infectious units/ml. Approximately 103–106 infectious units of concentrated virus were used to infect 3×105 Jurkat cells. Six days after infection, gene expression of infected cells was transactivated by incubating Jurkats with a combination of 20 ng/ml TNFα, 400 nM TSA, and 12.5 mg exogenous Tat protein [10]. After stimulation for 18 hours, GFP expression was measured by flow cytometry, and titering curves were constructed by determining the percentages of cells that exhibited GFP fluorescence greater than background levels. This titering curve was used to attain the desired MOI (∼0.05–0.10).
Forty-eight single GFP+ LTR-GFP (LG) Jurkats (clones) were sorted on a DAKO-Cytomation MoFlo Sorter into 96-well plates and cultured for at least 4 weeks to allow for clonal expansion. Infected cultures were analyzed via flow cytometry on a Beckman-Coulter EPICS XL-MCL cytometer (http://biology.berkeley.edu/crl/cell_sorters_analysers.html). Thirty-one single-integration clones, whose expression histograms were sufficiently distinguishable from an autofluorescence profile for model fitting (with mean fluorescence exceeding twice the autofluorescence mean), were selected for further analysis.
Cytometry measurements on 104 cells for each clone quantified GFP fluorescence as well as forward and side scatter (FSC and SSC). Live cells were selected by standard gating of FSC and SSC, and further gated to select the mid 60% of FSC and SSC values. This gating was optimized using a bootstrap approach to resolve the GFP profile at the mean scattering measure, while eliminating significant correlation between GFP distribution and scattering (see Text S1 Sec. S.I for further discussion, Fig. S1, and Table S1). GFP histograms were smoothed using an optimized low-pass Fourier filter, and normalized to obtain probability distributions, that were used for model fitting. Distribution deconvolution, for the transformation applied in Fig. 2B, was accomplished using a Weiner filter. Model fits were also obtained for distributions resulting from a 10% scattering gate, and indicate no significant effect on our parameter inferences (Fig. S4).
The model in Fig. 2A represents a continuous-time discrete-state Markov process described by a chemical master equation [87], which was solved using an in-house Matlab routine (The MathWorks, Inc.; code available upon request) for steady-state protein distributions, which were then convolved with a separately measured autoflorescence profile and converted to cytometer-based RFU (Relative Fluorescence Units) values for comparison with smoothed experimental distributions. Briefly, the master equation was truncated at large protein and transcript numbers to specify a finite system. A graded coarse-graining approach was applied, whereby neighboring states at higher transcript and/or protein number, where distributions admit a continuum approximation, were binned together (probabilities summed), and transition rates between binned states were approximated by interpolation to estimate probability fluxes at the boundaries. The coarse graining scheme reproduces the master equation at small transcript and protein numbers (where no coarse-graining is applied), and specifies a second-order approximation to the corresponding Fokker-Planck equation at large protein and transcript numbers. The resulting linear system was then integrated in time until an effectively stationary distribution was achieved by using a forward/backwards Euler method that alternates treating transcript and protein transitions implicitly or explicitly; this represents a fast and stable method, appropriate to stiff systems and multi-dimensional PDEs [88]. Marginal coarse-grained protein distributions were then calculated by summing calculated probabilities over transcript numbers and gene states, and the resulting distributions were interpolated. Solution accuracy was established by comparing the first three moments of the calculated distributions to their theoretical values (calculated analytically, see Text S1, Sec. S.III), by varying the coarse graining and the time step for the integrator, and by comparing our solutions to those calculated using the Finite State Projection algorithm developed by Munsky and Khummash [89], which allows a rigorous calculation of numerical error for finite times, for several test cases. Further details may be found in Text S1 Sec. S.II.
Fit parameters (κa, b = κt+/κr, and κr) were varied using the MATLAB minimization function ‘fmincon’ to identify the combination that minimized the fit deviation, defined as , where is the predicted/measured probability of counting a cell in cytometer bin i for the data/fit.
A number of model parameters quantify processes occurring at spatially separate locations from the integrated LTR. These were assumed to be the same for all integrations, and were specified separately. Methods developed independently from this study allowed us to calibrate relevant non-fit model parameters via comparison between the measured transcript distribution for a single clone, and the corresponding cytometry-based GFP distributions (Foley, et al. manuscript in preparation). A conversion factor between transcript number and RFU could be estimated from the measured ratio of means, as = 2.5 ( = measured cytometry-based RFU/transcript mean). By assuming transcriptional bursting, the ratio of transcript to protein degradation rates could be calculated as , yielding a value of for our measurement. These constitute the remaining quantities necessary to specify our model. While uncertainties in these quantities would affect the values inferred for model fit parameters, they would approximately affect the inferred fit parameters for each clone by the same scale factor, preserving inferred trends in parameter variation over the set of integration clones. These uncertainties were therefore not explicitly considered in our analysis (see Text S.1, Sec. S.VII for further discussion). Quantifying the dilution of a synthetic non-degraded fluorescence marker allowed us to estimate a cell-division rate of 0.05 h−1, which served as an effective protein degradation rate () in our model, and thus specified ; the absolute values of these degradation rates were not essential to specifying our model because steady-state distributions only depend on ratios of rate constants, and all rates were therefore scaled relative to the transcript degradation rate in our analysis. The relatively large protein numbers in our system dictate that fluctuations in protein production and degradation do not significantly influence distribution shapes, and as long as the ratio was chosen to be a sufficiently large value, its specific value did not affect our analysis; we chose . See Text S1, Sec. S.V for further discussion of non-fit model parameters.
A bootstrap procedure was used to estimate a 95% upper-bound on the value of Dev for our processed experimental distributions (Devdata) that included uncertainties due to the finite number of cells sampled and to specifying distributions at a single scattering measure. Other sources of uncertainty, such as cytometer PSF and distribution variability over time, were found not to significantly affect our determination of trends in model-parameter variations over the set of integration clones and were not included (see Text S1, Sec. S.VIII, and Fig. S3). Model fits whose deviations (Devfit) differed from each other by less than Devdata were considered effectively indistinguishable, as the differences in their quantified deviations might be accounted for by uncertainty in our experimental distributions. 95% confidence intervals about best-fit model parameters were calculated as maximum variations for which the increase in Devr = Devfit/Devdata was less than 1 (assuming simultaneous parameter variations), as estimated using a Hessian-based quadratic approximation for variation of Devr with respect to burst parameters and based on the parametric sampling in Fig. 3 for κr.
|
10.1371/journal.ppat.1001097 | The Microbiota Mediates Pathogen Clearance from the Gut Lumen after Non-Typhoidal Salmonella Diarrhea | Many enteropathogenic bacteria target the mammalian gut. The mechanisms protecting the host from infection are poorly understood. We have studied the protective functions of secretory antibodies (sIgA) and the microbiota, using a mouse model for S. typhimurium diarrhea. This pathogen is a common cause of diarrhea in humans world-wide. S. typhimurium (S. tmatt, sseD) causes a self-limiting gut infection in streptomycin-treated mice. After 40 days, all animals had overcome the disease, developed a sIgA response, and most had cleared the pathogen from the gut lumen. sIgA limited pathogen access to the mucosal surface and protected from gut inflammation in challenge infections. This protection was O-antigen specific, as demonstrated with pathogens lacking the S. typhimurium O-antigen (wbaP, S. enteritidis) and sIgA-deficient mice (TCRβ−/−δ−/−, JH−/−, IgA−/−, pIgR−/−). Surprisingly, sIgA-deficiency did not affect the kinetics of pathogen clearance from the gut lumen. Instead, this was mediated by the microbiota. This was confirmed using ‘L-mice’ which harbor a low complexity gut flora, lack colonization resistance and develop a normal sIgA response, but fail to clear S. tmatt from the gut lumen. In these mice, pathogen clearance was achieved by transferring a normal complex microbiota. Thus, besides colonization resistance ( = pathogen blockage by an intact microbiota), the microbiota mediates a second, novel protective function, i.e. pathogen clearance. Here, the normal microbiota re-grows from a state of depletion and disturbed composition and gradually clears even very high pathogen loads from the gut lumen, a site inaccessible to most “classical” immune effector mechanisms. In conclusion, sIgA and microbiota serve complementary protective functions. The microbiota confers colonization resistance and mediates pathogen clearance in primary infections, while sIgA protects from disease if the host re-encounters the same pathogen. This has implications for curing S. typhimurium diarrhea and for preventing transmission.
| Numerous pathogens infect the gut. Protection against these infections is mediated by mucosal immune defenses including secreted IgA as well as by the competing intestinal microbiota. However, so far the relative importance of these two different defense mechanisms remains unclear. We addressed this question using the example of non-typhoidal Salmonella (NTS) gut infections which can be spread in stool of infected patients over long periods of time. We used a mouse model to reveal that the intestinal microbiota and the adaptive immune system hold different but complementary functions in fighting NTS infections. A primary Salmonella infection disrupts the normal microbiota and elicits Salmonella-specific sIgA. sIgA prevents disease when the animal is infected with NTS for a second time. However, sIgA was dispensable for pathogen clearance from the gut. Instead, this was mediated by the microbiota. By re-establishing its normal density and composition, the microbiota was necessary and sufficient for terminating long-term fecal Salmonella excretion. This establishes a novel paradigm: The microbiota clears the pathogen from the gut lumen, while sIgA protects from disease upon re-infection with the same pathogen. This has implications for the evolutionary role of sIgA responses as well as for developing microbiota-based therapies for curing infected patients.
| Bacterial diarrhea is a global cause of morbidity and mortality. In most cases, the acute disease symptoms cease after a few days and the pathogen is eliminated from the gut. However, the mechanisms eliminating enteropathogenic bacteria from the gut lumen are poorly understood. Most “classical” effector mechanisms of the immune system are ineffective in the gut lumen (i.e. complement-mediated killing, opsonophagocytosis, T-cell mediated toxicity). In the gut, innate and adaptive immune responses such as antimicrobial peptides, natural and pathogen-specific mucosal secretory IgA (sIgA) antibodies are considered to be cardinal defense mechanisms. In addition to the host's immune system, the highly dense and diverse bacterial community in the gut (the microbiota; >500 different species [1], [2]) plays a key role by inhibiting pathogen growth in the gut lumen right from the beginning. This phenomenon is referred to as ‘colonization resistance’ and efficiently blocks infections by Clostridium difficile, Salmonella spp. and many other pathogenic bacteria [3]. Colonization resistance might be based on nutrient limitation, release of inhibitory metabolites, production of bactericidal compounds, the competition for binding sites and other, unidentified features of the dense microbial community [4], [5].
Much less is known about the mechanisms clearing enteropathogenic bacteria from the gut lumen once they have established an infection in this niche. ‘Pathogen clearance’ differs significantly from colonization resistance as both, the mucosa [6] and the microbiota, must recover from pathogen-inflicted disturbance while eliminating the pathogen [7]. Here, we have studied the mechanisms of pathogen clearance from the gut lumen using the example of non-typhoidal Salmonella (NTS) diarrhea.
NTS infections, including S. enterica spp. I serovar Typhimurium (S. tm), account for a significant share of food-borne diarrhea in Europe and Northern America. In sub-Saharan Africa, NTS are also an important cause of invasive disease with high mortality, particularly in HIV infected individuals [8]. In humans, colonization resistance confers partial protection, but antibiotic treatment increases the risk of Salmonella diarrhea [9], [10]. In the typical cases of NTS diarrhea, the pathogen begins to grow in the gut and disease symptoms manifest eight to 24h after consumption of contaminated food or water. Usually, the pathogen remains limited to the gastrointestinal tract and diarrhea subsides within several days. After cessation of symptoms, Salmonella remains detectable in the stool for weeks, several months or sometimes even longer [11], [12]. Pathogen clearance seems to fail in these long-term ‘asymptomatic excretors’. This is problematic, as ‘asymptomatic excretors’ pose a significant risk of transmission, in particular when food workers in restaurants or the food industry are affected [13].
So far, we can only speculate about mechanisms mediating pathogen clearance from the gut lumen. Antimicrobial peptides might be involved in some infections, but should not affect S. tm clearance, as this pathogen is particularly resistant against this type of compound [14], [15]. Antibody responses, i.e. pathogen-specific secretory IgA (sIgA), might also clear pathogens from the gut lumen. S. tm elicits profound antibody responses against LPS and protein antigens [16]. In systemic infection models antibody responses can confer some degree of protection [17], [18]. Previous work on the role of sIgA in intestinal S. tm infection yielded conflicting results. sIgA protected cultured epithelial cells from S. tm infection, but did not reduce intestinal pathogen densities [18]. Similar findings were made for the enteropathogenic bacterium Citrobacter rodentium [19]. However, the role of sIgA in pathogen clearance in models of acute Salmonella enterocolitis with high intestinal pathogen loads has not been addressed so far. Finally, we reasoned that the microbiota itself might contribute to pathogen clearance. It remained to be shown which mechanisms contribute to pathogen clearance.
We have used a Salmonella diarrhea mouse model to analyze the relative importance of sIgA and the intestinal microbiota in S. tm clearance after infection. In mice, the intestinal microbiota confers colonization resistance. Normally, <10% of mice permit pathogen growth and get mucosal inflammation upon oral S. tm infection [20]. Oral antibiotic-treatment alleviates colonization resistance and wild type S. tm grows up to very high densities in the intestinal lumen and induces mucosal inflammation (colitis) in 100% of the animals [6]. The gut inflammation allows S. tm to out-compete the microbiota thus promoting pathogen overgrowth [21]. Here, we have extended this mouse model to study pathogen clearance at later phases of the primary infection when acute mucosal inflammation has ceased. We analyzed the levels of pathogen shedding, sIgA responses and the role of the microbiota. This revealed that the microbiota plays an essential role in pathogen clearance. The implications for curing asymptomatic excretors and preventing S. tm diarrhea are discussed.
In sm-treated mice, infection with an attenuated S. typhimurium strain (S. typhimurium SL1344 sseD; termed S. tmatt; Table S1) is known to recapitulate key aspects of the early stages of human NTS diarrhea, i.e. gut inflammation 8h after orogastric exposure with infection confined to the gastrointestinal tract [22]. Symptoms of the acute gut inflammation usually decline by 5–7 days after infection [23]. In order to assess, if this model may be useful to dissect the role of pathogen-specific sIgA and the intestinal microbiota in pathogen clearance at the final stage of a primary infection we analyzed the outcome of long-term S. tmatt infections [6], [24].
We monitored S. tmatt shedding for up to 60 days after infection. S. tmatt shedding in stool began to decrease after a few days, varied extensively between different animals and lasted for 2 to 8 weeks (Fig. 1A). At 60 day p.i., S. tmatt shedding was reduced below 105 cfu/g (p<0.05 day 1 vs. day 60 p.i.). At all stages, the infection remained largely confined to the gastrointestinal tract and draining mesenteric lymph nodes (MLN) and gut inflammation subsided after 7–44 days (Fig. 1B,C). Interestingly, we observed a high incidence of ‘asymptomatic excretors’ around day 44 post infection (p.i.). These mice were characterized by a low pathological score (≤3) and high cecum pathogen loads (≥105 cfu/g stool; Fig. 1C, right panel, green symbols). This may indicate that pathogen clearance from the gut lumen is not necessary in order to resolve gut inflammation. In fact, both might be independent from each other. We concluded that this model could be useful to analyze the mechanism of pathogen clearance from the gut lumen after infection.
Next, we wanted to address if mice that had experienced a primary S. tmatt infection in our model developed an adaptive immune response that would protect against gut inflammation upon re-infection with the same pathogen. This would be a pre-requisite for functional analysis of antibody responses in pathogen clearance. Therefore, we extended the infection model as depicted in Fig. 2A (‘immunization-challenge’ protocol). Sm-treated mice were infected with S. tmatt for 39 days as in the standard protocol ( = experimental group; mock-immunization = negative control). This allowed sufficient time for recovering from acute inflammation and the generation of a S. tm-specific adaptive immune response (Fig. 1C; see also Fig. 2C,D). At day 39, the mice were treated with ampicillin to transiently suppress the microbiota and eliminate any S. tmatt which may have persisted in the gut. We then challenged the animals with wild type S. typhimurium (wt; ampicillin resistant; 200 cfu by gavage). In the mock-immunized mice, wt S. typhimurium efficiently colonized the gut and elicited acute intestinal inflammation within two days after challenge (Fig. 2B). In contrast, the S. tmatt-immunized mice were generally protected against wt S. typhimurium-inflicted disease (8/10 mice with cecal pathology score ≤3; p = 0.0099; Fig. 2B).
Bacterial surface structures and secreted proteins are dominant targets of adaptive immune responses [18], [25], [26], [27], [28], [29]. Therefore, we analyzed whether surface-protein or O-antigen specific immune responses might explain the protection of S. tmatt-immunized mice. No protection was observed against challenge with the NTS serotype S. enteritidis (S. enwt), harboring a different LPS-O-antigen, or an O-antigen deficient isogenic S. typhimurium mutant (S. tmΔO; ΔwbaP; Table S1; p>0.05 vs. colitis in mock immunized controls). Thus, S. tmatt-immunized mice mounted an adaptive immune response which protected from mucosal disease on re-infection with the pathogen in an O-antigen-dependent way.
The exquisite O-antigen specificity of protection from a second round of inflammation suggested that adaptive immunity and particularly sIgA may be the crucial mechanism not only for preventing inflammation on re-infection, but also for clearing pathogens from the gut. Therefore, we determined the kinetics of the Salmonella-specific humoral immune response by measuring specific Ig via surface staining of live, intact bacteria by flow cytometric analysis (Fig. 2C). This assay accurately differentiates S. tm specific antibodies from antibodies directed against closely related species, such as E. coli [30] (Fig.S1A–C). S. tm-specific IgM, IgG and IgA were detectable in the serum as early as 7 days post immunization. By day 14, all mice secreted S. tm -surface-specific sIgA into the gut lumen. Mucosal sIgA responses were confirmed by immunohistochemistry (Fig.S2). Salmonella antigens targeted by this strong, specific humoral immune response were analyzed by Western blotting. The antibody response was indeed pathogen-specific, as Lactobacillus reuteri RR and Enterococcus faecalis, two commensals isolated from our mouse colony, were not recognized (Fig. 2D; Fig.S3). In analogy to the human infection (Fig.S4), the antibody response included sIgA recognizing the O-antigen of S. tm (protease resistant ladder-like bands in the Western blot; Fig. 2D), a highly repetitive sugar structure of the lipopolysaccharide (LPS), coating the surface of the pathogen. In contrast, the O-antigens from S. enteritidis and E. coli, which have a different sugar structure or LPS from the O-antigen deficient mutant S. tmΔO were not recognized. In addition, antibodies to several prominent protein antigens were detected. Most of these protein antigens were conserved in different Salmonella and E. coli strains, but not in L. reuteri RR or E. faecalis.
It should be noted that acute mucosal inflammation seems necessary to elicit immune responses protecting from enterocolitis. It was also shown previously, that invasive Salmonella strains triggered more potent adaptive immune responses [31]. Mice not pretreated with sm before immunization (low antigen loads, no gut inflammation), sm-treated mice immunized with S. tmavir (high antigen loads, no gut inflammation) and parenterally immunized mice (S. tmatt i.v.; systemic antigen loads, no gut inflammation) did not mount detectable levels of O-antigen-specific sIgA. None of the mice were protected against wild type S. tm (S. tmwt) mediated enteropathogenesis (Fig.S5).
Overall, these data demonstrated that the LPS O-antigen was the dominant protective antigen and that mice mount a robust pathogen-specific sIgA response during the first round of infection. This is in line with earlier data from studies in the mouse typhoid fever model, in chicken and data from human patients [18], [25], [26], [27] (Fig.S4). However, from these first sets of experiments we could not conclude whether pathogen-specific sIgA was sufficient for S. tm clearance from the gut.
In order to address sIgA functions in pathogen clearance, we analyzed the outcome of S. tmatt infection in different KO-mice lacking key mediators of functional adaptive immune responses. We determined whether T-cell dependent or -independent mucosal sIgA immune responses [30], [32], [33] were critical for termination of inflammation, pathogen clearance and protection from inflammation on re-infection. ‘Immunization-challenge’ experiments were performed on mice lacking the T-cell receptor (TCRβ−/−δ−/−; T-cell deficient), B-cells (JH−/−), IgA (IgA−/−) or sIgA and sIgM-transport into the gut lumen (pIgR−/−; Table S2). Two days after initial infection with S. tmatt, all knockout mice displayed pronounced gut inflammation (data not shown) and gut inflammation subsided by day 40 (Table 1). This demonstrated that the acute mucosal inflammation can be efficiently terminated in the absence of T-cells, B-cells, antibodies or sIgA. Furthermore, several IgA−/− (3/4) and pIgR−/− (2/5) animals managed to clear S. tmatt from the gut lumen by day 40 p.i. This indicated that pathogens can (at least in some cases) be cleared from the gut lumen, in the absence of pathogen-specific sIgA (and sIgM) in the gut lumen. In order to exclude differences attributable to alterations in microbiota composition between different mouse lines, we have compared the S. tmatt clearance kinetics between IgA−/− and wild type littermates (IgA+/−, IgA+/−, IgA+/+; Fig. 3). This verified that kinetics of pathogen clearance was not affected by presence or absence of sIgA.
Strikingly, none of the S. tmatt -immunized knockout mice developed O-antigen specific antibodies and none were protected from intestinal inflammation upon challenge with S. tmwt (pathological score ≫3; Table 2 and Fig.S6). Thus, a T-cell dependent, adaptive mucosal sIgA response is essential for protection from secondary disease, but is dispensable for resolving the initial inflammatory response to S. tmatt and for clearing the pathogen from the intestinal lumen.
Though dispensable, our findings did not exclude that sIgA exerts an effector function [34] which contributes in some way to pathogen clearance. To identify such mechanisms, we analyzed the effects of sIgA on pathogen growth and its interaction with the host's intestinal mucosa in greater detail. First, we applied a modified ‘immunization-challenge’ protocol. Sm-treated mice were infected with S. tmatt, an equivalent S. enteritidis strain (S. enatt; S. enteritidis 125109 sseD; [35]) or mock. Antibody responses and S. tmatt/ S. enatt loads in the stool were monitored (Fig.S7 and data not shown). After 39 days, immunized mice were treated with ampicillin (elimination of microbiota and remaining S. tmatt or S. enatt) and challenged with a 1∶1 mixture of S. tmavir and S. enavir (ampicillin resistant; sseDinvG mutants; 200 cfu each by gavage; Table S1). These latter mutants can colonize the gut lumen of naïve mice for up to four days, remain confined to the gastrointestinal tract and they do not elicit enteropathogenesis, thus mimicking the situation in the intestines of ‘asymptomatic excretors’ [21], [24]. We decided not to use S. tmΔO for this type of competition experiments as it displays a pronounced competitive growth defect in mice when co-infections are performed with an isogenic wild type strain [36]. In the gut lumen of S. tmatt immunized mice, S. enavir out-competed S. tmavir (Fig. 4A; black symbols). In S. enatt immunized mice, S. tmavir out-competed S. enavir (red symbols), and in mock-immunized mice, both strains colonized with equal efficiency. Therefore, O-antigen specific sIgA may help controlling pathogen growth or survival in the gut lumen. Furthermore, S. tmavir (but not S. enavir) was aggregated in the gut lumen and occluded from the mucosal surface of S. tmatt immunized mice (Fig. 4A, right panels). Pathogen occlusion was confirmed by assessing pathogen loads in the gut tissue of challenged mice. In S. tmatt -immunized animals, S. tmwt tissue loads were 100-fold lower than in mock-immunized controls (Fig. 4B). In contrast, S. tmatt immunization did not prevent the invasion of S. enteritidis. Furthermore, S. tmatt immunized pIgR−/− mice, which cannot transport sIgA across the gut epithelium, failed to prevent gut tissue invasion by wt S. typhimurium into the mucosal tissue (Fig. 4B). Thus, the O-antigen-specific sIgA response conferred protection by restricting pathogen growth in the gut lumen and preventing the interaction of the pathogen with the intestinal mucosa. To some extent, this may also contribute to pathogen clearance from the gut lumen.
While O-antigen-specific sIgA was indispensable to prevent disease, it did not seem to be a major determinant in pathogen clearance from the gut lumen. The onset of adaptive sIgA responses and cessation of symptoms seemed to occur well ahead of S. tmatt elimination from the intestines. Moreover, IgA deficiency did not affect pathogen clearance kinetics (Table 1; Fig. 3). This was different from most well studied paradigms of acute systemic infection where the onset of protective immunity coincides with declining pathogen loads. This strongly suggested that sIgA-independent mechanisms may underlie pathogen clearance from the gut.
Thus, we hypothesized that the microbiota might play a crucial role in pathogen clearance. The microbiota is a dense bacterial community composed of approx. 500–1000 different species [9], [37]. It confers numerous beneficial effects to the host [38] including ‘colonization resistance’, i.e. a generalized interference with the growth of many pathogens in the gut of a naïve host [3]. Antibiotic treatment disrupts the normal microbiota, alleviates colonization resistance and constitutes a known risk factor for Salmonella infections in humans and mice [7], [9], [10], [21], [39]. Furthermore, the species composition of the microbiota - and by inference the degree of colonization resistance - can vary significantly between different individuals [40]. Therefore, the microbiota composition might explain why Salmonella shedding by ‘asymptomatic excretors’ can last for months or years.
In sm-treated mice, the microbiota is transiently reduced, but rapidly returns to pretreatment community composition, re-establishes ‘colonization resistance’ and ‘asymptomatic excretion’ occurs just transiently (Fig. 1C; [21], [41]). For this reason, our original infection model was not optimally suited for dissecting the differential role of the microbiota and sIgA in pathogen clearance.
To overcome this problem we used ‘L-mice’ which harbor a well defined, low complexity microbiota (L = ‘LCM mice’; [20]). L -mice are ex-germ free mice that are stably associated with the ‘Altered Schaedler Flora’ [42] comprising <20 species. The representatives with the highest abundance are ASF500 (Firmicutes; Clostridia; Clostridiales; Lachnospiraceae; unclassified_Lachnospiraceae) and ASF519 (Bacteroidetes; Bacteroidia; Bacteroidales; Porphyromonadaceae; Parabacteroides). Thus, the L microbiota resembles the conventional (C) microbiota of mice and men at broad lineages levels [43]. However, in spite of an equally high bacterial density as the C microbiota, the L microbiota does not confer colonization resistance [20]. Accordingly, S. tmatt efficiently colonized the gut lumen of L-mice at high levels (≥108cfu/g) and elicited pronounced enteropathogenesis by day 2 p.i. even without previous antibiotic treatment (Fig. 5A). After 40 days, all immunized L-mice had resolved acute inflammation, but kept on shedding S. tmatt at high levels for at least 83 days (Fig. 5A; see also below). This was not due to a defective O-antigen-specific sIgA response: sIgA responses in L-mice were as pronounced as in C-mice as indicated by the increased numbers of IgA+ cells in the cecal mucosa (Fig. 5B,C and Fig.S2) and by Western Blot analysis (Fig.S8 and Fig 2D). The strong adaptive mucosal immune response was also confirmed by gene expression profiling of the cecal mucosa (Fig. 5D, Fig.S9, Table S3). Furthermore, challenge experiments confirmed the O-antigen-specific protection from enteropathogenesis (Fig. 6). However, despite this O-antigen-specific sIgA response, high-level pathogen shedding persisted in all analyzed animals (Fig. 5A; see also below). Therefore, O-antigen-specific sIgA was insufficient for luminal S. tmatt clearance. This was in line with our hypothesis that elements of the normal, complex microbiota (which is lacking in L-mice) may play a key role in terminating fecal S. tmatt shedding.
To formally define the importance of the commensal microbiota in pathogen clearance, two groups of L-mice were infected with S. tmatt for 83 days. The first group was kept under strict hygiene isolation and shed high loads of S. tmatt until the end of the experiment (S. tmatt→L; Fig. 7A; open symbols). The second group was exposed to C microbiota at day 40 by placing C donor mice into the same cage (S. tmatt→L/C; 6 independent cages). Both groups of mice developed the typical pathogen-specific, adaptive sIgA response by day 83 p.i. (Fig. 7B,C). Upon introduction of the C donor mice, fecal shedding decreased gradually and ceased in most of the S. tmatt→L/C mice by day 83 (<105cfu/g; Fig. 7A; black symbols) but not in the S. tmatt→L group. This suggested that pathogen clearance was mediated in some way by the complex microbiota.
In order to verify microbiota-transfer, we assessed microbiota composition using high-throughput 16S rRNA gene sequence analysis (Materials and Methods). S. tmatt→L/C mice displayed significantly higher diversity than S. tmatt→L mice as well as LCM mice at day 2 and 40 after S. tmatt immunization (Fig. 7D). The rarefaction curves indicated that S. tmatt→L/C mice had acquired a microbiota of the similar complexity as the C donor mice. This was confirmed by assessing the richness (actual diversity) of the samples by calculating the Shannon index (H) and species evenness (E) as well as the Chao1 diversity estimate (Table S4). Accordingly, the taxonomy assignment confirmed that the number of bacterial taxa in the stool increased significantly in S. tmatt→L/C mice. All S. tmatt→L mice carried high loads of Enterobacteriaceae (i.e. Salmonella spp., E. coli spp.; red colors) in their stools. In contrast, no Enterobacteriaceae were detected in the stools of 3 (out of 6) S. tmatt→L/C mice and the remaining 3 animals carried low levels of this family (yellow colors, Fig. 7E). In addition, the microbiota composition was similar between all S. tmatt→L/C animals as demonstrated by hierarchical cluster analysis of eubacterial family profiles (Fig. 7E). This indicated that pathogen displacement occurs in a reproducible, stereotypic fashion and may not result from random transfer of only few members of the conventional microbiota. Most importantly, these data demonstrate that members of the conventional microbiota can upon transfer lead to the termination of sustained pathogen shedding in L-mice.
It remained unclear whether pathogen clearance was mediated directly by the microbiota or by microbiota-induced mucosal responses. Recently, it has been shown that parts of the microbiota (i.e. segmented filamentous bacteria) induce mucosal TH-17 cell responses that can protect from pathogen infection [2]. However, we did not observe differences in IL-17A or IFN gamma-producing CD4 T-cells in the MLN of S. tmatt→L and S. tmatt→L/C animals by day 83 (Fig. 8). Furthermore, we tested if total MLN cells obtained from S. tmatt→L/C animals would, upon transfer into S. tmatt→L (d.40) induce clearance of intestinal S. tmatt. However, this was not the case (Fig.8). This was in line with the notion that the microbiota may directly mediate pathogen clearance.
In this study, we have defined the contributions of sIgA and the microbiota in protecting the host from NTS infection. During the first encounter with the pathogen, the microbiota mediates at least two different protective functions, colonization resistance and pathogen clearance. The former is well established and prohibits the growth of diverse incoming pathogens, thus preventing colonization right on [3], [5], [20], [44]. Here, we identified pathogen clearance as a second protective function attributable to the microbiota. Pathogen clearance eliminates the pathogen from the gut lumen after an episode of acute infection, i.e. after Salmonella diarrhea. This differs from colonization resistance as the pathogen starts out at high density and the normal microbiota must re-grow from a state of depletion and disturbed composition (i.e. caused by the pathogen and the inflammatory response). Compared to the microbiota, an adaptive sIgA response mounted during the later stages of an acute infection contributes little to clearing the pathogen from the gut lumen. However, pathogen-specific sIgA protects from mucosal inflammation if the same pathogen is encountered for a second time. Thus, the microbiota and sIgA have complementary functions which jointly protect against enteropathogenic bacteria during the initial infection and subsequent exposure.
How does the microbiota mediate pathogen clearance? The lack of suitable assay systems has hampered addressing this question in the past. Clearly, S. tm clearance starts out in a situation where the pathogen has grown up in the gut lumen, inflicted disease and thereby slashed microbiota density, composition and/or function [45]. This situation is gradually reversed, involving the decrease of luminal pathogen loads as well as microbiota re-growth. Finally, normal microbiota composition, density and function are restored. Conceivably, some of the mechanisms conferring colonization resistance, i.e. bacteriocin production, inhibitory metabolites, oxygen depletion, receptor blocking, stimulating mucin- or antimicrobial peptide release, stabilization of the mucosal barrier, improvement of gut motility and/or nutrient limitation [5], might also contribute to different phases of pathogen clearance. Also, microbiota-mediated stimulation of the mucosal cellular immune system may be involved [2], [46] even though TH-17 mediated responses do not seem to contribute significantly to S. tm clearance, as indicated by adoptive transfer experiments (Fig. 8). The mechanisms mediating clearance and the relative importance of the microbiota, sIgA and other mucosal immune responses may differ between different pathogens or even between different strains of a given pathogen. Identifying the commensal species (or consortia) involved and the molecular mechanisms mediating pathogen clearance will be an important task for future research.
Does sIgA contribute to pathogen clearance? Pathogen specific sIgA is produced by wild type animals during the phase of pathogen clearance. O-antigen-specific sIgA led to aggregation of luminal pathogens, prevented access to the enterocyte surface and reduced net pathogen growth as indicated by a reduced competitive index. Surprisingly, this had little effect on S. tm clearance. Wild type mice and IgA−/− littermates displayed equivalent rates of pathogen clearance. Moreover, sIgA did not reduce pathogen loads in the stool, at least in the L-mice in the absence of a complex gut flora. Thus, for pathogen clearance at the end of a primary enteric S. tm infection, a pathogen-specific sIgA response is neither necessary nor sufficient.
Instead, sIgA protected from mucosal inflammation upon re-infection with the same pathogen. The LPS O-antigen was the key protective antigen of this adaptive immune response. Protection was attributable to pathogen-aggregation in the gut lumen, reduced net pathogen growth and pathogen-exclusion from the epithelial surface, thus inhibiting pathogen invasion into the gut tissue. This required sIgA transport into the gut lumen, as immunized pIgR−/− mice, which fail to transport sIgA across the intestinal epithelium, have equivalent pathogen loads in the gut mucosa as non-immunized littermates or wild type animals. Thus, pathogen specific antibodies do not seem to contribute much to protection, once S. tm has breached the mucosal barrier. This is in line with earlier work on the roles of antibody responses in systemic S. tm infection [17], [47]. In conclusion, our experiments show that O-antigen-specific sIgA responses protect against Salmonella-mediated gut inflammation upon re-infection.
It is interesting to consider the protective function of sIgA and the microbiota from an evolutionary perspective. The intestines of most animals are colonized by bacterial communities [43]. It seems safe to assume that microbiota have an evolutionary ancient function in protecting from infection. We speculate that this pertains to colonization resistance and to pathogen clearance and that both have evolved to provide protection against a broad range of pathogens. The elaborate adaptive immune system of modern mammals, including sIgA responses, evolved much later. It evolved in the presence of the protective functions provided by the microbiota, i.e. colonization resistance and pathogen clearance. The high efficiency of this microbiota-mediated protection may explain why sIgA responses have not evolved to affect the first round of infection with a given pathogen. This was simply not necessary. In contrast, evolving the sIgA response to protect in the case of repeated exposure to the same pathogen may have represented a strong benefit which cannot be accomplished by the microbiota. This evolutionary history may explain why the sIgA response contributes little during the primary infection. Anyhow, in modern mammals the microbiota and sIgA have quite different protective functions which complement each other during the initial- and subsequent encounters with a given pathogen.
An ‘unfavorable’ microbiota composition, e.g. in L-mice, can result in long term shedding. Asymptomatic NTS excretion is also observed in humans recovering from acute diarrhea. This period of asymptomatic excretion normally lasts for two to eight weeks, but may last for more than a year in a few patients. This poses a risk of transmission. In analogy to the long term shedding by L-mice, we propose that these individuals might lack some unidentified component of the normal microbiota. In L-mice, the pathogen is cleared upon transferring microbiota from a healthy donor. This may have implications for managing human long term asymptomatic excretors. Traditionally, patients are advised to adhere to strict personal hygiene and might even be isolated in order to reduce the risk of transmission. However, at the same time this deprives the patients from exposure to conventional microbiota from healthy individuals which might enhance pathogen clearance. So far, we do not know the species of the microbiota, the cellular interactions, and molecular mechanisms explaining pathogen clearance. However, the experimental systems presented in our study may provide the tools to address these important issues. Our findings provide a basis for future research on optimal management of ‘asymptomatic excretors’, NTS vaccine development and microbiota-directed therapy for acute diarrheal NTS infections.
Specified pathogen-free (SPF) wild type C57BL/6 mice, JH−/− [48] and IgA−/− [49], pIgR−/− [50] and TCRβ−/−δ−/− mice [51] (7–10 weeks old; all C57BL/6 background) were bred at the Rodent center HCI (RCHCI) under barrier conditions in individually ventilated cages (Ehret). IgA−/−, IgA+/− and IgA+/+ littermates were generated by crossing IgA−/− with C57BL/6 mice and breeding IgA+/−×IgA+/− animals. L-mice were generated by colonizing germfree C57BL/6 mice with the Altered Schaedler flora (ASF). Mice, housed in a bubble isolator, were inoculated at eight weeks of age by intra-gastric and intra-rectal administration of 107–108 cfu of ASF bacteria on consecutive days (www.taconic.com/library). Later, L-mice (C57BL/6 background) were maintained under barrier conditions in IVCs with autoclaved chow and autoclaved, acidified water. Mice with complex microbiota were never housed together with these in the same room to prevent contamination with additional commensal bacteria.
All animal experiments were approved (license 201/2004 and 201/2007 Kantonales Veterinäramt Zürich) and performed according to local guidelines (TschV, Zurich) and the Swiss animal protection law (TschG).
Salmonella infections were performed in individually ventilated cages at the RCHCI, Zurich, as previously described [52]. In brief, wild type C57BL/6 mice, JH−/−, IgA−/−, pIgR−/− and TCRβ−/−δ−/− mice were pretreated with 20mg of streptomycin (sm) by gavage. 24h later, the mice were inoculated with 5×107 CFU of S. tmatt, PBS (mock) or as indicated. ‘Challenge infections’ were performed 40 days later (or as indicated). Mice were treated with ampicillin (20mg; by gavage) and 24h later infected with a dose of 200 CFU of the respective ampicillin-resistant (pM973) challenge strain. Samples of cecal tissue were cryo-embedded, and inflammation was quantified on cryosections (5 µm, cross-sectional) stained with hematoxylin and eosin (H&E). Pathogen-colonization was assessed as described, below.
H&E-stained cecum cryosections were scored as described, evaluating submucosal edema, PMN infiltration, goblet cells and epithelial damage yielding a total score of 0–13 points [53].
Mesenteric lymph nodes (MLN), spleen and liver were removed aseptically and homogenized in cold PBS (0.5% tergitol, 0.5% BSA). The cecum content was suspended in 500ìl cold PBS and bacterial loads were determined by plating on MacConkey agar plates (50 ìg ml−1 streptomycin) as described [21]. Colonization levels of the challenge strain (carrying pM973 with an antibiotic marker) and immunization strain (S. tmatt; kmR) were determined by selective plating (100ìg ml−1 ampicillin or 30 ìg ml−1 kanamycin; levels of challenge strain: ampR-kmR Salmonella). For co-infection experiments shown in Fig. 2C, competitive indices were calculated according to the formula CI = ratio S. tm:S. encecal content/ratio S. tm:S. eninoculum.
MLN were harvested from S. tmatt→L (total of 5 mice) and S. tmatt→L/C (total of 5 mice) at day 80 post S. tmatt infection. Single cell suspensions were prepared using 100µm cell strainers and 40µg/ml DNAse (Roche). S. tmatt→L mice (at day 40 post S. tmatt infection) were injected intravenously with 3×107 cells (pooled for each of the two groups) in 200µl PBS. Fecal S. tmatt shedding was monitored for another 40 days.
Single-MLN-cell suspensions were prepared as described above. For intracellular staining of IFN-γ and IL-17A, 1×107 nucleated MLN cells were cultured for 3 h in 1 ml of RPMI 1680 supplemented with 10% heat-inactivated FCS and stimulated with PMA (5pg/ml) /ionomycin (500pg/ml). After adding 20 µg/ml brefeldin A, the cells were incubated for another 3h at 37°C. Cells were harvested and washed in ice-cold FACS buffer (PBS, 2% heat-inactivated FCS, 5 mM EDTA, and 0.02% sodium azide). Cells were resuspended in FACS buffer and stained on the surface with fluorescently-labeled antibodies for 30 min on ice. For intracellular staining of IL-17A and IFN-γ, cells were washed once and fixed/permeabilized for 10 min at room temperature using 500 µl of FIX/perm solution (FACSLyse; BD Biosciences; diluted to 2× concentration in distilled water and 0.05% Tween 20). Cells were washed once and stained with directly conjugated Abs against IFN-γ-APC (BD) and IL-17A-PE (Biolegend). Cells were then washed again and resuspended in PBS. Data were collected on a LSRII flow cytometer (BD Biosciences) and analyzed using FlowJo software (Tree Star).
The intestine was flushed with 2 ml of a washing buffer containing PBS, 0.05M EDTA pH8.0 and 66µM PMSF. Intestinal wash was briefly vortexed and centrifuged at 4°C, 30 min, 40.000 rpm (Eppendorf centrifuge). Aliquots of supernatants were stored at −80°C.
Bacteria harboring pM973 (GFP expression after tissue entry) in the lamina propria and epithelium were enumerated by fluorescence microscopy as described [22] using cryo-sections of PFA-fixed cecal tissue stained with Armenian hamster anti-CD54 (clone 3E2; stains lamina propria) antibody (Becton Dickinson), Cy3-conjugated goat anti–Armenian hamster Ig (Jackson ImmunoResearch Laboratories), DAPI (stains DNA; Sigma-Aldrich), and Alexa647-conjugated phalloidin (stains polymerized actin; Fluoprobes). We evaluated three 20 µm thick sections of the cecum per mouse and plotted for each mouse the average of the three values.
For detecting S. tm and S. en pM979 in the gut lumen in situ, cecal tissues were recovered and treated as described recently [54]. Briefly, the tissues were fixed in paraformaldehyde (4% in PBS, pH 7.4 over night, 4°C), washed with PBS, equilibrated in PBS (20% sucrose, 0.1% NaN3 over night, 4°C), embedded in O.C.T. (Sakura, Torrance, CA), snap-frozen in liquid nitrogen and stored at −80°C. Cryosections (7ìm) were air-dried for 2 h at room temperature, fixed in 4% paraformaldehyde (5 min), washed and blocked in 10% (w/v) normal goat serum in PBS for 1h. S. tm was detected by staining for 1h with a polyclonal rabbit á-Salmonella-O-antigen group B serum (factors 1, 4, 5 and 12, Difco; 1∶500 in PBS, 10% (w/v) goat serum) and Cy3-conjugated secondary goat-α-rabbit antibody. S. en pM979 expresses gfp under the control of a constitutive promoter and bacteria were detected in the green channel. F-Actin (epithelial brush border) was visualized by staining with Alexa-647-conjugated phalloidin, as indicated (Molecular Probes). Sections were mounted with Vectashield hard set (Vector laboratories) and sealed with nail polish. Images were recorded with a microscope (Axiovert 200; Carl Zeiss, Inc.), an Ultraview confocal head (PerkinElmer), and a krypton argon laser (643-RYB-A01; Melles Griot). Infrared, red, and green fluorescence was recorded confocally, and blue fluorescence was recorded by epifluorescence microscopy.
Frozen consecutive sections of spleen, liver, cecum, colon and small intestine (7ìm thick) were briefly fixed (10 min) in acetone and blocked for 30 min with phosphate buffered saline (PBS) containing 0.5% bovine serum albumin (BSA). Sections were then incubated with the primary antibody for 1 h at room temperature. Primary antibodies included: B220/CD45R (Pharmingen 553084; 1∶200) , CD4 (clone YTS191; 1∶200) and CD8 (clone YTS169; 1∶200) for T-cells (kindly provided by Rolf Zinkernagel; 1∶50), Ly-6G (Gr-1) for neutrophils (clone RB6-8C5; 1∶600), F4/80 for macrophages (Serotec MCAP 497; 1∶50), CD11c for dendritic cells (BD Biosciences 553800; 1∶100) and IgA (rat-anti-mouse IgA; Pharmingen 556969; clone C10-3; 1∶4000). Secondary antibodies and detection chromagens were applied and visualized using standard methods (see also [55]).
Statistical analysis was performed using the exact Mann-Whitney U test (Prism 4.0c). A P value of <0.05 (two tailed) was considered to be statistically significant. In mouse experiments, values were set to the minimal detectable value (10 cfu for cecum; 10 CFU for MLNs; 20 CFU for the spleen) for samples harboring “no bacteria.” Two figures (Fig. 1A and 4F) were generated using the statistical software package R. To assess the distribution of Salmonella loads in mice during the 60 day infection experiments, median and quantiles (corresponding to 0.05, 0.25, 0.75 and 0.95 probabilities) were plotted for each day or group of days. We performed a linear regression on medians and both 0.05 and 0.95 quantiles, weighted by the number of data points sampled for each day.
The OTUs abundance heatmap represents the mouse normalized OTU abundances (log2) clustered by average linkage clustering on Euclidean distances. This was generated using the function ‘heatmap.2’ from the ‘gplots’ R library.
The equivalent of 1 OD600 units/ml (where OD600 is the optical density at 600nm) of liquid o.n. cultures of S. tmΔO, Lactobacillus reuteri RR, Enterococcus faecalis, S. enwt, E.coli, S. tmwt, S. tmwt proteinase K treated (Gibco/Life Technologies; 0.4 mg/ml; 1h 57°C) or S. tm M933 was pelleted by centrifugation at 14,000×g for 2 min, and the supernatant was discarded. Cells were resuspended in Laemmli sample buffer (0.065 M Tris-HCl [pH 6.8], 2% [wt/vol] sodium dodecyl sulfate [SDS], 5% [vol/vol] β-mercaptoethanol, 10% [vol/vol] glycerol, 0.05% [wt/vol] bromophenol blue) and lysed for 5 min at 95°C. Equal amounts of the different strains and purified S. tm flagellin FliC were loaded on a 12% SDS-polyacrylamide gel and proteins wer separated by electrophoresis. Immunoblots were stained with mouse serum (diluted 1∶200 in PBS) or intestinal lavages (diluted 1∶20 in PBS) from naïve or immunized mice, goat-α-mouse-IgA HRP (Southern Biotech), goat-anti-mouse-IgG HRP (Bethyl Laboratories) and developed using an ECL kit (Amersham). The same protocol was used for the analysis of the human patient serum (dilution 1∶20 in PBS), where goat-anti-human-IgA-HRP (2050-05, NEB) and goat-anti-human-IgG-HRP (2040-05, NEB) were used as secondary antibodies.
Analysis was performed as described in [30]. 3ml LB cultures were inoculated from single colonies of plated bacteria and cultured overnight at 37°C without shaking. 1ml of culture was gently pelleted for 4min at 7,000 rpm in an Eppendorf minifuge and washed 3× with sterile-filtered PBS (1% BSA, 0.05% sodium azide) before resuspending to yield a final density of 107 bacteria per ml. Mouse serum was diluted 1∶20 in PBS (1% BSA, 0.05% sodium azide) and heat-inactivated at 60°C for 30min. The serum solution was then spun at 13,000 rpm in an Eppendorf minifuge for 10min to remove any bacteria-sized contaminants and the supernatant was used to perform serial dilutions (1∶20, 1∶60, 1∶180). 25µl serum solution and 25µl bacterial suspension were mixed and incubated at 4°C for 1h. Bacteria were washed twice before resuspending in monoclonal FITC-anti-mouse IgA (559354; BD Pharmingen), PE-anti-mouse total IgG (715-116-151; Jackson Immunoresearch Europe) and APC-anti-mouse IgM (550676; BD Pharmingen). After a further hour of incubation bacteria were washed once with PBS (1% BSA, 0.05% sodium azide) and then resuspended in PBS (2% PFA) for acquisition on a FACSCalibur using FSC and SSC parameters in logarithmic mode. Data were analysed using FlowJo software (Treestar, USA). Analysis of specific IgA in intestinal lavages was achieved using an identical protocol, using a dilution of 1∶2, 1∶6 and 1∶18 of gut wash.
The cecum tissue was excised (3 biological replicates per group), washed in cold PBS, placed in 300µl RLT-buffer (RNeasy Mini Kit, Qiagen; 1% β-Mercaptoethanol) and snap-frozen in liquid nitrogen. Total RNA was extracted with the Nucleospin RNA II kit (Macherey Nagel, Germany) and prepared for hybridization as recommended by the manufacturer (Applied biosystems, USA). Briefly, 2µg of total RNA and a T7-oligo(dT) primer were used for reverse transcription. The double-stranded cDNA was purified and converted to DIG labeled-cRNA by in vitro transcription using DIG-UTP (Roche, Germany). The cRNA was purified, fragmented and hybridized on ABI Mouse Genome Survey v2.0 microarrays for 16h. The microarray was washed and incubated with anti-DIG antibodies conjugated to alkaline phosphatase and a chemiluminescent substrate. The microarrays were scanned with the Applied Biosystems 1700 chemiluminescent microarray analyzer [56]. Normalization was achieved using the NeONORM method [57]. Significance of log2 fold changes (log2Q) were determined based on a double-log normal distribution hypothesis of signal intensities using mixture ANOVA methodology [56]. A change in the gene expression profiles was considered as significant if p<0.001. Heat maps were created according to standard methods [56]. Gene Ontology (GO) annotations were analyzed using the Panther Protein Classification System (http://www.pantherdb.org). Microarray data were deposited in the publicly available database: http://mace.ihes.fr with accession number: 2947924142.
Total DNA was extracted from cecal contents using a QIAmp DNA stool mini kit (Qiagen). Bacterial lysis was enhanced using 0.1mm glass beads in buffer ASF and a Tissuelyzer device (5 minutes, 30Hz; Qiagen). V5-V6 regions of bacterial 16S rRNA were amplified using primers B-V5 (5′ GCCTTGCCAGCCCGCTCAG ATT AGA TAC CCY GGT AGT CC 3′) and A-V6-TAGC (5′GCCTCCCTCGCGCCATCAG [TAGC] ACGAGCTGACGACARCCATG 3′). The brackets contain one of the 20 different 4-mer tag identifiers [TAGC, TCGA, TCGC, TAGA, TGCA, ATCG, AGCT, AGCG, ATCT, ACGT, GATC, GCTA, GCTC, GATA, GTCA, CAGT, CTGA, CAGA, CTGT, CGTA]. Cycling condition were as follows: 95°C, 10min; 22 cycles of (94°C, 30s; 57°C, 30s; 72°C, 30s); 72°C, 8min; 4°C, ∞; Reaction conditions (50µl) were as follows: 50ng template DNA; 50 mM KCl, 10 mM Tris-HCl pH 8.3, 1.5 mM Mg2+, 0.2mM dNTPs; 40pmol of each primer, 5U of Taq DNA polymerase (Mastertaq; Eppendorf).
PCR products of different reactions were pooled, ethanol-precipitated and fragments of ∼300bp were purified by gel electrophoresis, excised and recovered using a gel-extraction kit (Machery-Nagel). Amplicon sequencing of the PCR products was performed using a 454 FLX instrument (70×70 Picotitre plate) according to the protocol recommended by the supplier (www.454.com). PCR to detect ASF bacteria in the feces was done as described in [42].
We applied quality control to 454 reads in order to avoid artificial inflation of ecosystem diversity estimates [58]. Reads containing the consensus sequence (‘ACGAGCTGACGACA[AG]CCATG’) of the V6 reverse primer were filtered with respect to their length (200nt≤length≤300nt). Quality filtering was then applied to include only sequences containing one of the exact 4nt tag sequences and displaying at maximum one ambiguous nucleotide ‘N’. The latter criterion has been reported as a good indicator of sequence quality for a single read [41]. We identified 6,754 reads with an incorrect primer sequence, 1,155 reads shorter than 200nt, 8 reads longer than 300nt and 119 reads containing more than one ‘N’. After filtering, 140,237 reads remained (out of an initial total of 149,786 raw reads) and were processed as described below for OTU definition and chimera filtering. To estimate the reliability of sample discrimination using our primer-tagging approach, we assessed the number of reads observed to have an illegitimate 4-mer tag (i.e., different from our set of 20 tags). The sequencing plate produced a total of 141,784 quality-filtered reads from which 1,547 contained an incorrect tag (1.09%). Given that 256 distinct 4-mer tags are possible and that we used only 20 of these, the majority of sequencing or primer errors in this region are detectable. Correcting for the small fraction of undetectable errors (20/256) and division by four yields an error rate of 0.296% per single nucleotide - at the position of the tag in the primer (this includes errors during primer synthesis as well as sequencing). Since most errors are actually visible as errors, the rate of unintentional ‘miscall’ of sample identity is 0.092%.
To reduce computational time and complexity, we built OTUs using the complete filtered dataset covering all non-redundant reads from the 20 samples. Exactly identical sequences were represented by one representative only; after OTU computation, redundant sequences were taken into account for OTU abundance analysis. For subsequent taxonomy classification, we included additional quality-filtered 16S rRNA reference sequences, selected from the Greengenes database (http://greengenes.lbl.gov/Download/Sequence_Data/Greengenes_format/greengenes16SrRNAgenes.txt.gz, release 01-28-2009 [59]). This reference database is based on full-length non-chimerical sequences with a minimum length of 1100nt (in order to fully cover the V6 region of all entries). No archaeal sequences were included in the analysis.
The alignment of non-redundant reads from all mice with the reference database was performed using the secondary-structure aware Infernal aligner (http://infernal.janelia.org/, release 1.0, [60]) and based on the 16S rRNA bacterial covariance model of the RDP database (http://rdp.cme.msu.edu/; [61]). Before defining OTUs, we first removed reference sequences for which the alignment was not successful (Infernal bit-score<0). The alignment was then processed to include an equivalent amount of information from every read. To do so, we identified the consensus reverse primer sequence of the V6 region within the aligned sequence of Escherichia coli K12, as a reference. The full alignment was then trimmed from the start position (defined by the E. coli V6 reverse primer) and ended after 200nt. This also ensured a further limitation of the effect of pyrosequencing errors by trimming the 3′ end of each read, a region which is more sequencing-error prone (the trimmed and aligned reads length ranged from 152 to 231nt) [58]. Using this alignment, OTUs were built by hierarchical cluster analysis at various distances (0.01, 0.03, 0.05 and 0.10) using the ‘complete linkage clustering’ tool of the RDP pyrosequencing pipeline http://pyro.cme.msu.edu/; [61]).
In a first step, taxonomy was inferred for all reads using the stand-alone version of the RDP classifier (http://sourceforge.net/projects/rdp-classifier, revision 2.0, [62]). Taxon-level predictions were considered reliable when supported by a minimum bootstrap value of 80%. In order to predict taxonomy for each OTU, we either used any reference sequences present within a cluster, or the taxonomy of the reads present in the cluster, as predicted by the RDP classifier. To increase the resolution of the prediction, we privileged any reference sequences over the reads. For each OTU, taxonomy was inferred by a simple majority vote: if more than half of the reference sequences (or reads) present within a cluster agreed on a taxon, the OTU was annotated according to this taxon. In case of conflicts, we assigned a consensus taxon to a higher phylogenetic level for which the majority vote condition was met.
Deep pyrosequencing on the 454 platform has revealed extensive microbial diversity that was previously undetected with culture-dependent methods [63]. Nevertheless, sequencing data generated from pools of PCR products have to be interpreted carefully; limitations and biases of the PCR technique have to be taken into account. This can lead to over-estimations of microbial diversity as has been recently reported [58], [64]. Moreover, during amplification, chimerical sequences can be generated.
On such short sequences, recombination points (recombination can occur from an incompletely extended primer or by template-switching; [65]) are extremely difficult to detect. Recently, a new tool to filter noise and remove chimera in 454 pyrosequencing data has been published [64]. There, the authors suggest that because of sequencing errors, diversity estimates may be at least an order of magnitude too high. To our best knowledge, at the time of our analysis, there were no available tools to detect chimera within libraries of short 454 reads. Therefore, in order to detect chimeras we decided to compare taxonomies assigned to N-terminal and C-terminal read fragments using BLASTn. In order to ensure a reasonable alignment length and a relatively high identity to the matching reference sequences, we only analyzed reads for which both fragments had a minimum identity of 95% and a minimum bit-score of 150 (these cutoffs were selected heuristically). A given read was deemed chimeric when the taxonomies of the best hits of each half were clearly not congruent (i.e., differing at the phylum level). Our simple chimera detection method resulted in a slightly higher rate of detected chimera compared to the method of Quince et al., 2009 (∼4.5% compare to ∼3% in their example), suggesting that our approach is at least of comparable stringency [64].
|
10.1371/journal.ppat.1004811 | Plasma Membrane Profiling Defines an Expanded Class of Cell Surface Proteins Selectively Targeted for Degradation by HCMV US2 in Cooperation with UL141 | Human cytomegalovirus (HCMV) US2, US3, US6 and US11 act in concert to prevent immune recognition of virally infected cells by CD8+ T-lymphocytes through downregulation of MHC class I molecules (MHC-I). Here we show that US2 function goes far beyond MHC-I degradation. A systematic proteomic study using Plasma Membrane Profiling revealed US2 was unique in downregulating additional cellular targets, including: five distinct integrin α-chains, CD112, the interleukin-12 receptor, PTPRJ and thrombomodulin. US2 recruited the cellular E3 ligase TRC8 to direct the proteasomal degradation of all its targets, reminiscent of its degradation of MHC-I. Whereas integrin α-chains were selectively degraded, their integrin β1 binding partner accumulated in the ER. Consequently integrin signaling, cell adhesion and migration were strongly suppressed. US2 was necessary and sufficient for degradation of the majority of its substrates, but remarkably, the HCMV NK cell evasion function UL141 requisitioned US2 to enhance downregulation of the NK cell ligand CD112. UL141 retained CD112 in the ER from where US2 promoted its TRC8-dependent retrotranslocation and degradation. These findings redefine US2 as a multifunctional degradation hub which, through recruitment of the cellular E3 ligase TRC8, modulates diverse immune pathways involved in antigen presentation, NK cell activation, migration and coagulation; and highlight US2’s impact on HCMV pathogenesis.
| As the largest human herpesvirus, HCMV is a paradigm of viral immune evasion and has evolved multiple mechanisms to evade immune detection and enable survival. The HCMV genes US2, US3, US6 and US11 promote virus persistence by their ability to downregulate cell surface MHC. We developed ‘Plasma Membrane Profiling’ (PMP), an unbiased SILAC-based proteomics technique to ask whether MHC molecules are the only focus of these genes, or whether additional cellular immunoreceptors are also targeted. PMP compares the relative abundance of cell surface receptors between control and viral gene expressing cells. We found that whereas US3, US6 and US11 were remarkably MHC specific, US2 modulated expression of a wide variety of cell surface immunoreceptors. US2-mediated proteasomal degradation of integrin α-chains blocked integrin signaling and suppressed cell adhesion and migration. All US2 substrates were degraded via the cellular E3 ligase TRC8, and in a remarkable example of cooperativity between HCMV immune-evasins, UL141 requisitioned US2 to target the NK cell ligand CD112 for proteasomal degradation. HCMV US2 and UL141 are therefore modulators of multiple immune-related pathways and act as a multifunctional degradation hub that inhibits the migration, immune recognition and killing of HCMV-infected cells.
| HCMV is the prototype betaherpesvirus and an important human pathogen. Following primary infection, HCMV persists for the lifetime of the host under constant control by the host immune system. In the face of this selective pressure, HCMV has evolved multiple mechanisms to evade immune detection and has emerged as a paradigm of viral immune modulation and evasion. Experimentally, only 45 of the ~170 canonical HCMV protein coding genes are required for in vitro replication [1, 2]; most HCMV genes appear to be directed at promoting virus persistence through targeting host defenses [3–5].
Four genes clustered in the HCMV unique short (US) gene region use independent mechanisms to suppress MHC-I dependent antigen presentation to CD8+ cytotoxic T lymphocytes [6]. US3 is an immediate early gene product that binds and retains newly synthesized MHC-I proteins in the endoplasmic reticulum (ER) and blocks tapasin-dependent peptide loading [7, 8], whereas US6 inhibits TAP-mediated peptide translocation into the ER [9, 10]. US2 and US11 both bind MHC-I in the lumen of the ER and hijack the mammalian ER-associated degradation (ERAD) machinery to promote retrotranslocation to the cytosol for proteasome degradation [11, 12]. US2 and US11 appropriate distinct cellular ERAD pathways for MHC I dislocation. US2 utilizes the cellular E3 ligase TRC8 (translocation in renal cancer from chromosome 8) to ubiquitinate and subsequently degrade MHC-I [13], whereas US11 uses a Derlin-1-associated ERAD complex centered around the newly characterized TMEM129 E3 ligase [14–16]. Functionally US2 and US11 are distinct as US11 has a combined ER retention and degradation function [13, 15, 17], while US2 is unable to retain MHC-I in the ER prior to degradation, but relies on US3 for enhanced degradation. In addition to downregulating MHC-I, US2 and US3 also target the MHC-II antigen presentation pathway [18, 19]. US3 retains MHC-II molecules in the ER while US2 initiates the retrotranslocation of MHC-II DR-α chain and DM-α chain from the ER back to the cytosol for subsequent degradation.
Since endogenous MHC-I molecules constitute the chief ligands recognized by NK cell inhibitory receptors, their downregulation has the potential to render cells more vulnerable to NK cell attack. To compensate, HCMV encodes its own MHC-I homologue (UL18) and a peptide present in the UL40 signal sequence acts to stabilize and maintain cell surface expression of the NK inhibitory ligand HLA-E [20–24]. Moreover, HCMV systematically suppresses cell surface expression of ligands for NK cell activating receptors. The HCMV glycoprotein UL141 plays a major role in such protection via interaction with TRAIL death receptors, as well as CD155 (PVR, necl5) and CD112 (PVRL2, nectin-2) which are both ligands for the ubiquitous NK activating receptor DNAM1 [25–27]. In isolation, UL141 is capable of suppressing both CD155 and TRAIL-R2 cell surface expression, but an additional HCMV-encoded factor is known to be required for efficient downregulation of cell surface CD112 [25]. Moreover, while CD155 and TRAIL-R2 accumulate in the ER during the course of infection, CD112 is degraded [25].
While US2, US3, US6 and US11 were originally defined by their capacity to inhibit cell surface MHC-I expression, classical MHC molecules are not necessarily their only targets. To gain an unbiased view of cellular receptors whose expression is altered upon viral gene expression, we recently developed ‘Plasma Membrane Profiling’ (PMP), a SILAC (Stable Isotope Labelling of Amino acids in Culture)-based quantitative proteomics technique which compares the relative abundance of cell surface receptors between infected and uninfected cells, and therefore identifies the range of cell surface proteins downregulated upon viral infection [28–30]. PMP demonstrated that whereas US3, US6 and US11 specifically downregulate MHC-I, US2 targets a series of novel substrates including the NK cell ligand CD112, the anti-coagulation factor thrombomodulin and at least six integrin family members, abolishing integrin signalling, cell adhesion and migration. While US2 alone is necessary and sufficient to target most substrates, effective downregulation of CD112 requires a synergistic interaction between US2 and UL141. UL141 retains CD112 in the ER and associates with US2 resulting in CD112 dislocation across the ER membrane for proteasome degradation. We therefore propose a role for US2 that is much broader than previously appreciated, but nevertheless depends on the common activity of TRC8 for impact. Furthermore, US2 and UL141 form a multifunctional and highly adaptive degradation hub with a substrate range much wider than previously appreciated, affecting cellular processes as broad as antigen presentation, NK cell killing, cell migration and coagulation.
To determine whether the HCMV-encoded US2, 3, 6 and 11 viral gene products downregulate cell surface proteins in addition to MHC molecules, we used Plasma Membrane Profiling (PMP), an unbiased, proteomic technique which compares the relative abundance of plasma membrane proteins [28]. Plasma membrane proteins were isolated from THP-1 monocytic cells stably expressing HCMV US2, US3, US6 or US11 by sequential cell surface glycoprotein biotinylation followed by streptavidin pull-down and their relative expression was quantified by high through-put mass spectrometry. To minimise differences in sample preparation, cells were metabolically labelled prior to biotinylation by SILAC, which allows early stage sample mixing without loss of sample identity [31]. The relative abundance of plasma membrane proteins in US2, US3, US6 or US11 expressing cells versus control is plotted in Fig 1A with individual proteins represented by single dots. Those proteins whose expression is unaltered by viral gene expression accumulate in the centre, whereas left and right shifts represent proteins down- or up-regulated at the plasma membrane of viral gene expressing cells.
In US3, US6 and US11 expressing cells, the majority of plasma membrane proteins identified (370–454) were unchanged, with HLA-A, B and C allotypes of MHC-I molecules being the predominant proteins lost from the cell surface (Fig 1A; left and top right panels). US3 also showed a decrease in MHC-II and C1q complement receptor expression. In contrast, US2 altered the expression of a multitude of cell surface receptors, with thirteen new proteins showing more than a four-fold downregulation (Fig 1A bottom right panel, S1 Table). In addition to MHC-I, six integrin family members were downregulated: α1 (ITGA1, 4.9 fold), α2 (ITGA2, 10.7 fold), α4 (ITGA4, 15.3 fold), α5 (ITGA5, 3.9 fold), α7 (ITGA7, 4.3 fold) and β1 (ITGB1, 5.2 fold). Other substrates downregulated by US2 include thrombomodulin (THBD, 31.9 fold), protein tyrosine phosphatase, receptor type, J (PTPRJ, 4.9 fold) and the interleukin-12 receptor β1 (IL12RB1, 8.9 fold).
We focused on novel US2 substrates and confirmed their downregulation by flow cytometry. Indeed US2 expressing THP-1 cells showed a robust downregulation of integrins α1, α2, α4, β1, thrombomodulin, PTPRJ and IL12 receptor β1 (Fig 1B, grey line), compared to control (black line). Other cell surface molecules, including the transferrin receptor and integrin αV, remained unaffected by US2 and none of the US2 substrates were affected by US11 expression (S1 Fig), thus confirming the specificity of substrate down-regulation. AXL, integrin αM (ITGAM) and αL (ITGAL) expression was dysregulated by lentiviral transduction and not followed further.
US2 is a type I membrane protein that co-opts the cellular ERAD degradation machinery to degrade MHC-I [12]. The ER-resident ubiquitin E3 ligase TRC8 is a critical component of the US2-mediated MHC-I degradation pathway [13]. US2 recruitment of TRC8 is required for the ubiquitination and subsequent degradation of newly synthesized MHC-I, and in the absence of this ligase MHC-I is rescued back to the cell surface [13]. In a similar manner, shRNA knock-down of TRC8 in US2-expressing THP-1 cells rescued cell surface expression of integrins, THBD, PTPRJ and IL-12Rβ1 (Figs 1B dashed line and S2). The recruitment of TRC8 thus appears to be a common and essential step in the US2 pathway.
To further examine the fate of these novel US2 substrates, we focused first on the integrin family. Integrins consist of an / chain heterodimer. All integrin alpha chains downregulated by US2 share the common beta-1 chain as their binding partner (Fig 1C). Integrins require α/β dimerisation prior to transport to the cell surface and loss of either may result in ER retention of the other subunit. It was therefore important to test whether the various α subunits or the β1 chain itself constitute the primary US2 substrate. On immunoblots integrin 4, 5 and 6 expression was strongly decreased in US2-expressing THP-1 cells (Figs 2A lane 3 and S3), suggesting not only down-regulation from the cell surface but US2-dependent degradation. Indeed integrin expression was rescued by proteasome inhibition (Fig 2B lane 7) or shRNA-mediated depletion of the TRC8 E3 ligase (Fig 2A and 2B lane 4). US11-expressing cells showed no change in integrin expression (Fig 2A and 2B lane 2). Unexpectedly, the integrin 1 chain was not itself degraded in US2-expressing cells, but accumulated in its faster migrating ER-resident immature form (Fig 2A lane 3). TRC8 depletion rescued this immature species to its mature form, while expression of the control integrin 3 was unaffected by US2 (Fig 2A lanes 3 and 4).
We used [35S]-methionine radiolabeling and pulse-chase analysis to further examine how US2 affects 4 and 1 integrin maturation. Integrin 4 was rapidly degraded in US2-expressing cells with a marked reduction in its half-life from more than 1 hour to less than 15 minutes (Fig 2C), which was prevented by the use of proteasome inhibitors (Fig 2D). In contrast, in the presence of US2, the 1 integrin was neither degraded, nor was it able to mature to its higher molecular species but remained in the ER in its immature form throughout the course of the 3 hour chase (Fig 2E). Our data suggest that α integrins are direct substrates for the US2/TRC8 pathway of proteasomal degradation, whereas ER retention of the β1 integrin is likely secondary to degradation of its α integrin interaction partners.
US2 rapidly degrades its target proteins, making it difficult to ascertain whether they physically interact with US2. A truncated US2 mutant (US2ΔC'), from which the cytoplasmic tail was deleted (aa 186–199) is reported to be functionally inactive, but can still bind its MHC-I substrate [32], providing a useful tool to probe US2 interactions. We initially tested whether the US2 cytosolic domain is responsible for TRC8 recruitment, which would explain the US2ΔC' loss of function. While wild-type US2 readily binds TRC8 [13], this association is lost in the US2ΔC' mutant (Fig 2F lanes 5 and 6), explaining the loss of function phenotype. Furthermore, the US2ΔC' mutant is now found associated with, but unable to degrade integrin α4 (Fig 2G lanes 3 and 6), suggesting that US2 binds its α integrin substrate prior to recruitment of TRC8.
Ubiquitination by TRC8 triggers the US2-induced retrotranslocation and degradation of MHC-I [13]. We therefore examined whether integrin α4 is also ubiquitinated in the presence of US2. Immune precipitation of radiolabelled integrin α4 visualized a smear of ubiquitinated species in the presence but not absence of US2 (Fig 2H). These species were recovered upon denaturation and re-precipitation of integrin α4 as well as ubiquitin. Integrin α4 is thus ubiquitinated in a US2-dependent manner triggering its proteasomal degradation.
To address the functional consequence of US2-mediated integrin downregulation, we focused on the signaling properties of the most dramatically down-regulated integrin, α4β1. The importance of the α4β1 integrin in embryogenesis and disease pathogenesis derives from its role in cell adhesion and cell migration [33, 34]. Binding of integrin α4β1 to its ligands fibronectin or vascular cell adhesion molecule-1 (VCAM-1) initiates focal adhesion complex assembly and phosphorylation of paxillin [35, 36]. This phosphorylation event is specific to integrin β1 and α4 tails, and is not stimulated by other α-integrins [33–35]. Fibronectin stimulation of vector-only and US11-transduced control cells led to the expected increase in paxillin phosphorylation (Fig 3A, top panel, lanes 5 and 7), while US2 inhibited phosphorylation of paxillin as effectively as shRNA-induced depletion of the β1 integrin (lanes 6 and 8). Since the α4β1 integrin is required for macrophage chemotaxis, we examined how US2 affects cell adhesion and migration. Adhesion of US2-expressing THP-1 cells to a fibronectin substrate was completely abolished (Fig 3B) and migration of these cells was significantly reduced (Fig 3C) compared to the empty vector and US11 controls. A similar phenotype was observed in β1 integrin depleted cells. Thus US2-induced downregulation of integrin α4β1 has dramatic functional consequences, inhibiting downstream integrin-mediated signalling, cell adhesion and migration.
US2 expression alone is both necessary and sufficient for downregulation and TRC8-dependent degradation of the novel substrates. We next sought to investigate US2 function in the context of a productive HCMV infection and, in an experiment complementary to US2 single gene expression, evaluated the effect of deleting US2 from the HCMV genome using PMP. Permissive human foreskin fibroblasts (HFF) were infected with the HCMV strain Merlin (HCMV wt) or a US2 deletion virus (HCMV ΔUS2), SILAC labeled and plasma membrane proteins were isolated and quantified. Proteins whose relative abundance is higher in cells infected with HCMV ΔUS2 compared to those infected with wt-HCMV require US2 for their downregulation (left shift in Fig 4A). Of 686 plasma membrane proteins identified from HCMV-infected HFFs, four integrin family members (α2 (3.6 fold), α4 (3.2 fold), α6 (7.7 fold), and β4 (6.4 fold) (Fig 4A and S2 and S3 Tables) as well as PTPRJ were significantly downregulated in a US2-dependent manner. In addition, other immunoglobulin superfamily members that had not been identified in THP-1 cells also required US2 for their downregulation. These included immunoglobulin superfamily member 8 (IGSF8) and epithelial cell adhesion molecule (EPCAM) as well as butyrophilin subfamily 2 member A1 (BTN2A1), cadherin-4 (CDH4) and endothelin-converting enzyme 1 (ECE1) (Fig 4A and S3 Table). To validate PMP results, flow cytometry was performed on HFFs infected with either an HCMV variant encoding a UL32-GFP fusion protein or the same virus additionally deleted for the US1-11 region. Gating on GFP+ HCMV-infected cells confirmed that the US1-11 region was required to downregulate MHC-I, integrin α4 and THBD (Fig 4B). In a further test with the specific HCMV ΔUS2 mutant used for PMP, infected cells were distinguished by their downregulation of MHC-I. MHC-I downregulation is unaffected with the HCMV ΔUS2 mutant due to redundancy with the US3/6/11 gene products (S2 Table). By gating on MHC-Ilo HCMV-infected cells we confirmed that downregulation of integrin α4 and THBD was specifically dependent on US2 (Fig 4C). Thus, we confirmed that many of the key US2 targets in THP-1 cells were also downregulated in HFFs in the context of HCMV infection.
HCMV infected myeloid cells are thought to play a key role in the spreading of virus in vivo. Effective mechanisms of immune evasion will be important for enabling viral reactivation in the presence of a primed immune system. We therefore tested whether the US2-induced substrate downregulation observed in whole HCMV-infected HFFs could be replicated in differentiated THP-1 cells, and specifically how US2-induced integrin downregulation affects cell adhesion. PMA-activated THP-1 cells were infected with the endothelial-tropic HCMV strain TB40 harboring a UL32-GFP marker [37]. Since a ΔUS2 mutant of this strain was not available, we inactivated US2 function by a stable TRC8 knockdown, prior to TB40 infection. Comparing control to TRC8 knock-down, we observe a striking US2/TRC8-dependent downregulation (Fig 5A) and degradation (Fig 5B lanes 3 and 4) of integrins α2, α4, α6 and thrombomodulin in HCMV-infected cells. This downregulation is therefore similar to that observed following HFF infection with the HCMV Merlin strain (Fig 4A, 4B, and 4C). Integrin α6, which is not detected in basal THP-1 cells (Fig 5B lanes 1 and 2), was markedly induced upon viral infection and concomitantly downregulated via the US2/TRC8 pathway (Fig 5B lanes 3 and 4), suggesting a potent anti-viral role counteracted by US2. A similar upregulation upon virus infection and subsequent downregulation by US2 was observed for thrombomodulin in HFF cells (Fig 4B).
To assess the functional consequence of US2-induced integrin downregulation in HCMV infected cells, TB40-infected THP-1 cells were allowed to adhere to fibronectin, recombinant VCAM-1, collagen or uncoated tissue culture wells. TB40-infected THP-1 cells with a control knock-down showed significantly decreased binding to both VCAM-1 and collagen, but not uncoated wells, compared to TRC8 knock-down cells (Fig 5C). VCAM-1 and collagen are respectively substrates for the US2-targetted integrins α4 and α2. Binding to fibronectin—a less specific substrate for integrin α4—was unaffected by TRC8 knock-down. This lack of an effect is likely due to up-regulation of integrin αV in PMA-activated THP-1 cells; this β3-associated integrin binds readily to fibronectin but is not a US2 substrate. In conclusion we show α integrins including α2, α4 and α6 are specific US2 substrates that are degraded in a TRC8-dependent manner upon both US2 single gene expression and whole HCMV infection, thereby reducing cell adhesion of myeloid cells to a variety of substrates.
HCMV UL141 is a powerful NK cell evasion gene that downregulates NK cell ligands CD112, CD155 and the death receptor TRAIL-R2 from the cell surface [25, 26, 38]. Our previous data indicated that UL141 requires an additional unmapped HCMV function to efficiently downregulate CD112 [25]. The PMP study revealed that US2 was also able to alter CD112 expression, both independently and in the context of HCMV infection (Figs 1A and 4A and S1 Table), so we further analyzed and compared specific UL141 and US2 deletion mutants. Proteomic analysis of HFF cells infected with a HCMV UL141 deletion mutant versus wild-type HCMV confirmed UL141-dependent down-regulation of the known UL141 targets: CD155, TRAIL-R2 and CD112. TRAILR4 was identified as a novel UL141 target (Fig 6A). While expression of the US2 gene alone caused only a modest downregulation of CD112 (2.4 fold) (Fig 1A and S1 Table), in the context of whole virus infection, the observed robust CD112 downregulation was clearly dependent on both US2 and UL141 (Figs 4A and 6A and S3 Table). A further PMP experiment using a dual ΔUS2ΔUL141 HCMV deletion mutant (Fig 6B and S3 Table) confirmed many of the changes we observed using single gene deletion viruses, and showed that CD112 downregulation was most efficient in the presence of both viral genes (Fig 6A, 6B, and 6C; S3 Table). These results suggest a requirement for both US2 and UL141 for effective CD112 downregulation.
UL141 is a predominantly ER-resident viral protein which, in contrast to US2, does not actively promote proteolysis of CD155 or TRAIL-R2 but sequesters them in the ER, thus preventing their further trafficking through the secretory pathway [25, 26, 38]. We hypothesized that UL141 might retain CD112 in the ER, and promote its transfer to US2 for TRC8-dependent ubiquitination and subsequent degradation. Whereas UL141 or US2 alone caused only a partial cell surface down-regulation of CD112 within the viral context (Fig 6D; Merlin ΔUS2–3.6x; Merlin ΔUL141–2.8x), their combined action showed a more than additive effect (wt Merlin -13.3x), suggesting synergy between the two immune evasion genes. Cooperativity was also observed between TRC8 and UL141. Cell surface expression of CD112 was partially rescued by TRC8 depletion of cells infected with wild-type HCMV or HCMVΔUL141, but not the HCMVΔUS2 deletion mutant (Fig 6C). TRC8-dependent downregulation of CD112 is thus dependent on US2 within the viral context.
Two versions of CD112 are produced by differential splicing; the short α and long δ variants can be distinguished using antibodies specific for their cytosolic tails [39] (S4 Fig). Following wild-type HCMV infection, both CD112 α and δ forms were degraded (Fig 7A, lane 3 vs. lane 1). However, TRC8 depletion, or the absence of US2 (HCMV ΔUS2), rescued the CD112 δ form in both its mature Endo H resistant form (Fig 7A, CD112 δ blots, lane 11 vs. 12 and 11 vs. 13, upper bands) and its immature, ER resident, Endo H sensitive forms (lower bands). A similar pattern was seen with the CD112 α isoform, which was degraded by wild-type HCMV and restored specifically in its immature form by a HCMV US2 deletion mutant or TRC8 depletion (Fig 7A CD112 α blots, lane 3 vs. 4 and 3 vs. 5). The mature form of both the α and δ isoform was only fully restored upon combined UL141 deletion (HCMV ΔUL141) and TRC8 depletion (Fig 7A lane 3 vs. 8 and 11 vs. 16). Therefore, the ability of UL141 to retain both CD112 isoforms in the ER was only revealed in the absence of US2 or following TRC8 depletion which abrogates US2-induced CD112 degradation. In contrast, CD155 downmodulation was solely dependent on UL141 for retention in the ER, and integrin α4 required only US2 in order to be degraded during virus infection (Fig 7A, CD155 and integrin α4 blots).
Together the data indicate that US2 and UL141 co-operate to prevent CD112 cell surface expression. Indeed UL141 and US2 appear capable of interacting, as evidenced by US2 co-immunoprecipitation with UL141 (Fig 7B, lane 8). UL141 also directly associated with its substrate CD112 (Fig 7B, lane 7). This interaction is lost in the presence of US2 due to CD112 degradation, but was rescued following TRC8 depletion (Fig 7B, lanes 8 and 9). UL141 co-precipitated US2 in both the presence and absence of TRC8 indicating UL141 itself is not degraded by US2 (Fig 7B, lanes 8 and 9). The control viral protein US11 is not found in association with UL141. Collectively our results provide a remarkable example of cooperativity between two unrelated viral proteins with diverse functions.
While the retention of CD112 in the ER by UL141 is inefficient and can easily be overcome, teaming up with US2 promoted CD112 degradation via the TRC8-dependent pathway and provides an efficient mechanism of controlling expression of this cellular protein.
This study demonstrates the power of Plasma Membrane Profiling (PMP) as an unbiased approach to establish a global picture of how individual viral genes modulate the cell surface proteome. Specific antibodies have traditionally been used to determine the changes in cell surface proteins upon viral infection. This straightforward approach has proven particularly useful in tracking changes in expression of critical immune effector cell ligands (e.g. MHC-I) during the course of an infection. By design, this candidate approach is inevitably selective and cannot, therefore, provide a complete picture of the effect of virus-encoded immunomodulatory functions on the cell. By deploying PMP we were able to demonstrate the precision with which US3, US6 and US11 specifically target MHC molecules, whereas US2 was revealed to be a pleotropic modulator of cell surface receptors whose function extends beyond T cell evasion to impact on NK cell function, cell adhesion, signaling and coagulation (Fig 8).
We identified many novel cellular substrates that require US2 for their downregulation. For all targets examined, US2-mediated degradation was dependent on the TRC8 E3 ligase, indicating that HCMV-mediated appropriation of this cellular ubiquitin ligase provides a common pathway for the ER-associated degradation of US2 substrates. While US2 alone is sufficient for the downregulation of the majority of new targets, effective removal of the NK cell ligand CD112 requires co-operation between UL141, to retain CD112 in the ER, and US2 to initiate CD112 degradation. The recruitment of US2 by UL141 greatly enhances the efficiency with which CD112 is downregulated and the US2-dependant degradative pathway provides a potential conduit by which host proteins retained in the ER can be targeted for degradation, and may be exploited by other HCMV proteins. US3, like UL141, retains MHC molecules in the ER from where they are degraded by the US2/TRC8 complex [17, 18]. Whereas US3 is an HCMV immediate early gene with expression peaking at 8 hours post-infection [40], UL141 reaches maximum expression at 4–5 days post-infection [26], suggesting US2 might change substrate specificity during the HCMV life cycle. Reliance on different viral retention factors might therefore enhance the flexibility of the US2/TRC8 degradation hub which may therefore be customized towards specific requirements at different stages of the viral life cycle. Antigen presentation may be an acute problem early in viral infection, requiring US3, while NK cell killing, and the requirement for UL141, becomes critical as MHC-I levels on the cell surface decline.
While the US2/TRC8 hub induces degradation of the UL141-substrate CD112, this mechanism is not deployed against UL141’s other targets: CD155 and TRAIL-R2 [26]. Since these three main cellular targets of UL141 are all implicated in distinct intracellular signaling pathways, there may be additional benefit to the virus in retaining CD155 and TRAIL-R2 in an intracellular compartment while targeting CD112 to the proteasome. Alternatively binding to CD155 and/or TRAIL-R2 might be incompatible with US2 binding to UL141. Recent structural analysis suggests the Ig-like domain of UL141 is a structural mimic of TIGIT thus allowing CD155 binding [41], whereas the interaction between UL141 and TRAIL-R2 involves a separate, non-canonical death receptor interaction site. [42]. As it is unusual for an immunomodulator to selectively target multiple unrelated receptors, further understanding of UL141 function may highlight important aspects of the evolution of immune recognition and modulation. It will be of particular interest to gain further structural insight into the CD112-UL141-US2 complex.
The largest group of novel targets downregulated by US2 were the integrins, a large family of 18 α and 8 β chains that assemble into 24 different heterodimers. HCMV-induced integrin down-regulation was first reported for integrin α1β1 [43], and we here show that US2 down-regulates a variety of integrins including α1, α2, α4, α5, α6, α7, β1 and β4. αβ-heterodimer assembly is a prerequisite for integrin maturation and transport to the cell surface [44–46]. Our data suggest that α integrins are targeted by US2 for TRC8-induced ubiquitination and proteasomal degradation, whereas the β1 subunit is retained in the ER due to the absence of its α integrin binding partner. Whether each α chain is individually recognized by US2 and degraded, or US2 binds the shared β1 integrin, which is itself protected, but leads to the degradation of any associated α chain, remains unclear. Our preliminary experiments favour the former scenario, as alpha chains were still degraded in cells depleted of β1 integrin, suggesting that alpha chains are indeed direct targets of US2. However, we cannot exclude that, despite an effective depletion, any remaining β1 integrin, might still target alpha chains for degradation. Furthermore, no common motif in α integrins targeted by US2 has been identified. It is even less clear how US2 recognizes the broad range of substrates (MHC-I, integrins, thrombomodulin, the IL-12 receptor β1 subunit) that share no apparent structural features.
Integrins mediate cellular attachment to a wide range of extracellular proteins, and control multiple cellular functions, including morphology, migration and differentiation [47]. US2-induced α integrin degradation is predicted to have a broad impact on the physiology of HCMV infected cells. In this perspective integrin α6 is of interest as it is induced following HCMV infection and concomitantly downregulated by US2, a pattern reminiscent of antiviral proteins. Indeed, integrin α6 is essential for dendritic cell (DC) migration across the laminin and collagen IV rich basement membranes to reach the draining lymph node for antigen presentation [48, 49]. Integrin α6 specific blocking antibodies inhibit DC migration to lymph nodes [48], and US2-induced integrin degradation should therefore counteract infection-induced DC migration. This might be particularly relevant during viral reactivation from latency when the immune system is already primed to eradicate early stage infection. Additional US2-targetted integrins may also inhibit DC migration: integrin α1 and α2 are receptors for the basement membrane component collagen IV, while integrin α4 and α5 are likely required for DC reverse migration across the endothelial cell layer into the lymph [50]. Furthermore, the α4 integrin is essential for leukocyte transendothelial migration from blood into peripheral tissue and plays a prominent role in immune surveillance [47]. Indeed a monoclonal antibody targeting α4β1 and α4β7 is in clinical use for the treatment of autoimmune diseases, and reduces inflammation by preventing leukocyte extravasation into the tissue [51, 52]. The downregulation of α4β1 by US2 may thus prevent circulating virus-infected myeloid cells from responding to chemoattractants and homing.
Cell migration remains an understudied area of viral immune evasion. A global comparison of plasma membrane proteins altered upon HCMV infection showed that cell surface proteins involved in adhesion and migration are a major target for HCMV [30], and additional HCMV genes likely contribute to HCMV’s modulation of cell migration. In addition to integrins, downregulation of VCAM-1, at least eight protocadherins, five plexins and two ephrins was seen [30]. Furthermore, the actin cytoskeleton of HCMV-infected cells is heavily reorganized by UL135 which hijacks the WAVE2 complex and prevents the formation of focal adhesions [53]. Furthermore, in latent HCMV infection, UL138 likely inhibits DC migration via degradation of the multidrug transporter MRP1 which is essential for leukotriene C4 (LTC4) secretion [29, 54].
Although of potential benefit to viral immune evasion, US2-mediated degradation of cell surface receptors could also potentially contribute to the pathophysiology associated with congenital HCMV infection. Several integrins targeted by US2 (α1, α4, α5, α6 and β1) are required for placentation or fetal development [47]. Integrin α4 deficiency for example is embryonically lethal in mice due to cardiac defects and defective placentation [55]. Depending on the cell types infected in pregnancy, US2’s ability to downregulate integrin family members could contribute to the fetal damage associated with HCMV infection. Indeed, HCMV-induced downregulation of integrin α1β1 has been associated with impaired cytotrophoblast invasion and placentation [56].
The endothelial cell surface protein thrombomodulin (THBD) is another US2 substrate that might contribute to HCMV-associated fetal defects. THBD alters thrombin’s substrate specificity from pro-coagulant and pro-inflammatory to anti-coagulant and anti-inflammatory [57] and its deficiency leads to lethal consumptive coagulopathy in embryonic blood vessel endothelium. Since, the endothelium is a common site of HCMV infection in vivo, US2-induced THBD degradation may contribute to both the coagulopathy and severe fetal thrombotic vasculopathy seen in congenital HCMV infection. THBD is also a marker for a human DC subset proficient in antigen cross-presentation, with similarities to mouse CD8+ DCs [58, 59]. While the exact role of THBD in these DCs is unknown, THBD regulates substrate affinity of the TLR4 co-receptor CD14 [60], suggesting a potential broader role in antiviral immunity, which may explain its induction in HCMV-infected cells and concomitant downregulation by US2.
Our study emphasizes the key role of US2 in combating different HCMV host defense pathways through the downregulation of multiple cell surface receptors. As we are particularly interested in HCMV evasion of the cellular immune response, this study focused on plasma membrane proteins. Additional US2 substrates may yet be identified by analysis of ER-resident proteins. Indeed, as shown for UL141, other viral proteins which retain host receptors in the ER may also cooperate with US2 to degrade their cargo through the common pathway of TRC8-dependent degradation.
THP-1 cells and HFF cells (System Bioscience) were grown in RPMI-1640 or DMEM respectively (PAA), with 10% heat-inactivated fetal bovine serum (FCS; PAA) and penicillin/streptomycin (pen/strep, Sigma).
The HCMV strain Merlin is designated the reference HCMV genome sequence by the National Center for Biotechnology Information [61] and is available as a BAC clone [62]. Merlin BAC derived clone Merlin wild-type used for this study contains point mutations in RL13 and UL128, enhancing replication in fibroblasts [62]. Generation of virus recombinants and stocks was described previously [62]. All recombinants were validated by whole genome Illumina sequencing (S4 Table). Merlin ΔUS2 has a deletion of the US2 ORF; Merlin ΔUL141 has deletions of the UL141 ORF; GFP tagged Merlin ΔUL16,18 ΔUS1-11 has deletions of the UL16, UL18, US1-11 ORFs and contains a UL32-GFP fusion; GFP tagged Merlin ΔUL16,18 (Merlin delta UL16/UL18, UL32-GFP) was described previously [25]. HFF cells were infected with HCMV at indicated MOI for 72h.
The GFP-tagged endothelial-tropic HCMV TB40 strain was originally created by Christian Sinzger (University of Ulm, Germany) [37] and a kind gift from Mark Wills (University of Cambridge, UK). It contains an intact US1-11 region and a UL32-GFP fusion. For THP-1 infections, THP-1 cells were starved for 12-16h in RPMI with 2% FCS, activated for 48h with 100ng/ml PMA (Sigma) and infected with TB40 at MOI 25. This resulted in infection of >95% of cells as estimated by GFP positivity and MHC-I down-regulation. Infected cells were harvested at 96h post-infection.
A C-terminal myc-tagged UL141 (HCMV Merlin strain) and HA-tagged human integrin α4 were cloned into the pHRSIN lentivirus vector with a hygromycin B selection cassette. Untagged US2, US2 with a deletion of the C-terminal cytoplasmic tail (aa1-186; US2ΔC'), US3, US6 and US11 were cloned into a pHRSIN lentivirus vector with an IRES CFP and puromycin selection cassette. N-terminal HA-tagged US2 (HA-US2) was cloned into pHRSIN with a puromycin cassette only. The pHRSIN lentivirus expression system was used as described previously [63].
For shRNA-mediated knockdown of TRC8 and integrin β1 expression, hairpin oligonucleotides were designed as described [13, 64], annealed, cloned into the pHR-SIREN lentiviral vector (a gift from Greg Towers, UCL, London). Lentivirus was produced as previously described in 293ET cells and used to transduce THP-1 cells.
Primary antibodies used for flow cytometry were: mouse α-conformational MHC-I (W6/32), mouse α-conformational HLA-A2 (BB7.2), mouse α-integrin β1 (Biolegend), mouse α-integrin α1, (BD), mouse α-integrin α2, (BD), mouse α-integrin α4 (BD), rat α-integrin α6 (Biolegend), rabbit α-integrin αV (Santa Cruz), mouse α-thrombomodulin (BD), mouse α-PTPRJ (Medical & Biological Laboratories), mouse α-IL12 receptor β1 (BD), mouse α-CD112 (Santa Cruz), mouse α-CD155 (Abcam) and FITC-conjugated mouse α-transferrin receptor (CD71; BD). AlexaFluor 647 conjugated goat anti-mouse (Life Biosciences) was used as a secondary antibody. Antibodies used for immunoblotting were: mouse α-MHC-I (HC10), rabbit α-calreticulin (Thermo), mouse α-β-actin (Sigma), mouse α-HA (Sigma), rabbit α-integrin α 4, α 5, β1, β3 (Integrin antibody sampler kit; Cell Signaling), rabbit α-Integrin α6 (Cell Signaling), goat α-CD112 (R&D Systems), rabbit α- CD112 δ form (Abcam), rabbit α-CD112 α form (LifeSpan Biosciences), mouse α-CD155 [26], mouse α-HCMV IE antigen (Argene), mouse α-HCMV US2 (a kind gift from Jack R. Bennink, NIH, US), rabbit α-HCMV US11 (a kind gift from Emmanuel Wiertz, University of Utrecht, Netherlands), mouse α-HCMV UL141 [26], mouse α-paxillin (BD) and rabbit α-phospho-paxillin Tyr118 (Cell Signaling). Antibodies used for immune precipitation were: mouse α–HA and α-myc agarose affinity gel (Sigma), mouse α-integrin β1 (Biolegend) and mouse α-ubiquitin (FK1; Millipore) in combination with protein A or protein G-conjugated sepharose (Sigma).
For siRNA gene depletion, cells were transfected using Oligofectamine (Invitrogen) at a final concentration of 40 nM and harvested at 96 h post transfection. The following siRNA oligonucleotides were used (Dharmacon, ON-TARGET PLUS):
CD112 siRNA-1, 5’-GCGCUGAGCAGGUCAUCUU-3’;
CD112 siRNA-2, 5’-GCAUGAGAGCUUCGAGGAA-3’.
For immunoblots, cells were lysed in TBS (pH7.4) with 1% NP40, 0.1% SDS, 5mM IAA, 0.5mM PMSF (Sigma) and 1X complete protease inhibitor (Roche) for 30 min on ice. For immunoprecipitations, cells were lysed in 1% digitonin (Calbiochem) in TBS with 5mM IAA, 0.5mM PMSF (Sigma) and 1X complete protease inhibitor (Roche) for 30 min on ice. Lysates were cleared of cellular debris by centrifugation and pre-cleared using IgG-sepharose (GE Healthcare). Individual proteins were immunoprecipitated using indicated antibodies in combination with Protein A or G sepharose (Sigma), washed extensively and eluted in SDS reducing sample buffer. All samples were heated for 10 min at 50°C, separated by SDS/PAGE and transferred to PVDF membranes (Millipore). Membranes were probed with the indicated antibodies, and reactive bands were visualized with Supersignal West Pico or West Dura (Thermo Fisher Scientific).
Cells were starved for 20 min in methionine-free, cysteine-free medium (Sigma), labeled with [35S]methionine/[35S]cysteine (Amersham) for the indicated time and then chased in medium containing an excess of cold methionine and cysteine (Sigma) at 37°C. Samples taken at the indicated time points were lysed in 1% Triton X-100/TBS with 5mM IAA, 0.5mM PMSF (Sigma) and 1X complete protease inhibitor (Roche) for 30 min on ice. Immunoprecipitations were performed as above.
For visualization of ubiquitinated integrin α4, primary α-HA immune precipitations from 35S-labelled cells were eluted at 50oC for 15min in 50μl TBS, containing 1% SDS. Eluates were taken off the beads and after addition of 20μM DTT fully denatured at 70oC for 10min to dissociate interacting proteins. SDS was quenched by the addition of 1ml 1% Triton X-100/TBS with IAA/PMSF/protease inhibitor followed by secondary α-HA or α-ubiquitin immune precipitation.
For cell adhesion assays, 96-well tissue culture plates were coated with 20 μg/ml fibronectin (Sigma), 20 μg/ml recombinant VCAM1/Fc (R&D Biosystems) or 200ug/ml collagen (Sigma) in PBS, for 16 h at 4°C and blocked with 0.5% bovine serum albumin (BSA) in PBS for 2 h at 37°C to block nonspecific binding. THP-1 cells expressing single HCMV genes were starved 3 h in serum free RPMI prior to assays, seeded on fibronectin-coated and uncoated plates for 1 h at 37°C and washed 3 times with PBS. Adherent cells numbers were quantified by CyQUANT-NF (Invitrogen) according to the manufacturer’s protocol.
HCMV TB40 infected THP-1 cells were harvested at 96h post-infection using enzyme-free cell dissociation buffer (Life Biosciences), counted and re-seeded on fibronectin, VCAM1, collagen or uncoated plates. Following 1 h (fibronectin, VCAM1, uncoated) or 2 h incubation at 37°C, wells were washed 5 times with PBS, containing 0.5% BSA. Adherent cell numbers were quantified as above.
For cell migration assays, Costar Transwell 24-well plates with 8 μm pore size (Thermo) were coated with 20 μg/ml fibronectin (Sigma) in PBS, for 16 h at 4°C. Cell migration was performed in RPMI with or without the chemotaxis agent MCP-1 (Sigma) at 10 ng/ml and 10% FBS in the lower chamber. After 6 h incubation at 37°C, cells migrated to the lower chamber were counted using a Neubauer counting chamber. All cell adhesion and migration assays were performed in triplicate and p-values were calculated using paired Student’s t-test based on independent experiments.
Plasma membrane profiling was performed as described previously for THP-1 cells, with minor modifications for adherent HFFs [28, 29]. Briefly, for THP-1 cells, 1.5 × 108 of each SILAC-labeled cell type were pooled in a 1:1 ratio. Labeling was as follows: (first experiment, Fig 1A, top left and bottom right panel) THP-control (light label); THP-US2 (medium label); THP-US11 (heavy label). (Second experiment, Fig 1A, top right and bottom left panel) THP-US6 (light label); THP-control (medium label); THP-US3 (heavy label). Surface sialic acid residues were oxidized with sodium meta-periodate (Thermo) then biotinylated with aminooxy-biotin (Biotium). The reaction was quenched, and the biotinylated cells incubated in a 1% Triton X-100 lysis buffer. Biotinylated glycoproteins were enriched with high affinity streptavidin agarose beads (Pierce) and washed extensively. Captured protein was denatured with DTT, alkylated with iodoacetamide (IAA, Sigma) and digested with trypsin (Promega) on-bead overnight. Tryptic peptides were collected and fractionated (described below). Glycopeptides were eluted using PNGase (New England Biolabs) then desalted using StageTips [65].
For HFF cells, one 150cm2 flask of HCMV-infected HFFFs per condition was washed twice with ice-cold PBS. Labeling was as follows: (first experiment, Fig 4A) HFF/Merlin-wt (light label), HFF-Merlin ΔUS2 (medium label). (Second experiment, Fig 6A) HFF/Merlin-wt (medium label), HFF-Merlin ΔUL141 (heavy label). (third experiment, Fig 6B) HFF/Merlin-wt (light label), HFF-Merlin ΔUL141ΔUS2 (medium label). Sialic acid residues were oxidized with sodium meta-periodate (Thermo) then biotinylated with aminooxy-biotin (Biotium). The reaction was quenched, and the biotinylated cells scraped into 1% Triton X-100 lysis buffer. Biotinylated glycoproteins were enriched and digested as described above.
HpRP-HPLC was performed on tryptic peptides as described previously [28]. 90% of each tryptic peptide sample was subjected to HpRP-HPLC fractionation using a Dionex Ultimate 3000 powered by an ICS-3000 SP pump with an Agilent ZORBAX Extend-C18 column (4.6 mm x 250 mm, 5 μm particle size). Peptides were resolved using a linear 40 min 0.1%-40% acetonitrile gradient at pH 10.5. Eluting peptides were collected in 15s fractions. For THP-1 experiments, a total of 30 combined fractions were generated then dried using an Eppendorf Concentrator for LC-MSMS using a NanoAcquity uPLC (Waters, MA, USA) coupled to an LTQ-OrbiTrap XL (Thermo, FL, UA). MS data was acquired between 300 and 2000 m/z at 60,000 fwhm with CID spectra acquired in the LTQ with MSMS switching operating in a top 6 DDA fashion. Fractionated HFF experiments were re-combined to give either 40 or 10 fractions. The 40 fraction experiments were acquired using the OrbiTrap XL as above. The 10 fraction experiments were acquired using a Q Exactive (Thermo) coupled to a RSLC nano3000 (Thermo) with MS data acquired between 400 and 1650 m/z at 75,000 fwhm with HCD fragment spectra acquired in a top 10 DDA fashion.
Raw MS files were processed using MaxQuant version 1.3.0.5. [66, 67]. Data were searched against concatenated Uniprot human and HCMV databases, and common contaminants [67]. Fragment ion tolerance was set to 0.5 Da with a maximum of 2 missed tryptic cleavage sites. Carbamidomethyl cysteine was defined as a fixed modification, oxidised methionine, N-terminal acetylation and deamidation (NQ) were selected as variable modifications. Reversed decoy databases were used and the false discovery rate for both peptides and proteins were set at 0.01. Protein quantitation utilised razor and unique peptides and required a minimum of 2 ratio counts and normalized protein ratios reported. Peptide re-quantify was enabled in all analyses. Summed intensity represents the sum of all heavy and light labelled peptide intensities for a given protein and was calculated by MaxQuant [68]. Significance B values were calculated and Gene Ontology Cellular Compartment (GOCC) terms added using Perseus version 1.2.0.16 (downloaded from http://maxquant.org). Significance B identifies the significance of outlier protein ratios from the distribution of all ratios with a greater significance given to proteins with a high intensity [68]. Ratios were excluded for proteins identified by <2 unique peptides, or with a variability >150%. We assessed the number of PM proteins identified as described previously [28].
|
10.1371/journal.pbio.2000841 | Nonlatching positive feedback enables robust bimodality by decoupling expression noise from the mean | Fundamental to biological decision-making is the ability to generate bimodal expression patterns where 2 alternate expression states simultaneously exist. Here, we use a combination of single-cell analysis and mathematical modeling to examine the sources of bimodality in the transcriptional program controlling HIV’s fate decision between active replication and viral latency. We find that the HIV transactivator of transcription (Tat) protein manipulates the intrinsic toggling of HIV’s promoter, the long terminal repeat (LTR), to generate bimodal ON-OFF expression and that transcriptional positive feedback from Tat shifts and expands the regime of LTR bimodality. This result holds for both minimal synthetic viral circuits and full-length virus. Strikingly, computational analysis indicates that the Tat circuit’s noncooperative “nonlatching” feedback architecture is optimized to slow the promoter’s toggling and generate bimodality by stochastic extinction of Tat. In contrast to the standard Poisson model, theory and experiment show that nonlatching positive feedback substantially dampens the inverse noise-mean relationship to maintain stochastic bimodality despite increasing mean expression levels. Given the rapid evolution of HIV, the presence of a circuit optimized to robustly generate bimodal expression appears consistent with the hypothesis that HIV’s decision between active replication and latency provides a viral fitness advantage. More broadly, the results suggest that positive-feedback circuits may have evolved not only for signal amplification but also for robustly generating bimodality by decoupling expression fluctuations (noise) from mean expression levels.
| A central and recurring feature of cell-fate regulating circuits is their ability to generate bimodal expression—2 alternate expression states that exist simultaneously—with each state corresponding to a different cell fate. To understand the mechanisms enabling bimodality in a natural decision-making circuit, we examined HIV’s fate-selection circuit, the Tat circuit. This bimodal circuit is sufficient and necessary to generate a bet-hedging decision between 2 alternate HIV fates: active viral replication and long-lived dormancy (proviral latency). The dormant state, which is resistant to most antiviral drugs, is the primary clinical barrier to curing an HIV infection. While the canonical role of positive feedback is to amplify a signal, surprisingly, we find that the HIV transactivator of transcription (Tat) positive-feedback architecture is instead optimized to expand the regime of HIV expression bimodality. From an evolutionary perspective, the results suggest that positive-feedback circuits may have evolved to robustly generate bimodality in certain contexts, and, given the rapid evolution of HIV, the presence of a circuit optimized to robustly generate bimodal expression patterns appears to support the hypothesis that HIV’s active-versus-latent decision confers viral fitness.
| Bimodality is a recurring feature in many biological fate-selection programs [1], such as the HIV active-versus-latent decision (Fig 1A). Bimodal expression is a population-wide distribution pattern comprises 2 gene-expression modes, each corresponding to a specific fate path [2]. The mechanisms that can generate bimodal phenotypes have long been studied, and the architecture of underlying gene-regulatory circuits appears to be a key driver of bimodality [3–11]. Classically, bimodality has been associated with deterministic bistability in gene circuits [12–15]. Deterministic bistability requires ultrasensitive input-output relations and can result from nonlinear positive feedback (i.e., Hill coefficient > 1) on a constitutively expressed promoter [16,17]. However, many promoters are nonconstitutive and instead toggle between inactive and active expression states, generating episodic bursts of mRNA production (for review, see [18]). The finding that promoters undergo episodic bursts of expression led to a proposal that this toggling alone could generate bimodality without deterministic bistability. Unlike constitutive expression, toggling increases the degrees of freedom in a system [19], and if promoter toggling occurs relatively slowly, the resulting expression bursts can potentially produce bimodality independent of ultrasensitivity [19,20]. However, the promoter toggling kinetics required to generate bimodality appeared to be in a small portion of the experimentally observed regime [18,21–23], with experimental measures of intrinsic promoter toggling exhibiting kinetics that are typically too fast to produce bimodal expression patterns (Fig 1B)—specifically, the measured promoter toggling rates were greater than the per capita protein and mRNA decay rates [18,24,25]. Nevertheless, synthetic positive-feedback circuits that slowed toggling could induce bimodality [26]. Thus, while computational models showed that promoter ON-OFF toggling was sufficient for bimodal expression [20] and synthetic transcriptional circuits lacking bistable feedback could generate bimodal expression [26], it remained unclear how natural biological circuits exploit this mechanism to generate bimodality without bistability. Here, we determine if promoter toggling can intrinsically generate bimodal distributions in a natural biological system (i.e., HIV) and the potential physiological relevance.
We focus on HIV as a physiological model system for expression bimodality driving a decision-making process (Fig 1A). Upon infection of a CD4+ T lymphocyte, HIV undergoes a fate-selection decision, either actively replicating to produce viral progeny and destroy the host cell or entering a long-lived quiescent state called proviral latency [28,29]. A viral gene-regulatory circuit is both necessary and sufficient to drive HIV fate selection [10]. At the core of this decision-making circuit is a virally encoded transcriptional positive-feedback loop comprises a single HIV protein—the transactivator of transcription (Tat)—that amplifies expression from the virus’s only promoter, the long terminal repeat (LTR) promoter. Molecularly, this positive-feedback loop functions because the LTR is a relatively weak promoter, in the absence of Tat, with RNA polymerase II (RNAPII) elongation stalling approximately 69 nucleotides after initiation [30]. Tat transactivates the LTR by binding to a short, approximately 69-nucleotide-long RNA-hairpin loop called the Tat-activation RNA (TAR) loop and recruiting the positive transcriptional elongation factor b (pTEFb)—principally composed of CDK9 and cyclinT1—which hyperphosphorylates the carboxy-terminal domain (CTD) of RNAPII, thereby relieving the RNAPII elongation block [30,31]. Thus, Tat acts much like a bacterial antiterminator enhancing transcriptional elongation rather than initiation.
Importantly, minimal LTR-Tat positive-feedback circuits are sufficient to generate bimodal expression patterns [32], and in the full-length viral context, this circuit is both necessary and sufficient to drive HIV’s active-versus-latent decision [27]. There are 2 specific quantitative features of the Tat-LTR feedback circuit that are curious, given its obligate role in viral fate selection. First, unlike many other positive-feedback circuits that control phenotypic decisions [33,34], the Tat positive-feedback loop is noncooperative (Hill coefficient ≈ 1) and not deterministically bistable [35]. Second, the LTR promoter itself displays large episodic expression bursts toggling between ON and OFF states at virtually all integration sites throughout the human genome [24,36,37], raising the possibility that the LTR itself may be sufficient to generate bimodal expression patterns independent of Tat feedback.
In this study, we construct minimal circuits to examine if the LTR itself is capable of generating bimodal expression patterns in the absence of Tat feedback and then computationally examine the precise role of Tat positive feedback in bimodality. The results indicate that the LTR is intrinsically capable of generating bimodal ON-OFF expression even in the absence of feedback but that Tat feedback shifts and expands the regime of LTR bimodality into physiological ranges by slowing LTR toggling. In fact, the architecture and parameters of the Tat circuit appear optimized to robustly generate bimodal expression. Given the rapid evolution of HIV, the presence of a circuitry that appears optimized to slow promoter toggling and generate bimodality may be consistent with the hypothesis that the circuit has been selectively maintained and that bimodal expression (between active replication and latency) provides a viral fitness advantage [38].
Previous studies demonstrated that Tat positive feedback can generate bimodal expression patterns from the HIV LTR [32]. However, given the large, episodic bursts of expression that characterize LTR activity [24,36,37], we set out to test if the LTR was capable of bimodal expression, even in the absence of feedback (i.e., whether feedback was dispensable for bimodality, possibly having an orthogonal function in HIV). Analysis of experimental and computational literature reports indicated that the regime for generating bimodality through promoter toggling alone fell outside the experimentally observed values of LTR toggling but that slightly slower LTR toggling transitions might generate bimodality without feedback (Fig 1 and S1 Fig).
To test this prediction that Tat feedback was dispensable for bimodality, HIV circuitry was refactored to split the Tat positive-feedback loop [27] into open-loop parts (Fig 2A). This minimal circuit system allows Tat concentrations to be modulated by doxycycline (Dox) and Tat protein stability to be tuned through Shield-1 addition [27]. As Tat is fused to Dendra, the Tat concentrations can be quantified, while LTR activity is simultaneously tracked in single cells. This open-loop doxycyline-inducible circuit was integrated into T cells by viral transduction, and cells were exposed to varying concentrations of activator (Dox) and Tat proteolysis inhibitor (Shield-1)—generating approximately 48 unique unimodal Tat inputs to the LTR (S2 and S3 Figs and S1–S23 Data). Expression profiles from the LTR are all unimodal in the absence of Tat (S2 Fig), in agreement with previous findings [32,36,37]. However, in striking contrast, the presence of Tat induces bimodality from the LTR despite the lack of cooperativity or feedback in this open-loop system (Fig 2B, S2 and S3 Figs and S1–S23 Data). In other words, despite a fixed, unimodal concentration of active Tat transactivator, bimodal LTR distributions can be generated, and single-cell time-lapse microscopy confirms that the activity of the LTR is dependent on Tat input (S4 Fig and S24 Data). From the known requirements for bimodality to arise from a toggling promoter (Fig 1), the data suggest that LTR toggling becomes sufficiently slow in the presence of Tat to produce bimodal expression patterns, even in the absence of positive feedback.
The bimodality in the minimal open-loop system (Fig 2) represents the 2 fate paths of the virus—active replication and proviral latency [40]—and suggests that positive feedback may also be dispensable for controlling viral fate in full-length HIV. Importantly, results from a Tat-deficient full-length HIV virus [27], where Tat is introduced in trans (S5 Fig), confirm that Tat feedback is not required to select between alternate HIV fate paths. Thus, unlike other decision-making circuits [17,26], fate selection can occur independent of positive feedback or cooperativity in HIV.
To understand the molecular mechanisms enabling LTR bimodality in the absence of feedback, we used a validated computational model of HIV [27] and adapted it to an open-loop system where Tat would either modulate (1) burst frequency alone, kON modulation; (2) burst frequency and burst size, kOFF modulation; or (3) burst size alone by affecting transcriptional efficiency, α modulation (top of Fig 2C, S1–S3 Tables and S1 Data). To model Tat modulation of kOFF alone, a third promoter state, termed Tat-LTRON, was added such that it maintained the same transcriptional efficiency, α, as the LTRON state. Thus, the transactivated LTR promoter must first transition from Tat-LTRON to LTRON and only then can it transition from LTRON to LTROFF and fully turn off. This third promoter state, Tat-LTRON, is necessary to generate changes in burst sizes without altering transcriptional efficiency or toggling from the LTROFF to LTRON state. The model results are consistent with previous findings that bimodality is not induced through frequency modulation of the LTR (i.e., kON modulation) or increases in burst size through transcriptional efficiency, α [24,36,37]. However, the model shows that slowing toggling kinetics, or increasing the dwell time in the LTRON and LTRTatON states (i.e., kOFF modulation), is required for bimodality, and if Tat only affects a single parameter, kOFF modulation is necessary and sufficient (bottom of Fig 2C, S6 Fig and S1 Data).
The interpretation of these results is that, while natural LTR promoter toggling is too quick to generate large enough expression fluctuations for bimodality, Tat transactivation is able to slow the kinetics of toggling, expanding the bimodal regime (Fig 1). The slowing of toggling kinetics reinforces the findings that Tat stabilizes transient pulses of expression from LTR fluctuations [40], by effectively reducing kOFF. If Tat does stabilize pulses of expression to control gene-expression variability, then the prediction is that altering Tat-feedback strength would, similar to the open-loop system, control the shape of the gene-expression distribution and bimodality.
To test the prediction that Tat-feedback strength shapes the expression distribution, we used a synthetic Tat circuit [27] where positive-feedback strength could be manipulated pharmacologically by the addition of a small-molecule, Shield-1, that stabilizes Tat proteolysis (Fig 3A). In this system, a subset of isoclonal cell populations carrying this synthetic circuit naturally generate bimodal distributions (Fig 3B and S25–S29 Data). These clonal differences are mainly due to the genomic location of HIV integration, which can dictate the transcriptional bursting parameters, and the effectiveness of Tat transactivation [24,36]. Though the differences in Tat transactivation potential are not clear, transcriptional parameters of the LTR in the absence of feedback vary due to promoter methylation status, nucleosome acetylation and methylation state, or gene-proximity dependencies [41]. When positive-feedback strength is increased, a significant fraction of the cells generate bimodal distributions and even convert from a unimodal (low peak) into a bimodal (low and high peak) distribution or from a bimodal (low and high peak) to a unimodal (high peak) distribution (Fig 3B, S7 Fig and S25–S30 Data).
Importantly, simulations of Tat positive-feedback circuitry corroborate this phenomenon of bimodal expression at intermediate feedback strength if Tat acts by decelerating LTR toggling kinetics (Fig 3C and S26 Data), in agreement with simulations of the open-loop circuit (Fig 2). Thus, these simulations indicate that Tat-feedback strength likely alters the natural LTR toggling kinetics set by the local integration site [42] to control HIV bimodal-expression patterns. To test if Tat feedback in fact extends pulses of expression (i.e., effective kOFF reduction), HIV gene-expression was activated to a high-expression state, using tumor necrosis factor alpha (TNFα), and the circuit was then allowed to relax back to the unperturbed state under varying feedback strengths. TNFα enhances HIV expression by stimulating recruitment of a p50-RelA heterodimer to nuclear factor kappa-light-chain-enhancer of activated B cells (NFkB) binding sites within the LTR [42]. The cells were exposed to TNFα for 24 hours and then allowed to relax back in the presence of strong or weak feedback (S8 Fig). The results show that increasing feedback strength, by dosing cells with increasing amounts of Shield-1, increases the transient in the expressive states, leading to slower transitions from ON to OFF states (S8 Fig and S31 Data), which corroborates previous findings [27,40]. Thus, relaxation to various baseline states is dictated by feedback acting on promoter toggling.
One simplifying assumption in the model is that Tat only modulates a single bursting parameter. To test how relaxing this assumption affects bimodal generation, new simulations in which Tat could modulate multiple bursting parameters were performed. The models allow Tat to alter both burst size and frequency through kON and kOFF, kON and α, or kOFF and α modulation (S9 Fig). Interestingly, the simulations show that any combination of parameters could yield bimodality (S9 Fig). In each scenario, Tat positive feedback yields nonexponentially distributed “OFF” times and slows toggling kinetics. This result is in agreement with the previous findings that slowing promoter toggling kinetics yields bimodal distributions (Figs 1–3 and S6 Fig).
A few alternate explanations are possible for the observed bimodality. The first is that the bimodality may arise from deterministic cell-to-cell variability [43] where the transcriptional parameters vary between cells, leading to bimodality. However, these minimal circuits display a high level of ergodicity [24,40], suggesting the cell-to-cell variability in the transcriptional parameters is minimal. Second, HIV feedback may be bistable (i.e., exist in 1 of 2 stable states [high or low] [17]). Bimodality observed from bistable circuits results from fluctuations around latching feedback strengths (S11 Fig). Previous studies analyzing fluctuations in noise to measure feedback strength, cooperativity in feedback, or stability of the “ON” state found that HIV feedback lacks the canonical features of bistability [34,35,40]. Last of all, HIV feedback may latch, meaning small increases in Tat would be drastically amplified to saturable levels upon which the system would then latch in a high state. Note that the latching behavior can be present in deterministically monostable feedback [40]. To test this, here, we directly quantified the feedback strength—to test if the feedback-induced bimodality results from latching feedback—by use of the small-signal loop gain, a direct measure of feedback strength [40,44,45]. The small-signal loop gain was quantified by measuring changes in LTR expression associated with changing Tat stability (S11 Fig) or increasing Tat concentration (S12 Fig and S32 Data). First, we verified that green fluorescent protein (GFP) fluorescence intensity was linearly correlated to GFP-protein abundance, as shown [32,46], by quantifying the fluorescence intensity of known concentrations of soluble GFP by microscopy and then comparing these values to the GFP fluorescence intensity of the LTR-GFP-IRES-Tat-FKBP circuit in 2 isoclonal populations when feedback was either inactive or active (S10 Fig and S33 and S34 Data). As expected, the GFP fluorescence intensity was well within the linear GFP-protein concentration regime for both microscopy and flow cytometry (S10 Fig and S33 and S34 Data). After verifying that fluorescence intensity scales linearly with protein abundance, we used fluorescence intensity to quantify changes in protein expression associated with altering Tat stability or Tat concentration. In agreement with other measures of HIV feedback strength [40], we find that Tat positive feedback appears to be nonlatching (S11 and S12 Figs and S32 Data). Interestingly, unlike systems that latch, nonlatching feedback strength inherently renders the system relatively insensitive to small fluctuations [47] (i.e., HIV will not drastically change expression profile or latch in response to a small fluctuation) lending a molecular explanation for the insensitivity of HIV circuitry to external cues [48,49].
The combination of nonlatching feedback coupled to a toggling promoter allows for bimodal generation across a wide range of Tat concentrations (Fig 2) and feedback strengths (Fig 3). Promoters driving nonlatching feedback can exhibit extended, transient pulses of expression before reverting back to the initial system state [8]. To test if this mechanism of extended-duration transient pulses was responsible for generating bimodality in the LTR, we built a specific model of the LTR to map out the phase space of feedback strengths that would allow for LTR bimodality given the known toggling parameters (S1 Table). The model specifically considers promoter toggling coupled to weak positive feedback and examined the effect of changing feedback strength (from weak nonlatching to strong nonlatching). In agreement with previous theoretical predictions [19,20], intrinsic slow promoter toggling is sufficient to generate bimodality, but only in a very narrow parameter regime (Figs 1 and 4A).
To explore if weak nonlatching positive feedback might explain the robust generation of bimodality that was experimentally observed, we incorporated dose-response data from the open-loop circuit into the model and generated an input-output function (S13 Fig) to quantify the relationship between Tat and kOFF values. This approach allows the open-loop data to be mapped onto a model containing feedback (Fig 4A). The output of the resulting model shows a striking dependence of bimodality on feedback strength (Fig 4B and S31 Data). Specifically, as feedback strength increases from zero, the bimodality regime significantly expands. However, as feedback increases further, to strong nonlatching feedback strengths, there is a drastic reduction in the potential for bimodal generation (Fig 4B, S14 Fig and S35 Data). This acute contraction of the bimodal regime likely results from drastic amplifications of small noise spikes that drive the system to stay on [17]. Interestingly, the model predicts that bimodality is generated across approximately 13% of the parameter values for the HIV system (Fig 4B and S35 Data), in agreement with experimentally observed frequencies for spontaneous bimodal generation across the HIV-integration landscape [32]. Thus, HIV’s moderate feedback strength (S11 and S12 Figs and S32 Data) appears optimized to slow promoter-toggling kinetics into the regime that enables bimodality.
Since the circuit’s bimodality is ultimately dependent upon fluctuation-driven (i.e., stochastic) extinction of Tat, we next sought to determine how increasing expression levels influenced bimodality. In the classical Poisson or super-Poissonian transcriptional burst models [50], the expression mean scales with variance (σ2 ∝ μ) such that the noise magnitude (CV2 = σ2 / μ2) decreases proportionally to the inverse of the mean squared (Fig 5A) and the extinction probability can be shown to be as follows (S1 Text):
Probextinct=∫−∞012πσ2e−(P−μ)22σ2dP
(1)
However, nonlatching positive feedback breaks the Poissonian relationship such that σ2 ∝ μN with 1 < N < 2 [44]. In the extreme case where N = 2, CV2 becomes independent of the mean and the extinction probability becomes the following (S1 Text):
Probextinct=∫−∞012πσ2NFBe−(x−μNFB)22σ2NFBdx,
(2)
where σ2NFB and μNFB are the variance and the mean for the nonfeedback case, respectively. Importantly, Eq 2 shows that stochastic extinction can be decoupled from the mean (when N = 2), and simulations verified that such perfect decoupling was possible (Fig 5B). Analysis of the experimental data in Fig 3 shows that the Tat circuit displays partial decoupling of noise and mean with N ≈ 1.5 (Fig 5C). Thus, Tat circuitry enables greater stochastic extinction over a broader range than other circuitries (e.g., no feedback or latching positive feedback) would be able to achieve.
In summary, HIV’s Tat circuit seems particularly well suited for generating bimodal expression patterns, and alternate single-parameter mechanisms for Tat function (e.g., increasing burst frequency alone rather than slowing toggling kinetics) appear to severely limit or completely abrogate the potential for bimodality. The precise architecture of this robust bimodal-generator circuit in such a rapidly adapting virus suggests that bimodality in HIV expression (i.e., latent and active replication modes) may be a beneficial trait that has been selectively maintained [38]. In contrast with other known roles for positive feedback (e.g., bistability and noise amplification), these findings demonstrate a further role for positive feedback as a mechanism for robust generation of bimodality [51]. On a conceptual level, this ability of positive feedback to expand the bimodal regime into physiological ranges may be related to positive feedback’s ability to expand the regime where sustained oscillations occur [52,53]. Consequently, positive-feedback circuits may have evolved not only for signal amplification but also to stabilize certain dynamic phenotypes (e.g., bimodality and oscillations) in diverse biological systems.
From a basic HIV biology standpoint, these results on Tat’s mechanism of action may have therapeutic implications for HIV cure approaches. Specifically, Tat protein addition reactivates HIV latency more potently than current chromatin remodeling latency-reversing agents (LRAs) such as histone deacetylase inhibitors (HDACis) [42]. Despite the known role of Tat as a transcriptional elongation factor, there has been no clear mechanistic explanation as to why Tat protein is more potent than LRAs that are transcriptional activators (e.g., HDACis). Conventional LRAs (e.g., protein kinase C [PKC] agonists and HDACis) only affect kON, and we have previously shown that agents that simultaneously reduce kOFF and kON potentiate reactivation [54]. Hence, the finding herein that Tat alters kOFF, coupled with the magnitude of the Tat-induced kOFF change, provides a mechanistic explanation as to why Tat is so effective for latency reversal. The findings also suggest that Tat-based strategies and conventional LRA strategies could be used synergistically, and new approaches aimed at simultaneously reducing kOFF and increasing kON would be optimal for “shock-and-kill” strategies, while conversely increasing kOFF and decreasing kON would be optimal for “block-and-lock” strategies.
The sequence of Tat from recombinant clone pNL4-3, GenBank: AAA44985.1, M19921, was used. To clone the LTR-mCherry-IRES-Tat-FKBP construct, d2GFP was swapped with mCherry using BamHI and EcoRI restriction sites [27]. To clone the Tet-Tat-Dendra-FKBP plasmids, Tat-Dendra or Tet-Tat-Dendra was swapped with YFP-Pif from the pHR-TREp-YFP-Pif plasmid (a gift from Wendell Lim’s laboratory at UCSF) using BamHI and NotI restriction sites. The full-length virus was generated as described [27].
For the GFP standard curve, a stock solution of 1 g/L (= 30.58 μM) recombinant eGFP (Cell Biolabs) was diluted 500-, 1,000-, 5,000-, and 10,000-fold (= 61.12, 30.58, 6.11, and 3.06 nM, respectively). These soluble GFP standards of known concentration were imaged in an 8-well chambered imaging dish using the same confocal microscope settings as subsequent cellular GFP imaging.
Isoclonal populations were incubated with shield for 20 hours (if applicable). Approximately 6 x 105 cells were washed with 2 mL of PBS solution and then immobilized on a Cell-Tak (Fisher) coated 8-well chambered imaging dish, using the manufacturer’s protocol. Both soluble GFP standards and cellular GFP were imaged on a Nikon Ti-E microscope equipped with a W1 Spinning Disk unit, an Andor iXon Ultra DU888 1k x 1k EMCCD camera, and a Plan Apo VC 100x/1.4 oil objective in the UCSF Nikon Imaging Center; the exposure time was 500 ms with 50% laser power. Approximately 15 xy locations were randomly selected for each isoclonal population. After background and autofluorescence subtraction from the cellular GFP images, the cellular GFP concentration was determined from the GFP standard curve. The cellular volume was approximated from the measured cellular dimensions, assuming a spherically shaped cell.
Lentivirus was generated in 293T cells and isolated as described [32,55]. To generate the isoclonal closed-loop circuit populations, lentivirus was added to Jurkat T Lymphocytes at a low MOI to ensure a single integrated copy of proviral DNA in the infected cells. The cells were stimulated with TNFα and Shield-1 for 18 hours before sorting for mCherry. Isoclonal and polyclonal populations were created as described [32]. The sorting and analysis of the cells infected was performed on a FACSAria II. Inducible-Tat cells were generated by transducing Jurkat cells with Tet-Tat-Dendra-FKBP and SFFV-rTta lentivirus at high MOI [27]. The cells were incubated in Dox for 24 hours and then FACS sorted for Dendra+ cells to create a polyclonal population. To create the Tet-Tat-Dendra-FKBP + LTR-mCherry cells, the polyclonal population was infected with LTR-mCherry lentivirus at a low MOI. Before sorting for mCherry+ and Dendra+ cells, Dox was added at 500 ng/mL for 24 hours, and single cells were FACS sorted and expanded to isolate isoclonal populations.
Flow cytometry data were collected on a BD FACSCalibur DxP8, BD LSR II, or HTFC Intellicyt for stably transduced lines and sorting. Flow cytometry data were analyzed in FlowJo (Treestar, Ashland, Oregon, United States) and using customized MATLAB code [27].
A simplified 2-state model of LTR toggling and Tat positive feedback was constructed based on experimental data of LTR toggling [24,36] and simulated using the Gillespie algorithm [56] in MATLAB to test how altering toggling kinetics and feedback strength would affect the activity of the circuit. At least 1,000 simulations were run for each condition.
Alternatively, to sweep the parameter space of different modulations of the Tat circuit, the accurate chemical master equation (ACME) method [57,58] was used to directly solve the chemical master equation (CME) to obtain the full probability landscapes of protein copy number. For each parameter pair in the sweeping, the protein probability landscape was computed at day 3 or at steady state. The phenotype of bimodality or unimodality at different parameter pairs was based on the numbers and locations of probability peaks in the landscape using the bimodality analysis approach described in the Materials and methods section.
Two approaches were taken to quantify whether a distribution from the experimental data or simulations was bimodal or unimodal. The first, applied to both simulations and experimental data, was to convert the fluorescence density data using the bkde function in the KernSmooth package in R to a binned kernel density [59]: the KernSmooth R package is available at https://cran.r-project.org/web/packages/KernSmooth/index.html. To filter out biologically irrelevant noise in the data, the data points with fluorescence density less than 1 or small peaks lower than 0.05 in calculated kernel density function were ignored. The number of modality peaks was determined by calculating the second-order derivative of the kernel density. The second approach, only applied to the experimental data, was to utilize the Hartigan Dip Test, a dip statistic that can test for multimodality by testing for maximal differences and ascertain the probability that a particular distribution is unimodal [39]. Code for the Hartigan Dip Test was obtained from http://nicprice.net/diptest/, adapted from Hartigan’s original Fortran Code for MATLAB.
|
10.1371/journal.pgen.1000973 | Sgs1 and Exo1 Redundantly Inhibit Break-Induced Replication and De Novo Telomere Addition at Broken Chromosome Ends | In budding yeast, an HO endonuclease-inducible double-strand break (DSB) is efficiently repaired by several homologous recombination (HR) pathways. In contrast to gene conversion (GC), where both ends of the DSB can recombine with the same template, break-induced replication (BIR) occurs when only the centromere-proximal end of the DSB can locate homologous sequences. Whereas GC results in a small patch of new DNA synthesis, BIR leads to a nonreciprocal translocation. The requirements for completing BIR are significantly different from those of GC, but both processes require 5′ to 3′ resection of DSB ends to create single-stranded DNA that leads to formation of a Rad51 filament required to initiate HR. Resection proceeds by two pathways dependent on Exo1 or the BLM homolog, Sgs1. We report that Exo1 and Sgs1 each inhibit BIR but have little effect on GC, while overexpression of either protein severely inhibits BIR. In contrast, overexpression of Rad51 markedly increases the efficiency of BIR, again with little effect on GC. In sgs1Δ exo1Δ strains, where there is little 5′ to 3′ resection, the level of BIR is not different from either single mutant; surprisingly, there is a two-fold increase in cell viability after HO induction whereby 40% of all cells survive by formation of a new telomere within a few kb of the site of DNA cleavage. De novo telomere addition is rare in wild-type, sgs1Δ, or exo1Δ cells. In sgs1Δ exo1Δ, repair by GC is severely inhibited, but cell viaiblity remains high because of new telomere formation. These data suggest that the extensive 5′ to 3′ resection that occurs before the initiation of new DNA synthesis in BIR may prevent efficient maintenance of a Rad51 filament near the DSB end. The severe constraint on 5′ to 3′ resection, which also abrogates activation of the Mec1-dependent DNA damage checkpoint, permits an unprecedented level of new telomere addition.
| A chromosomal double-strand break (DSB) poses a severe threat to genome integrity, and budding yeast cells use several homologous recombination mechanisms to repair the break. In gene conversion (GC), both ends of the DSB share homology to an intact donor locus, and the break is repaired by copying the donor to create a small patch of new DNA synthesis. In break-induced replication (BIR), only one side of the DSB shares homology to a donor, and repair involves assembly of a recombination-dependent replication fork that copies sequences to the end of the template chromosome, yielding a nonreciprocal translocation. Both processes require that the DSB ends be resected by 5′ to 3′ exonucleases, involving several proteins or protein complexes, including Exo1 and Sgs1-Rmi1-Top3-Dna2. We report that ectopic BIR is inhibited independently by Sgs1 and Exo1 and that overexpression of Rad51 recombinase further improves BIR, while GC is largely unaffected. Surprisingly, when both Sgs1 and Exo1 are deleted, and resection is severely impaired, half of the cells acquire new telomeres rather than completing BIR or GC. New telomere addition appears to result from the lack of resection itself and from the fact that, without resection, the Mec1 (ATR) DNA damage checkpoint fails to inactivate the Pif1 helicase that discourages new telomere formation.
| DNA double-strand breaks (DSBs) are generated by normal cellular processes including DNA replication or by exposure to DNA damaging agents or ionizing radiation. To maintain cell viability and preserve genomic integrity, cells employ multiple pathways of homologous recombination (HR) to repair DSBs [1]–[4]. A key initial step in HR is 5′ to 3′ resection of DSB ends to create single-stranded DNA (ssDNA) that recruits formation of a Rad51 filament, which engages in a search for homologous sequences. The predominant HR pathway is gene conversion (GC), a conservative mechanism in which both ends of the DSB share homologous sequences on a sister chromatid, a homologous chromosome, or at an ectopic location. Rad51-mediated strand invasion of the 3′-ended ssDNA allows the initiation of new DNA synthesis to copy a short region of the template and patch up the DSB. When only one DSB end shares homology to a template elsewhere in the genome, a less-efficient HR mechanism, break-induced replication (BIR), can be used to repair the break [5], [6]. In BIR, recombination is used to establish an uni-directional replication fork that can copy the template DNA to the end of the chromosome. If homologous sequences are located ectopically, BIR will result in formation of a non-reciprocal translocation with loss of the distal part of the broken chromosome and may be a significant source of gross chromosomal rearrangements (GCRs) and genomic instability [7]. BIR requires the non-essential subunit of the Polδ polymerase, Pol32, and all of the essential replication machinery except those excluisvely required for formation of the pre-replicative complex [8], [9]. BIR can be used to restart stalled or collapsed replication forks during DNA replication [10] and elongate telomeres in the absence of telomerase [8]. An alternative way to repair the DSB is through de novo telomere addition through the action of telomerase [11]–[13], although this is a very inefficient process that is improved by elimination of the Pif1 helicase [14].
Genetic and in vivo molecular biological experiments indicate that the early steps of GC and BIR are shared [15]–[17]. Following the generation of a DSB, the Tel1/ATM kinase is loaded at sites of DSBs in an Mre11-Rad50-Xrs2 (MRX)-dependent manner [18], [19]. Tel1 in turn phosphorylates MRX [20], [21]. The Sae2 and MRX proteins mediate the initial resection [22], [23] which is continued via two alternate pathways, one using the Exo1 nuclease and the other employing the multifunctional RecQ family helicase Sgs1, in concert with Top3, Rmi1 and the essential helicase/nuclease Dna2 [22]–[24]. DNA resection is also essential to activate the Mec1-dependent DNA damage checkpoint kinase cascade that triggers a cell cycle arrest, allowing time for the cell to repair the beak prior to mitosis [25].
Following resection, Rad51-mediated strand invasion of the donor template occurs with similar kinetics, but the initiation of DNA synthesis at the 3′-end of the invading strand is greatly delayed in BIR as compared to GC [16], [17]. Recently, Jain et al [16] showed that a “Recombination Execution Checkpoint” (REC) delays the initiation of BIR synthesis if a second DSB end has not become engaged nearby on the same template. It is unclear if the delay in BIR synthesis is due to a restructuring of the strand invasion D-loop and/or the recruitment of BIR associated proteins. The efficiency of BIR is inhibited by Sgs1, as there is an increase in BIR in sgs1Δ cells [16]. Sgs1 also has been shown to disrupt HR intermediates [26], inhibit homeologous recombination [27]–[29], and to dissolve double Holiday Junctions (dHJ) to yield noncrossovers [30]–[32].
To better understand the role of Sgs1 in BIR, we examined mutations of non-essential genes that either cooperate or act redundantly with Sgs1 in many of its roles in DNA metabolism, including DNA resection. Here we show that deletion of SGS1 or EXO1 increases the efficiency of BIR whereas overabundance of Sgs1 or Exo1 strongly inhibits it. Overexpression of Exo1 also inhibits GC. Deletion of other non-essential factors responsible for DNA resection, TEL1 or SAE2, modestly increases the efficiency of BIR whereas deletion of MRX impairs BIR. Additionally, we find that overexpression of Rad51 markedly improves the efficiency of BIR but has little effect on GC. Finally, we show that Sgs1 and Exo1 redundantly prevent remarkably efficient de novo telomere addition at broken chromosome ends, a pathway dependent on both telomerase and Sae2.
To study BIR we used the haploid Saccharomyces cerevisiae strain JRL346. A galactose-inducible HO endonuclease is expressed to induce a DSB at a modified CAN1 locus approximately 30 kb from the telomere in the non-essential terminal region on Chromosome V (Ch V) (Figure 1A). The HO endonuclease cut site and an adjacent hygromycin-resistant marker, HPH-MX, was integrated into the CAN1 locus, deleting the 3′ portion of the gene but retaining the 5′ portion of the gene (denoted as CA). A 3′ portion of the gene (denoted as AN1) with 1,157 base pairs of shared homology to CA on Ch V was introduced in the same orientation into Ch XI, 30 kb from its telomere. Prior to HO induction, these cells are canavanine-resistant (CanR) because CAN1 is disrupted. Completion of BIR results in a non-reciprocal translocation that duplicates the donor sequences and the more distal part of the left arm of Ch XI, thus restoring an intact CAN1 gene. These cells become canavanine-sensitive (CanS) and hygromycin sensitive (HphS). About 20% of cells are viable with 99.85% of these cells repairing by BIR and a small fraction by nonhomologous end-joining (NHEJ). The efficiency of BIR repair allows us to physically monitor the kinetics of repair by PCR, Southern blot and pulse-field gel electrophoresis (PFGE), as described in Materials and Methods.
To compare the effects of mutations on GC, we used the isogenic strain JRL475 (Figure 1B). The GC strain was modified from the BIR strain by introducing 2,404 bp of homology marked by URA3 to the other end of the break (denoted as 1, for the 3′-end of CAN1). The insertion of the URA3-1 sequences also deleted 376 bp in the middle of the CAN1 so there is a gap between the homology shared by the two DSB ends created by HO cleavage (CA-URA3-1) with the donor sequences on Ch XI (AN1). Repair by GC results in restoration of the CAN1 gene, rendering cells CanS, but, unlike BIR, the Ch V arm distal to the cut site is retained. When there is a second end of homology to a DSB break, the cell strongly favors GC over BIR [16], [17], [33], so that after induction of a DSB cell viability increases from 20% in the BIR strain to nearly 70% when there are two ends of homology and GC is used to repair the break (Figure 1B and Figure 2B).
To better understand the role of Sgs1 in BIR, we first measured the viability of sgs1Δ cells after inducing a DSB (Figure 1A). As previously shown [16], sgs1Δ cells are 1.5 times more efficient in BIR compared to wild type cells (Figure 2A), repairing the break with 33% efficiency (p<0.001). To confirm that the increase in viability directly correlates with an increase in repair product, we monitored the kinetics of repair using the PCR assay that detects the first 242 bp of new DNA synthesis. The maximum amount of product detected by PCR (18% at 12 hours) in wild type cells (Figure 2D) is comparative to the viability of cells (21%) following induction of the DSB (Figure 2A). As expected, deletion of SGS1 increased the efficiency of product formation compared to wild type cells (Figure 2D). Using the previously described BIR system involving the LEU2 sequences [16] we also showed that a helicase-dead allele of Sgs1 [34] behaves like the complete deletion of Sgs1 (Figure S1). We have previously shown that deletion of sgs1Δ does not increase the efficiency of GC events in which there is perfect homology or when there is a small gap in homology of 1.2 kb or less [16], [35]. We confirmed that sgs1Δ does not affect the efficiency of GC in the ectopic assay used here (Figure 1B and Figure 2B).
To better understand the role of Sgs1 in BIR, we investigated a number genes that have previously been shown to interact genetically with Sgs1 [27], [29], [36]–[41]. Deletions of MSH6, MUS81, YEN1, RAD27, ESC2, DIA2, YBR094w, or RNH202 did not have a statistically significant effect on BIR when tested for viability after inducing a DSB that can only be repaired by BIR (Table S1). However, we found that the other non-essential genes required for 5′ to 3′ resection of DSB ends all affect the efficiency BIR. A deletion of SAE2 resulted in a slight, but statistically significant, increase in viability (p = 0.02). In contrast, deleting subunits of the MRX complex, mre11Δ or rad50 Δ, decreased viability nearly 2 fold (both p = 0.003) (Figure 2A). The effect of deleting mre11Δ or rad50Δ is consistent with results previously seen in a diploid BIR assay in which a DSB is induced at the MAT locus on Ch III [17], [42], but differs from a transformation-based BIR assay that saw no requirement for MRX in BIR [15].
Because Tel1 plays a role in suppressing gross chromosomal rearrangements and enhances Sae2 and MRX activity in DNA resection [43] we asked if deletion of TEL1 would affect BIR. Similar to sae2Δ, deletion of TEL1 resulted in a small but statistically significant increase in viability (p = 0.008) (Figure 2A). Complementation of a tel1Δ strain with the kinase-dead allele [20] partially restored viability to wild type levels (Figure 2A).
The Exo1 nuclease acts redundantly with Sgs1 in DNA resection after the initial trimming of the ends by Sae2 and MRX, although by itself exo1Δ has a minimal impact on 5′ to 3′ resection [22]–[24]. Similar to sgs1Δ, deletion of EXO1 (p = 0.001) increased viability nearly 1.5 times compared to wild type (Figure 2A). Also like sgs1Δ, deletion of EXO1 increased the efficency of BIR when measured by PCR (Figure 2D) and does not affect the efficiency of GC (Figure 2B).
Plamids overexpressing Sgs1 pYES2-SGS1 [44] or Exo1 (pSL44) [45] were expressed under the control of a galactose-inducible promoter on a high copy plasmid. These overexpression plasmids are denoted as pGAL::SGS1 and pGAL::EXO1, respectively. Expression is induced concomitantly with HO induction. In cells carrying pGAL::SGS1, the efficiency of BIR decreased 5 fold (p<0.001) whereas in pGAL::EXO1 the efficiency of BIR decreased 10 fold (p<0.001) (Figure 2A). Overexpression of these genes did not affect cell viability in cells that lacked an HO cleavage site (data not shown). Furthermore, we found that Exo1 overexpression inhibited BIR prior to inhibition of new DNA synthesis, by monitoring the kinetics of repair by PCR (Figure S2). The strong inhibition of BIR by overexpressing Exo1 depends on the nuclease activity of this protein, as there is no such inhibition when we overexpressed plasmids carrying exo1 mutations that are required for exonuclease activity (Figure 2C). As shown previously [8], increasing the homology in our BIR assay more than two fold to 2,977 bp increases the efficiency of BIR (Figure 3C). The increase in homology results in slightly higher viability but does not significantly suppress the effects of overexpressing SGS1 or EXO1 (Figure 3C). When tested in the GC assay, overexpressing Sgs1 had no effect on viability but overproduction of Exo1 decreased viability by half (Figure 2B).
The initiation of BIR is delayed several hours after the ends of the DSB begin to be resected at a wild type rate of about 4 kb/hr [22], [46]. We have also previously shown that the abundance of Rad51 is sufficient to continuously coat only about 10 kb of ssDNA on either side of the break [47]; consequently it is possible that excess ssDNA would interfere with forming or maintaining a stable and efficient Rad51 filament that is needed to promote strand invasion and initiation of new DNA synthesis. Excess ssDNA has been previously shown to interfere with recombination in meiotic cells [48]. We therefore asked if overexpression of Rad51 would also increase the efficiency of BIR, using well-characterized high-copy plasmids in which RAD51 was expressed under the ADH1 promoter (pDBL(RAD51)) [49] or under the PGK promoter (pSJ5). Strikingly, overexpressing RAD51 in wild type cells caused a 2.5-fold increase in viability (p<0.001) when expressed under control of either promoter (Figure 3D). When we tested the same plasmids in the GC assay we found that there was a slight but not statistically significant decrease in viability (Figure 3E). These results clearly indicate that Rad51 overexpression preferentially stimulates BIR. Overexpression of RAD51 in the BIR assay with longer homology further increased the efficiency of BIR (Figure 3C). We also find that the efficiency of BIR is increased when we tested the kinetics of repair by Southern blot (Figure 3A) and PCR (Figure 3B). However, when normalized to the percent of final product the kinetics of repair are not different from wild type cells (data not shown).
An elevated level of Rad51 increased the viability of sgs1Δ, exo1Δ or tel1Δ cells to the level seen for overexpressed RAD51 alone (Figure 3D), so the effects of RAD51 expression and deleting SGS1 or EXO1 are not additive. However, overexpressing RAD51 in cells also overexpressing SGS1 or EXO1 did not significantly suppress the inhibition of BIR that is seen with overexpressing SGS1 or EXO1 alone (Figure 3D). These results could suggest that Sgs1 and Exo1 act prior to the rate-limiting step carried out by Rad51. In the case of Sgs1, it could be in dismantling transient strand invasion encounters; for Exo1, there is no evident mechanism at this point unless a modest increase in resection [45] would overwhelm excess Rad51.
We examined a a dramatic 2-fold increase in viability in an sgs1D exo1D double mutant compared to sgs1Δ or exo1Δ alone when tested in the BIR assay (Figure 4A); however this increase is not in the level of BIR. Instead, it is due to a dramatic increase in new telomere addition, as described below. There is in fact no increase in BIR events compared to the single mutants and repair appears to be no better than wild type cells when repair was monitored by PCR (Figure S3). As has previoulsy been reported [22]–[24], we found that resection is severely impaired in sgs1Δ exo1Δ cells as evident by the persistence of the cut chromosome band seen by Southern blot (data not shown). Although TEL1 and SAE2 moderately inhibit BIR and are involved in DNA resection like SGS1 and EXO1 [50], deleting TEL1 did not cause new telomere additions at the DSB when ablated in combination with sgs1Δ or exo1Δ nor did deletion of SAE2 in combination with exo1Δ (Figure 4A).
As mentioned above, when we analyzed the viablity of sgs1Δ exo1Δ cells, we found that half of the survivors did not have the CanS HphS phenotype indicative of repair by BIR (Figure 4A). Instead, the new survivors were HphS but CanR, suggesting that they might have lost the terminal non-essential portion of Ch V distal to the cut site but failed to restore a functional CAN1 locus.
Sgs1 has previously been shown to inhibit homeologous recombination [27], [29], specifically the formation of translocations between CAN1 and two highly diverged CAN1 homologs, LYP1 and ALP1, on Ch XIV [51]; these rearrangements might be further elevated by the absence of Exo1. Alternatively, given that sgs1Δ exo1Δ severely retards 5′ to 3′ resection, the chromosome end could be stabilized, allowing new telomere addition. To distinguish between these possibilities, we performed pulse field gel electrophoresis (PFGE) on 12 independent CanR HphS colonies, comparing them to the starting strain and a survivor that repaired by BIR (CanS HphS) (Figure 5). The ethidium bromide-stained agarose gel (Figure 5A) shows that the majority of the CanR HphS survivors (lanes 1–11) have a smaller chromosome than the starting (ST) strain or one repaired by BIR (B). (There is no size difference in Ch V size prior to DSB induction and after BIR because the 30 kb of non-essential region distal to the cut site on Ch V is replaced by a duplication of 30 kb from Ch XI.) We confirmed by Southern blot that the band remaining at the original position of Ch V is Ch VIII, which is approximately the same size as Ch V in this strain background (data not shown). One CanR HphS colony (lane 12) increased in size from the original strain. These data indicate the CanR colonies are not due to mutations in a restored CAN1 gene, and are therefore not repaired by BIR nor by NHEJ that could have deleted a small region including HPH. To confirm that none of the CanR HphS colonies were repaired by BIR, we probed with the MCH2 probe that hybridizes proximal to the telomere on Ch XI (Figure 5B). The MCH2 probe hybridized to sequences on Ch XI in every sample, but only to Ch V in the CanS HphS colony that repaired by BIR.
To determine what sequences of Ch V were retained in the CanR HphS colonies, we next probed the blot with a CAN1 probe that hybridizes to the donor sequences on Ch XI and just proximal (1 kb) to the cut site on Ch V (Figure 1A, Figure 5C). The CAN1 probe hybridized to sequences on Ch XI in all samples and to Ch V in the starting and BIR strains, but only to three CanR HphS colonies (1, 9 and 12). This result indicates that at least 1 kb of sequence was deleted in the 9 other CanR HphS survivors. To determine approximately how much sequence was deleted in the other CanR HphS colonies we probed the Southern blot with a NPR2 probe that specifically hybridizes to Ch V 4 kb proximal to the cut site (Figure 1A and Figure 5D). In this case, the NPR2 probe hybridized to all CanS samples except lanes 3, 5, 6, and 7. When we probed with PRB1 that hybridizes approximately 9 kb proximal to the cut site on Ch V, the probe hybridized to Ch V in all CanS survivors (Figure 1A and Figure 5E). We also probed the blot with the highly diverged ALP1 and LYP1 sequences on Ch XIV with which CAN1 forms translocations in sgs1Δ cells [51], but these sequences did not hybridize to the novel chromosome in lane 12 (data not shown). We have not explored further the structure of this translocation.
Based on our PFGE and Southern blot analysis we conclude that the great majority of the CanR HphS survivors result in a truncation of Ch V after limited resection. To show if the sequences at the terminus of the truncations are indeed new telomeres, we determined the breakpoint of five independent sgs1Δ exo1Δ CanR HphS repaired colonies by PCR, using a Ch V-specific primer and a telomere-specific primer as previously described [52], [53]. As shown in Figure 6, the presence of a new telomere is indicated by a laddered PCR product. We then sequenced the PCR product using the Ch V-specific primer. As shown in Table 1, all five sgs1Δ exo1Δ CanR HphS colonies have new telomere sequences directly added to the Ch V sequences. Consistent with the PFGE and Southern blot analysis, the breakpoints were not at a uniform location. Based on our results, we hypothesize that in the absence of both Sgs1 and Exo1, a DSB frequently results in a truncated chromosome with newly added telomeres and that these additions can occur at several different sites, often as far as between 1 and 4 kb away from the DSB end. To confirm that these events are telomerase-dependent, we deleted EST2, an essential components of telomerase. As shown in Figure 4A, deletion of EST2 does not affect repair by BIR but eliminates recovery of CanR colonies.
We next asked if NHEJ or HR pathways contributed to de novo telomere formation (Figure 4A). Telomere addition was not dependent on NEJ1, which is required for NHEJ. We next deleted RAD51, which is required for both BIR and GC. We confirmed that nearly all BIR is eliminated in sgs1Δ exo1Δ rad51Δ cells but also found a 20% increase in the number of cells with new telomeres. Although overexpression of RAD51 increased the efficiency of BIR it did not suppress new telomere addition (Figure 4A). We then tested if the MRX-associated exonuclease Sae2 plays a role in new telomere addition. Recently, Sae2 and Sgs1 have also been shown to act in parallel telomere processing pathways [54]. Interestingly, when resection is nearly eliminated by deletion of sae2Δ in combination with sgs1Δ exo1Δ, new telomere addition is eliminated and BIR is significantly reduced (Figure 4A). When TEL1 was deleted in combination with sgs1Δ exo1Δ there was no change in levels of BIR or de novo telomeres compared to sgs1Δ exo1Δ cells.
It has previously been seen that sgs1Δ exo1Δ cells are defective in GC when tested for the ability to successfully complete MAT switching [23]. When we tested the viability of sgs1Δ exo1Δ cells in our GC assay there was no discenrable effect on viability. However, when the phenotypes of the viable colonies were examined only 5% were CanS, which is indicative of repair by GC, while the remaining viabile colonies were CanR, consistent with a truncated chromosome (Figure 4B). The drastic decrease in GC is consistent with previously published defects seen in sgs1Δ exo1Δ cells. We analyzed 10 independent CanS colonies by PCR to ascertain if the break was repaired by GC (Figure 4C). In fact, only 5 of the 10 colonies analyzed (samples S2, S3, S4, S5, S8) repaired by GC whereas 4 of the colonies repaired the break by BIR (S1, S6, S7, S10). One colony (S9) had PCR products consistent with repair by both GC and PCR. The use of BIR to repair half of the sgs1Δ exo1Δ colonies is consistent with the failure of these cells to activate the DNA damage checkpoint and thus to enter mitosis in the absence of DSB repair.
To verify that that the DNA damage checkpoint was impaired by the lack of normal 5′ to 3′ resection of the DSB ends we microscopically monitored the length of the cell cycle of individual cells plated on YEP-Gal to induce HO endonuclease, from the time that an unbudded G1 cell formed a bud until the dumbbell-shaped mother-daughter pair formed the next bud [55]. Wild type cells in which the DSB cannot be repaired remain arrested prior to anaphase for approximately 6 cell division times relative to an isogenic strain lacking the HO cleavage site [55]. In contrast, cells of the BIR strain lacking SGS1, EXO1 and RAD51, so that they could not repair the DSB by homologous recombination, show a brief, but significant arrest. These cells extend the cell cycle 1.8 times the length of time of a derivative that lacks the cut site (6.2 h versus 3.5 h). Thus, there is still a brief activation of DSB-induced cell cycle arrest but much shorter than when extensive resection activates Mec1.
As was the case with CanR sgs1Δ exo1Δ colonies found in the BIR assay, the CanR colonies in the GC assay appear to be chromosome truncations with de novo telomere formation. PCR analysis showed that the broken chromosomes were truncated at different points proximal of the DSB (Figure S4). When representative isolates were tested by PCR as mentioned above we found that consistent with new telomere addition there was a laddered PCR product as seen in sgs1Δ exo1Δ cells in the BIR assay (Figure S4).
We conclude that eliminating both Sgs1 and Exo1, by markedly reducing 5′ to 3′ resection and most likely by preventing full activation of the Mec1-dependent DNA damage checkpoint (see Discussion), allows a dramatic increase in new telomere formation, rescuing almost half of all cells suffering a DSB.
In this work we show that the RecQ family helicase, Sgs1, and the Exo1 exonuclease negatively regulate BIR to maintain genomic integrity. From the observation that the efficiency of BIR was no greater in sgs1Δ exo1Δ than in a single mutant one might conclude that the helicase/endonuclease (Sgs1-Rmi1-Top3/Dna2) and Exo1 act in the same pathway, but since the sgs1Δ exo1Δ double mutant has such distinctly different phenotypes from sgs1Δ or exo1Δ it is difficult to know precisely why the double mutant does not show an increase in BIR similar to that seen when Rad51 is overexpressed in sgs1Δ or exo1Δ alone. We note also that other proteins responsible for 5′ to 3′ DNA resection, Sae2 and MRX, do not inhibit BIR in the same fashion; but the behavior of sae2Δ or mre11Δ may be explained by their other important roles in other steps in HR [1], [3],[4].
Sgs1 and Exo1 likely do not act in precisely the same way in inhibiting BIR. Sgs1-mediated inhibition of BIR may involve unwinding of a nascent strand invasion D-loop, as demonstrated in vitro for the human Sgs1 homolog, BLM [56], [57]. In vivo it is clear that the Sgs1 helicase can dismantle strand annealings and strand invasions if the heteroduplex DNA contains mismatches [27]–[29]. In meiotic recombination, Sgs1 prevents independent strand invasions of alternative templates [58], [59]. If Sgs1 dismantles heteroduplex DNA, we might expect that increased homology between the DSB end and the donor template would lead to a more stable D-loop that would counteract Sgs1. Increasing the extent of homology from 1.1 kb to ∼3 kb did not significantly change the response of cells to overexpression of Sgs1. It is also possible that Sgs1 inhibits the recruitment of some of the BIR-associated proteins. We note that the effect of deleting Sgs1 or Exo1 is not apparent in a different BIR assay system in a diploid in which nearly all homologous sequences distal to the DSB are deleted [17], [60]; and where there are 100 kb of homologous sequences centromere-proximal to the DSB that can be used to initiate BIR. However, even in this case, many BIR events fail to retain a marker 3 kb proximal to the DSB, suggesting either that more extensive homology increases BIR or that some more proximal sequences are especially favored in initiating BIR [61].
Rather than acting on D-loop stability, Exo1 may act on the assembly of the BIR replication fork. In response to DNA damage or defective checkpoint activation, Exo1 has also been shown to process stalled replication forks and resect nascent strands [62], [63]. The mechanism by which Exo1 interferes with fork integrity is unclear; it may be possible that the intermediate steps at which the BIR replication fork is assembled are an Exo1 substrate. We have previously shown that overexpression of Exo1 increases the rate of resection [45]; this has not been tested for Sgs1 overexpression.
A unifying hypothesis would be that BIR is severely limited if resection of the DSB ends is too extensive. There is a limited amount of Rad51 in the cell (about 3,500 molecules), enough to cover continuously about 10 kb of ssDNA [47]. Although Rad51 will initially form a filament with sequences close to the DSB (including the relevant “CA” sequences that engage in BIR), as resection proceeds the continuous polymerization and depolymerization of Rad51 may leave patches of Rad51 along much of the ssDNA so that by the time BIR is seen, many DSBs will not have a continuous Rad51 filament near the 3′ end to promote the completion of recombination. Thus, even in wild type cells, overexpressing Rad51 would ensure that there would be a functional filament over the CA sequences and BIR would consequently be more efficient. Deletions of Sgs1 or Exo1 would partially suppress the problem by slowing down resection (hence BIR is increased 1.5 times wild type), although we again note that exo1Δ by itself has little visible effect on resection. Overexpression of Rad51 is apparently unable to suppress the consequences of overexpressing Exo1 or Sgs1. It is important to note that Exo1 overexpression is only effective if nuclease activity is preserved; at least some of Exo1's functions in meiosis are independent of nuclease activity (N. Hunter, personal communication; L. Symington, personal communication). Increasing homology in our assay does not suppress these effects but further increases in homology may do so, as noted above.
It is possible that overexpressing Rad51 could ensure that the 3′-ended single-stranded DNA was better protected against degradation over the long time required to enact BIR, as previously suggested [64]. However, we have previously shown that in single-strand annealing where one of the flanking 1-kb homologies is very close to the DSB and the other is exposed only after 6 hr of 5′ to 3′ resection, at least 85% of cells are able to accomplish SSA, which would be impossible if even 1 kb of the 3′-end were degraded in the 6-hr period. Moreover, SSA was equally possible with and without Rad51 [16], arguing that Rad51 did not provide end-protection to the 3′-ended single-strand.
Eliminating both Sgs1 and Exo1 had a marked defect in completing GC but did not impair BIR so severely. Because resection is severely impaired in the sgs1Δ exo1Δ double mutant, it is possible that the more severe defect in GC is attributable to the need to resect more than 1 kb of intervening sequence before the “1” end of homology would be single-stranded (see Figure 1B). However, it is also possible that the difference reflects still another defect in sgs1Δ exo1Δ strains, a failure to activate the DNA damage checkpoint because of a lack of sufficient ssDNA [25], [65]. If mitosis is not arrested, then cells that have an unrepaired DSB will proceed through mitosis. This may lead to the loss of the acentric fragment, as we have shown in other assays [66], so that only the centromere-proximal DSB end will be inherited. This situation is not fatal for BIR, which only uses homology on that side of the DSB; indeed previous studies [17], [67] have shown that BIR may actually increase in a checkpoint-deficient situation whereas GC will be defective. Thus, even when GC should be possible, half of the HR outcomes of the sgs1Δ exo1Δ GC assay proved to be BIR events.
Strikingly, Sgs1and Exo1 also redundantly inhibit new telomere formation. In a previous study [12], when an HO-induced DSB was generated in a rad52Δ strain that could not carry out recombination but had apparently normal 5′ to 3′ resection, only about 1% of cells created new telomeres, and this was only in a situation where a “seed” of T2G4 telomere sequences was located centromere-proximal to the DSB. In the absence of the T2G4 repeats, new telomeres arose less than 0.1% of the time. The remarkably high level of new telomere formation (up to 50% of all cells) must be attributable to the elimination of vigorous resection in the double mutant strain, but it is also likely that the failure to activate the Mec1 DNA damage checkpoint also plays a key role. Recently, Makovets and Blackburn [68] have shown that the Pif1 helicase, which antagonizes new telomere formation [69], is phosphorylated in a Mec1-dependent fashion; hence if sgs1Δ exo1Δ block resection and that prevents Mec1 activation, new telomeres should increase. However, in the assay used by Makovets and Blackburn [68] the level of new telomeres added near an HO endonuclease-induced DSB was only about 2%. Moreover, Chung et al [60] also find that new telomere addition is much less efficient in cells lacking MEC1 compared to sgs1Δ exo1Δ cells. Hence, it is likely that the 40–50% level of de novo telomere formation we find reflects both the failure to activate Pif1 when the checkpoint is not strongly activated and the severe block on resection itself.
Apparently de novo telomere formation does not require the recruitment of the MRX-Tel1 complex, as a tel1Δ mutant does not affect the formation of new telomeres in an sgs1Δ exo1Δ strain. When resection is blocked by deletion of SAE2 in sgs1Δ exo1Δ cells, new telomeres are absent. The fact that new telomeres were added as far as 4 kb from the DSB site indicates that there is a residual resection activity that–over a period of perhaps many hours–can chew away the chromosome end and expose sites suitable for new telomere addition. However, we show that the MRX-asociated endonuclease SAE2 is required for de novo telomere formation.
In this work we have expanded our understanding of the genetic relationships of factors that negatively regulate BIR. Furthermore, we have provided evidence for a novel repair pathway that is redundantly impaired by Sgs1 and Exo1. Understanding the interplay of these factors in response to DNA damage and uncovering the molecular details of signaling between them to maintain genomic integrity will be an area of much future research.
The wild type JRL346 was derived from JRL092 [8] by first disrupting the LEU2 marker with a leu2::hisG construct from pNKY85 [70] to generate strain JRL187. The HMRa-stk gene was then knocked out with an hmr::ADE3 fragment generated by PCR with mixed oligos to generate JRL346. All strains used to study BIR are isogenic to JRL346 and were created by standard gene disruption methods and confirmed by PCR unless otherwise stated [71]. In order to generate an assay to study GC that is isogenic with JRL346, an HOcs-HPH cassette [8] was integrated into Ch V between nucleotides 31,644 and 32,020, resulting in a truncation of the CAN1 ORF at nucleotide 1,146 to create strain JRL017 (CL11-7 can1,1-1446::HOcs::HPH). JRL017 was then modified by transforming in a hphmx::URA3 “marker swap” cassette [72] to generate JRL472 (CL11-7 can1,1-1446::HOcs::URA3::AVT2). To introduce another 2,404 bp of homology to the donor, the can1,1-1446::HOcs::URA3::AVT2 region with Ch V sequences 29,146 to 32,976 was amplified from JRL472 and integrated distal to the HO cut site into Ch V in strain JRL346 to generate JRL475 (can1,1-1446::HOcs::URA3::AVT2 ykl215c::leu2::hisG::can1DEL1-289::AVT2). As a result, there are Ch V sequences 33,177–32,020 shared between the donor and sequences proximal to the break, Ch V sequences 31,644–29,240 shared between the donor and sequences distal to the break and a 376 bp gap of homology. All mutant strains were created by standard gene disruption methods and confirmed by PCR. Plasmid pSJ5 was constructed by subcloning a XhoI-NotI fragment containing the RAD51 ORF under the PGK promoter form pNSU256 [47] into pRS314 [73].
The wild type JRL346 was derived from JRL092 [8] by first disrupting the LEU2 marker with a leu2::hisG construct from pNKY85 [70] to generate strain JRL187. The HMRa-stk gene was then knocked out with an hmr::ADE3 fragment generated by PCR with mixed oligos to generate JRL346. All strains used to study BIR are isogenic to JRL346 and were created by standard gene disruption methods and confirmed by PCR unless otherwise stated [71]. In order to generate an assay to study GC that is isogenic with JRL346, an HOcs-HPH cassette [8] was integrated into Ch V between nucleotides 31,644 and 32,020, resulting in a truncation of the CAN1 ORF at nucleotide 1,146 to create strain JRL017 (CL11-7 can1,1-1446::HOcs::HPH). JRL017 was then modified by transforming in a hphmx::URA3 “marker swap” cassette [72] to generate JRL472 (CL11-7 can1,1-1446::HOcs::URA3::AVT2). To introduce another 2,404 bp of homology to the donor, the can1,1-1446::HOcs::URA3::AVT2 region with Ch V sequences 29,146 to 32,976 was amplified from JRL472 and integrated distal to the HO cut site into Ch V in strain JRL346 to generate JRL475 (can1,1-1446::HOcs::URA3::AVT2 ykl215c::leu2::hisG::can1DEL1-289::AVT2). As a result, there are Ch V sequences 33,177–32,020 shared between the donor and sequences proximal to the break, Ch V sequences 31,644–29,240 shared between the donor and sequences distal to the break and a 376 bp gap of homology. All mutant strains were created by standard gene disruption methods and confirmed by PCR. Plasmid pSJ5 was constructed by subcloning a XhoI-NotI fragment containing the RAD51 ORF under the PGK promoter form pNSU256 [47] into pRS314 [73].
Logarithmically growing cells grown in YEP+2% Raffinose, or the appropriate drop-out media +2% Raffinose, were plated on either YEPD or YEP-Gal, and grown into colonies. Colonies were counted and were then replica plated onto plates containing either canavanine or hygromycin to confirm repair occurred by BIR. Experiments were performed at least 5 times for each strain unless otherwise indicated. To determine the statistical significance between strains the student's t-test was used (paired, two-tailed, n≥4 for all strains).
Strains were grown in YEP+2% Raffinose to a cell density of 3×10e6 to 1×10e7 cells/mL. A 50 mL aliquot of cells was removed for the zero time point. Freshly made galactose was added to final concentration of 2% to induce HO expression. Cell aliquots were taken at the indicated time points throughout the time course.
PCR analysis of BIR was performed as previously described [8]. Briefly, we monitor the initiation of new BIR DNA synthesis using a PCR assay in which one primer is specific to Ch V and the other primer is specific to the donor sequence on Ch XI. Once a covalent molecule is formed, corresponding to the first 242 bp of new DNA synthesis, we see PCR product. At least three PCR reactions from three different experiments were performed for wild type, sgs1Δ and exo1Δ strains. For all other strains tested, at least three PCR reactions from two experiments were performed. The technical replicates from each biological experiment was first averaged and then the technical averages were averaged among the two experiments to obtain a biological average. Data were graphed as the biological averages normalized to the maximum product obtained by amplifying DNA from a strain that has repaired the DSB by BIR. Error bars represent the data range between the biological averages.
Repair is also measured by Southern blot that detects approximately the first 3 kb of new DNA synthesis was performed as previously described [8]. The analysis by Southern blot or pulse-field (CHEF) gel electrophoresis followed by Southern blot was performed as described [8] using the probes indicated in Figure 1. The breakpoints and sequences of sgs1Δ exo1Δ CanR HphS repaired colonies were performed as described [52], [53].
|
10.1371/journal.ppat.1003932 | Unifying Viral Genetics and Human Transportation Data to Predict the Global Transmission Dynamics of Human Influenza H3N2 | Information on global human movement patterns is central to spatial epidemiological models used to predict the behavior of influenza and other infectious diseases. Yet it remains difficult to test which modes of dispersal drive pathogen spread at various geographic scales using standard epidemiological data alone. Evolutionary analyses of pathogen genome sequences increasingly provide insights into the spatial dynamics of influenza viruses, but to date they have largely neglected the wealth of information on human mobility, mainly because no statistical framework exists within which viral gene sequences and empirical data on host movement can be combined. Here, we address this problem by applying a phylogeographic approach to elucidate the global spread of human influenza subtype H3N2 and assess its ability to predict the spatial spread of human influenza A viruses worldwide. Using a framework that estimates the migration history of human influenza while simultaneously testing and quantifying a range of potential predictive variables of spatial spread, we show that the global dynamics of influenza H3N2 are driven by air passenger flows, whereas at more local scales spread is also determined by processes that correlate with geographic distance. Our analyses further confirm a central role for mainland China and Southeast Asia in maintaining a source population for global influenza diversity. By comparing model output with the known pandemic expansion of H1N1 during 2009, we demonstrate that predictions of influenza spatial spread are most accurate when data on human mobility and viral evolution are integrated. In conclusion, the global dynamics of influenza viruses are best explained by combining human mobility data with the spatial information inherent in sampled viral genomes. The integrated approach introduced here offers great potential for epidemiological surveillance through phylogeographic reconstructions and for improving predictive models of disease control.
| What explains the geographic dispersal of emerging pathogens? Reconstructions of evolutionary history from pathogen gene sequences offer qualitative descriptions of spatial spread, but current approaches are poorly equipped to formally test and quantify the contribution of different potential explanatory factors, such as human mobility and demography. Here, we use a novel phylogeographic method to evaluate multiple potential predictors of viral spread in human influenza dynamics. We identify air travel as the predominant driver of global influenza migration, whilst also revealing the contribution of other mobility processes at more local scales. We demonstrate the power of our inter-disciplinary approach by using it to predict the global pandemic expansion of H1N1 influenza in 2009. Our study highlights the importance of integrating evolutionary and ecological information when studying the dynamics of infectious disease.
| The emergence and worldwide dispersal of novel human pathogens is increasingly challenging global public health [1]. Notable recent examples include novel influenza strains, severe acute respiratory syndrome (SARS) virus and Methicillin-resistant Staphylococcus aureus, which all exploit today's complex and voluminous transport networks to rapidly disseminate in a globalized world. In the context of human infectious diseases, the worldwide air transportation network is by far the best studied system of global mobility [2]. Air travel likely drives the global circulation of seasonal influenza A (H3N2) viruses [3], and may explain seasonal dynamics in the absence of locally-persistent strains between epidemic seasons. Retrospective modeling of the ‘Hong Kong flu’ H3N2 pandemic in 1968 indicates that the virus spread through a global network of cities interconnected by air travel [4]. Numerous modeling and simulation studies have subsequently explored the potential influence of air travel on influenza virus spread, e.g. [5]–[8], but few have attempted to verify such models against underlying empirical data on human movement patterns [9].
Two studies on the timing and rate of seasonal influenza transmission across the United States of America (USA) highlight the difficulty of using standard epidemiological data to disentangle the relative contributions of different human transportation systems to influenza spread. Using weekly time series of excess mortality due to pneumonia and influenza (P&I), Viboud et al. [9] demonstrated that the patterns of timing and incidence of outbreaks across the continental USA are most strongly associated with rates of movement of people to and from their workplaces, and to a lesser extent with the distance between locations and various measures of domestic transportation. In contrast, Brownstein et al. [10] concluded that the rate of inter-regional spread and timing of influenza in the USA, as measured using weekly P&I mortality statistics, is predicted by domestic airline travel volume in November. These discordant findings generated significant debate [11], especially in the context of a potential pandemic of pathogenic influenza [12], which would require rapid decisions to be made on the implementation of travel restrictions.
As a historical record of epidemic spread, viral genetic sequence data may offer a valuable source of information for the empirical verification of epidemiological models. Several studies have demonstrated their utility and power, for example by revealing the genetic dynamics of influenza A H3N2 seasonality [13] and the spatial patterns of global H3N2 circulation [3], [14]. More generally, it is recognized that the genetic diversity of rapidly evolving viruses like influenza should be analysed in a framework that unifies evolutionary and ecological dynamics [15]. Current attempts to reconstruct viral spread through time and space from genetic data, however, typically fit parameter-rich models to sparse spatial data and result in phylogeographic patterns that are difficult to relate directly to underlying ecological processes [16]. Together with potential sampling bias, this complicates phylogeographic tasks, such as the characterization of source-sink dynamics in seasonal influenza. It is therefore unsurprising that different studies on the global circulation of H3N2 are sometimes inconsistent [3], [14], [17], despite the importance of such work for influenza surveillance and vaccine strain selection.
Here we use a model-based approach to explicitly tests spatial epidemiological hypotheses by integrating empirical data on human movement patterns with viral genetic data. This framework enables us to measure the relative contribution of different predictive variables to viral spatial spread. We apply this approach to seasonal H3N2 dynamics and use it to identify key drivers of the global dissemination of influenza viruses. Analysis of different sampling schemes, including one that represents the community structure in global air transportation, provides consistent support for air travel governing the spatial dynamics of seasonal H3N2 infections. Using epidemiological simulations, we further demonstrate that estimates resulting from the merger of human air travel and H3N2 influenza genetics best capture the observed global expansion of pandemic H1N1 influenza in 2009.
We complemented a previously collected hemagglutinin sequence data set, comprising 1,441 sequences sampled globally from 2002 to 2007 [3], with publicly available sequences sampled within the same time interval. The allocation of the sequence data into 15 and 26 geographic regions as well as into 14 air communities is described in detail in Supporting information Text S1.
The worldwide air transportation network is defined by a passenger flux matrix that quantifies the number of passengers traveling between each pair of airports. We use a dataset provided by OAG (Official Airline Guide) Ltd. (http://www.oag.com), containing 4,092 airports and the number of seats on scheduled commercial flights between pairs of airports during the years 2004–2006. We take the number of seats on scheduled commercial flights from airport i to j to be proportional to the number of passengers traveling.
To identify air transportation communities, we approximate a maximal-modularity subdivision of the 1,227-largest-airport network by employing a recently described stochastic Monte-Carlo approach [18]. Modularity provides a measure of how well the connectivity of a network is described by partitioning its nodes into non-overlapping groups; for a definition we refer to [19]. For any given partition, modularity will be high if connectivity within groups is high and connectivity among groups is low. For large networks, a variety of methods have been introduced to approximate their optimal subdivision. The method we employ here generates an ensemble of high modularity subdivisions and computes the consensus in this ensemble by superposition. For further details we refer to [18], [20] and in Text S1 we describe how we incorporate subdivision uncertainty in our phylogeographic approach.
We employ a novel approach to simultaneously reconstruct spatiotemporal history and test the contribution of potential predictors of spatial spread. The approach extends a recently developed Bayesian method of phylogeographic inference [21] into a generalized linear model (GLM), by parameterizing each rate of among-location movement in the phylogeographic model as a log linear function of various potential predictors. For each predictor j, the GLM parameterization includes a coefficient , which quantifies the contribution or effect size of the predictor (in log space), and a binary indicator variable , that allows the predictor to be included or excluded from the model. We estimate the variables using a Bayesian stochastic search variable selection (BSSVS) [22], [23], resulting in an estimate of the posterior inclusion probability or support for each predictor. This approach uses the data to select the explanatory variables and their effect sizes from a pre-defined set of predictors that can explain the phylogenetic history of among-location movement while simultaneously reconstructing the ancestral locations in the evolutionary history. In Text S1, we (i) provide more mathematical detail of the GLM model, (ii) describe novel transition kernels for efficient statistical inference, (iii) propose prior specifications and (iv) explain how Bayes factors can be calculated for each predictor based on estimates. The method introduced here is implemented in the BEAST software package [24].
The GLM approach offers many statistical advantages over other approaches [25] in efficiently testing spatial hypotheses (see Text S1 for a detailed comparative analysis). Commonly-used Bayesian measures of model fit (such as marginal likelihood estimation using the harmonic mean), which can be applied to models with among-location movement rates fixed to a particular predictor, have been shown to perform poorly [26]–[28]. Although more accurate alternatives have recently been proposed [26]–[28], they are computationally prohibitive on large data sets such as those studied here. Importantly, the previous approach provides only a relative ranking of different models and, unlike the GLM model, cannot identify which of the top-ranked predictors need to be jointly considered as explanatory variables. A further advantage of the GLM approach is that in addition to providing a measure of support for each predictor, it can also quantify the contribution or effect size of each predictor by estimating the associated coefficients ().
For the spread of seasonal influenza, we consider several potential predictors of global migration, including different log-transformed measures of geographical distance, absolute latitude, air transportation data, demographic and economic data, viral surveillance data, antigenic evolution and sequence sample sizes (described in more detail in Text S1). Text S1 also reports the evolutionary and demographic models used in BEAST and describes how phylogenetic uncertainty is approximated during phylogeographic inference.
Phylogeographic movement events among locations are modeled by a continuous-time Markov chain (CTMC) process along each branch of the viral phylogeny. Although both the transitions among locations (Markov jumps) and the waiting times between transitions (Markov rewards) are not directly observed, posterior expectations of these values can be efficiently computed [29], [30]. Here, we implement posterior inference of the complete Markov jump history through time in BEAST and use these estimates to assess the source-sink dynamics of influenza and to evaluate the predictive performance of phylogeographic models.
To compare the performance of different migration rate models in predicting global pandemic spread, we simulate a stochastic meta-population susceptible-infected-recovered (SIR) model with n = 14 populations, matching the 14 air communities analyzed in the phylogeographic model. The model tracks the number of susceptible (S), infected (I) and recovered (R) individuals in each population each day of the simulation. The simulations begin with a single initial infection in Mexico on January 5th 2009 [31]. Infection spreads through mass-action within each population according to the following epidemiological parameters. Population-specific host population size is equal to human population size (Text S1). Basic epidemiological parameters are based on empirical estimates from H1N1: the duration of infection was chosen as 3 days [31] and the basic reproductive number () or average number of secondary infections arising from a primary infector during their infectious period in a completely susceptible population was chosen as 1.3 [31]. This results in a transmission rate . Although estimates of for pandemic H1N1 vary across studies, the exact value is unlikely to affect the comparative simulations we perform as this is expected to equally impact the overall expansion rate and not the relative migration dynamics across populations. Force of infection within population scales with infected frequency across populations following , where the coupling coefficient represents the rate of contacts from population i to population j relative to within-population contacts and . Other pairwise coupling coefficients are taken to be proportional to pairwise migration estimates, so that , where is the air travel based or phylogenetically estimated rate of migration from population i to population j per year and parameter c is fitted to the data. Parameter c is the only free parameter in this model and we set this to the value that maximizes correspondence between simulations and observations (see below). This ensures that we can use phylogeographic migration rates as per capita migration rates in the simulation model, despite their different scales. Compartments are updated according to a -leaping algorithm [32] with one-day intervals.
Migration rates between populations in the SIR model are defined according to four scenarios, as follows: (A) equal rates, (B) rates proportional to the amount of air travel occurring between them (in terms of the number of passengers moving from one population to another), (C) rates proportional to Markov jump estimates based on a standard phylogeographic model (undertaken with and without BSSVS to reduce the number of rate parameters) and (D) a GLM model that only considers air travel as a predictor. To compare the spread of influenza under these simulated models to recorded H1N1 pandemic spread, we measure the relative correspondence between the mean peak times (across 100 simulations) and the observed peak times for all locations except Mexico (based on World Health Organization data; Text S1). Correspondence was measured using the Spearman's rank correlation coefficient, and tested with associated -values obtained using a permutation test (Text S1), as well as using the mean average error (MAE; in days). We consider the Spearman's rank correlation coefficients to be more appropriate for our comparison because they are more robust to outliers, which are clearly present in the observed peaks. Therefore, the scaling of between-population coupling c for the various migration matrices was also adjusted so as to maximize Spearman's rank correlation.
To identify key factors in the seasonal dispersal of human influenza viruses, we use a Bayesian model selection procedure to estimate the phylogeographic history of H3N2 viruses sampled worldwide between 2002 and 2007 (Text S1), while concurrently evaluating the contribution of several potential predictors of spatial spread. In addition to considering two geographic discretizations of the available data, we also identify community structure in global air travel by determining partitions with high intra-community connectivity and low inter-community connectivity (Methods). Although this approach is blind to the airports' geographic locations, the 14 resulting global air communities are spatially compact with few exceptions (Fig. 1). We find air communities that are largely specific to Oceania, China, Japan, Sub-Saharan Africa, Mexico and Canada. Madagascar, Réunion and some Caribbean destinations are examples of exceptions that are, as non-European locations, connected to a European air community.
Our analysis reveals that many potential predictors of global influenza virus spread are not associated with viral lineage movement, specifically, geographical proximity, demography and economic measures, antigenic divergence, epidemiological synchronity and seasonality do not yield noticeable support (Fig. 2). Instead, we find consistent and strong evidence that air passenger flow is the dominant driver of the global dissemination of H3N2 influenza viruses. This is reflected in both the estimated size of the effect of this variable ( on a log scale) and the statistical support for its inclusion in the model (posterior probability >0.93 and Bayes factor >760). This effect size means that viral lineage movement rates are about 15 times higher for connections with the highest passenger flow compared to connections with the lowest flow, controlling for all other predictors. The result is robust when we repeat the analysis (i) using different partitions of sampling locations (air communities and different geographic partitions, Fig. 2), (ii) using different sequence sub-samples for the air communities (Fig. S1), (iii) using the full data set or a small but more balanced number of sub-samples (Fig. S2), and (iv) using a more liberal prior specification on predictor inclusion (Fig. S3). We down-sampled particular air communities or geographic regions relative to their population sizes (Text S1), which still leaves considerable heterogeneity in sample sizes, explaining why they are included as an explanatory variable in the GLM model. Our aim is not to demonstrate a role for sample sizes in phylogeography, but by explicitly including them as predictive variables, we raise the credibility that other predictors are not included in the model because of sampling bias. We note that the sample size predictors may in fact absorb some of the effect of air travel because a GLM model that only considers passenger flux as a predictor of H3N2 movement among the air communities results in a higher mean effect of size of about 1.5.
To also explore spatial dynamics at smaller scales, we further partition large geographical regions that are administratively coherent, such as the USA, China, Japan and Australia, resulting in 26 global sampling regions (Text S1). In this analysis, air travel again predicts viral movement (posterior probability >0.99 and Bayes factor >18000), but the movement is also inversely associated with geographical distance between locations (posterior probability = 0.76 and Bayes factor = 87), and, less intuitively, with origin and destination population densities (although the size of the latter effects are weaker, Fig. 2). The negative association of population density with viral movement may suggest that commuting is less likely, per capita, to occur out of, or into, dense subpopulations.
Although not the main focus of the current study, our integrated approach also provides phylogeographic reconstructions that offer insights into the global source-sink dynamics of human influenza. The trunk or backbone of phylogenies reconstructed from temporally-sampled hemagglutinin genes (Fig. 3) represents the lineage that successfully persists from one epidemic year to the next [14], [33]. We determine the spatial history of this lineage using Markov rewards in the posterior tree distribution, thereby estimating the contribution of each location to the persistence of the trunk lineage from 2002 to 2006 (Fig. 3). These estimates provide strong support for mainland China as the principal H3N2 source population, occupying close to 60% of the trunk time in the H3N2 phylogenies (Fig. 3), followed by Southeast Asia, which comprises about 15% of the trunk time. We further examine temporal heterogeneity in the source-sink process by combining a summary of the estimated trunk location through time together with an phylogenetic summary in Fig. 3, which suggests that the above-mentioned proportions arose from the presence of the trunk lineage in China during 2002 to mid 2003 and late 2004 to 2006, interrupted by a period when the virus appeared to have a Southeast Asian H3N2 source. However, we cannot rule out the impact of temporal sampling heterogeneity on these estimates because the Southeast Asian trunk dominance precedes a period of higher sampling availability for Southeast Asia relative to mainland China (Fig. 3). The important role of mainland China in seeding the global seasonal spread of human influenza results in a high net migration out of this air community (Fig. S4). However, air communities that do not contribute significantly to the trunk can also maintain high net outflow, in particular the USA, which may be seeded by relatively few introductions each year whilst exporting comparatively more viruses to other locations during the epidemic season.
In order to assess the extent to which evolutionary analyses such as ours benefit from integrating host mobility data, we examine their predictive performance by using them to predict the relative timing of the geographic spread of the pandemic H1N1 influenza variant that emerged in 2009. We conduct simulations of the spread of a novel pathogen out of Mexico using an SIR model whose transmission parameters are informed by epidemiological estimates obtained for pandemic H1N1 [31] and whose spatial spread is determined by one of four different migration rate models, each defined by a different matrix of movement rates among all pairs of locations (Methods). We measure the relative correspondence between the simulated and observed H1N1 peaks for each location except Mexico using a Spearman's rank correlation coefficient () and mean absolute error (MAE; in days)(Fig. 4).
An equal rates model (A), which does not express any migration rate preference, results in a weak match (, P = 0.73, MAE = 40.9 days) between the simulations and the observed spatial spread of H1N1 (Fig. 4), indicating that the population sizes included in the SIR model for each region offer limited predictive performance. As expected, adding information on the number of airline passengers (model B) yields a large improvement in correspondence between simulations and observations (, P = 0.03, MAE = 35.8 days). In contrast, a standard parameter-rich phylogeographic model that is only informed by sequence data and not air traffic information (model C) yields only part of this improvement in predictive performance (, P = 0.10, MAE = 39.4 days). However, if inference under model C is made more efficient by focusing on a small set of parameters (using BSSVS, [21]; see Methods) then phylogeographic estimates yield a predictive performance (, P = 0.02, MAE = 36.4 days, Fig. S5) that is close to that of the air travel model (B). Finally, the GLM model (D) predicts the observed spread of H1N1 more accurately than all other models (, P<0.01, MAE = 32.3), suggesting that global influenza transmission is best predicted by combining passenger flux data with the information on viral lineage movement contained in sequence data. The simulations generally correspond better with observed H1N1 peaks during the initial period of pandemic expansion, while the epidemic peaks for Russia and Africa occur significantly earlier in the simulations than in reality. This is likely due to the multi-peaked character of the regional epidemics (Text S1); the H1N1 virus spreads to most of the world during the first pandemic wave, whereas regions like Russia and Africa appeared to miss the first wave entirely. Seasonal effects that are unaccounted for by our simulation may at least partly explain the outliers, but they affect the models we aim to compare in a very similar way. Because of the outliers, we consider the non-parametric Spearman's to be a more appropriate measure of correspondence than the MAE, but they are consistent in their model ranking. We note that absolute prediction errors can be considerably improved by only considering the 9 air communities that peaked prior to September, 2009, which returns a MAE of 11.2 day for the GLM model. However, because of the difficulties in establishing initial waves and their peaks, and the uncertainty in our epidemiological model, we caution against more detailed interpretation of these simulations beyond the general trends we extract here.
The prevention and control of influenza at the global scale relies critically on our understanding of its mode of geographical dissemination. Here, we demonstrate that such dynamics are most powerfully investigated by combining phylogeographic history with empirical data on the patterns of human movement worldwide. Our analysis strongly suggests that air travel is key to global influenza spread, an intuitive result that has long been predicted by modeling studies (e.g. [5]), but has, until now, remained difficult to obtain from empirical data. The dominant predictors of influenza spread will undoubtedly be scale-dependent, as indicated here by the importance of geographic distance as a predictor within more confined geographic areas (Fig. 2), which may represent forms of human mobility other than air travel, such as workplace commuting [9]. This indicates that our statistical framework could also prove valuable in testing hypotheses at smaller scales, where the underlying spatial processes may be less obvious, provided adequate sequence and empirical movement data are available. One of the limitations of the current heterogeneous sampling of H3N2 sequences worldwide is that geographic partitions need to be adjusted to account for the number of samples per location, which results in regions of widely different areas and population sizes. More representative sampling across the globe, or within a more geographically confined area of interest, will allow for more appropriate geographic partitioning and may facilitate more detailed spatial hypothesis testing based on the associated demographic and mobility measures. In particular, if sequences were sampled appropriately then our inference method could incorporate the rich geographic data that is currently available as global gridded population data sets [34]. In addition, many of the predictors used here can be improved in accuracy and resolution, for example by accounting for seat occupancy and actual origin-destination flows in air traffic passenger fluxes.
Due to the difficulties associated with geographic partitioning, we used algorithms to optimally define communities in the global air transportation network as an alternative strategy to specify phylogeographic states, and subsequently show that our GLM results are robust to the different partitions used. Because air travel is a consistent and highly supported explanatory variable for global influenza dispersal, communities within the air transportation network are likely to provide the most appropriate spatial structuring of our data. However, in addition to the partitioning itself, further research is also needed to select the appropriate number of samples from the resulting regions to improve on ad hoc down-sampling based on population size.
Although identifying the causes of pathogen spread is of great importance in spatial epidemiology, integrating this information in evolutionary models also offers major advantages for phylogeographic reconstructions and their relevance to infectious disease surveillance and pandemic preparedness. By capturing a more realistic process of spatial spread, our novel approach results in more credible reconstructions of spatial evolutionary history, which may shed further light on the persistence and migration dynamics of human influenza viruses. Because of the importance of influenza dynamics for vaccine strain selection, different phylogeographic reconstructions have attempted to characterize the global population structure of the virus and have arrived at somewhat mixed findings [3], [14], [17]. This may be explained by the use of both different sampling and different methodology. The data and methods used here corroborate the explorations of antigenic and genetic divergence by [3] and demonstrate the prominence of mainland China and Southeast Asia as locations of trunk lineage persistence. Our findings are however based on roughly the same genetic data, and our approach of inferring the spatial history of the trunk lineage through Markov reward estimates may be viewed as the more direct, statistical equivalent of measuring strain location distance from the trunk [3]. Although we find a strong signal for the presence of the trunk lineage in mainland China and Southeast Asia, our analysis is restricted to the period 2002 to 2006, and thus we make no conclusions about the location of the trunk lineage outside of this period. The degree of temporal stochasticity in the source location of seasonal influenza and its heterogeneity among different influenza variants has yet to be determined and requires datasets of longer duration. Moreover, we suggest that analyses of future data sets that are more comprehensively sampled through time will also benefit from phylogeographic models that can accommodate temporal heterogeneity in movement rates. Such models may also improve the performance of some explanatory variables. For example, in the analysis presented here, we do not consider the absence of support for seasonality as a predictor in our GLM model as evidence against seasonality in H3N2 spread. Rather, it simply reflects the difficulty in incorporating seasonality into a time-homogeneous model of lineage movement. Developments are now underway to appropriately accommodate heterogeneity in spatial spread through time.
By using models to predict the observed global emergence of pandemic H1N1, we demonstrate that an approach that integrates passenger flux data with viral genetic data provides a more accurate prediction of global epidemic spread than those which include only one source of information. Although the prediction improvement of the combined data over the passenger flux data alone is not very large, it remains remarkable because we attempt to predict the spatial expansion of an epidemic lineage (pandemic H1N1) from the seasonal dynamics of another lineage (H3N2) and because the main process underlying the global dispersal of H3N2 influenza appears to be air travel itself. Passenger flux data among pairs of locations is symmetric, thus it is possible that the phylogeographic data is capable of capturing asymmetry in the seasonal process of viral spread, which may also be important in explaining the spatial expansion of pandemic H1N1. Investigations using more advanced simulation techniques, e.g. [35], may be able to build upon the conceptual bridge between genetic data and epidemiological modeling implied by our findings. Future prediction efforts may also need to focus on alternative scenarios of spatial spread, as highlighted by the recent emergence of a novel avian influenza H7N9 lineage in China [36]. Should this virus evolve sustained human-to-human transmissibility, then airline-passenger data and flight routes from the outbreak regions in particular, would be able to pinpoint worldwide regions of immediate risk. If the virus remains restricted to avian hosts, however, risk maps for the transmission of avian influenza viruses (perhaps based on predictors calibrated against H5N1 avian influenza) may help to target H7N9 surveillance and control efforts. In conclusion, our framework is applicable to different infectious diseases and provides new opportunities for explicitly testing how host behavior and ecology shapes the spatial distribution of pathogen genetic diversity.
|
10.1371/journal.pmed.1002531 | Causes of death and infant mortality rates among full-term births in the United States between 2010 and 2012: An observational study | While the high prevalence of preterm births and its impact on infant mortality in the US have been widely acknowledged, recent data suggest that even full-term births in the US face substantially higher mortality risks compared to European countries with low infant mortality rates. In this paper, we use the most recent birth records in the US to more closely analyze the primary causes underlying mortality rates among full-term births.
Linked birth and death records for the period 2010–2012 were used to identify the state- and cause-specific burden of infant mortality among full-term infants (born at 37–42 weeks of gestation). Multivariable logistic models were used to assess the extent to which state-level differences in full-term infant mortality (FTIM) were attributable to observed differences in maternal and birth characteristics. Random effects models were used to assess the relative contribution of state-level variation to FTIM. Hypothetical mortality outcomes were computed under the assumption that all states could achieve the survival rates of the best-performing states. A total of 10,175,481 infants born full-term in the US between January 1, 2010, and December 31, 2012, were analyzed. FTIM rate (FTIMR) was 2.2 per 1,000 live births overall, and ranged between 1.29 (Connecticut, 95% CI 1.08, 1.53) and 3.77 (Mississippi, 95% CI 3.39, 4.19) at the state level. Zero states reached the rates reported in the 6 low-mortality European countries analyzed (FTIMR < 1.25), and 13 states had FTIMR > 2.75. Sudden unexpected death in infancy (SUDI) accounted for 43% of FTIM; congenital malformations and perinatal conditions accounted for 31% and 11.3% of FTIM, respectively. The largest mortality differentials between states with good and states with poor FTIMR were found for SUDI, with particularly large risk differentials for deaths due to sudden infant death syndrome (SIDS) (odds ratio [OR] 2.52, 95% CI 1.86, 3.42) and suffocation (OR 4.40, 95% CI 3.71, 5.21). Even though these mortality differences were partially explained by state-level differences in maternal education, race, and maternal health, substantial state-level variation in infant mortality remained in fully adjusted models (SIDS OR 1.45, suffocation OR 2.92). The extent to which these state differentials are due to differential antenatal care standards as well as differential access to health services could not be determined due to data limitations. Overall, our estimates suggest that infant mortality could be reduced by 4,003 deaths (95% CI 2,284, 5,587) annually if all states were to achieve the mortality levels of the best-performing state in each cause-of-death category. Key limitations of the analysis are that information on termination rates at the state level was not available, and that causes of deaths may have been coded differentially across states.
More than 7,000 full-term infants die in the US each year. The results presented in this paper suggest that a substantial share of these deaths may be preventable. Potential improvements seem particularly large for SUDI, where very low rates have been achieved in a few states while average mortality rates remain high in most other areas. Given the high mortality burden due to SIDS and suffocation, policy efforts to promote compliance with recommended sleeping arrangements could be an effective first step in this direction.
| High infant mortality rates in the US compared to other high-income countries have been well documented in the literature.
Most of this literature primarily attributes high infant mortality in the US to the high rates of prematurity.
Relatively little is known regarding the survival of infants born full-term.
We compared state-level mortality rates among full-term infants in the US to that of 6 European countries with low mortality rates.
We showed that infants born full-term in the US face 50%–200% higher risks of infant mortality compared to these European countries.
We found that the largest proportion of infant deaths among children born full-term in the US was due to sudden unexpected deaths of infants, which comprised both sudden infant death syndrome and other unexpected causes such as suffocation and violence.
Major improvements in full-term infant mortality (through increases in full-term infant survival and increases in pregnancy terminations) seem possible in the US.
More research is needed to identify the most effective policies to achieve this objective.
| Despite some progress made in recent years, infant mortality rates in the US continue to be high compared to other high-income countries [1]. According to the latest estimates, the US currently ranks 44th among 199 countries of all income levels, with an infant mortality rate of 5.6 deaths per 1,000 live births in 2015, about 3 times the rate observed for countries at the very top of the ranking [1].
While the high rates of prematurity and prematurity-related mortality in the US have been well documented in the literature [2,3], the US performs comparably to other high-income countries when it comes to the survival of preterm infants. Fig 1 compares gestation-specific mortality rates in the US and 6 leading European countries (in terms of low infant mortality rates) with data available for 2010. On average, infant mortality appeared to be very similar for premature births in the US and in these European countries. The same was not true for children born after 36 weeks of gestation, where children born in the US faced more than twice the mortality risk of children in European countries with low infant mortality rates (odds ratio [OR] 2.02, 95% CI 1.84, 2.22). A recent US Centers for Disease Control and Prevention (CDC) report suggests that this mortality gap among full-term births now accounts for almost 50% of the infant mortality gap between Sweden and the US [4].
In this study, we used complete and geocoded birth records from the period 2010–2012 to better understand the high burden of mortality among full-term infants in the US. We identified the main causes underlying the high mortality rates among full-term infants overall in the aggregate data in a first step, and then explored differences in actual and potential birth outcomes across US states in a second step. By first reviewing the causes of death in this population, we could identify the main risk factors for infants in this generally low-risk population, and could clearly distinguish the relative importance of preexisting conditions such as malformations relative to perinatal and post-neonatal conditions (those arising in the 28–364 days after birth). In order to provide a better sense of feasible outcomes in this population, we estimated and compared cause-specific full-term mortality rates at the state level both unconditional and conditional on maternal characteristics. While these state-level comparisons did not allow us to identify the specific reasons why certain states have particularly high rates of mortality, they did allow us to identify areas where major improvements were possible in principle.
The study was designed as a cross-sectional study using birth and death records of all infants born in the US between January 1, 2010, and December 31, 2012. No pre-analysis plan was developed for this study. The main objective of the project was to identify the primary causes underlying the high infant mortality rates observed in the US nationally as well as at the state level.
Linked birth and death records including restricted geographic identifiers were obtained from the National Center for Health Statistics (NCHS) for the years 2010 to 2012. All infants in the birth and death records could be directly linked to the geographic identifiers in these datasets (100% match rate). Additional data for Fig 1 were downloaded from the Euro-Peristat webpage at http://www.europeristat.com/our-indicators/euro-peristat-perinatal-health-indicators-2010.html.
Our primary outcome measure of interest was the infant mortality rate among full-term births defined as the number of deaths per 1,000 children born alive between 37 and 42 weeks of gestation within the first year of their life. For the purpose of this study, we used the traditional definition of full-term, which includes early-term (37 and 38 weeks), full-term (39 and 40 weeks), late-term (41 weeks), and some post-term (42 weeks) births according to the more recent definition of the American College of Obstetricians and Gynecologists Committee on Obstetric Practice [5]. To adjust for differential outcomes in this relatively wide 5-week gestational window, we controlled for differences in gestational age by including binary indicators for gestational age category (37 weeks, 38 weeks, 41 weeks, 42 weeks) in our multivariable analysis, using the more narrow, revised full-term definition (39 and 40 weeks) as our reference group. Gestational age was computed by the NCHS based on last menstrual period reported by the mother. To ensure gestational age was not measured differentially across states, we compared prematurity rates with rates of low birth weight in the full sample at the state level. The correlation of these measures at the state level was 0.97; the strong alignment between birth weight and reported gestational age is further supported by the descriptive statistics provided in S1 Table.
Causes of death for all children who died under the age of 1 year were based on death certificates, which are required to be completed by either a coroner or medical examiner in all US states, following CDC guidelines. Even though regulations vary by state, deaths due to violence or suspicious circumstances are further investigated and certified by a medical legal officer [6].
Death certificates were reviewed and coded following ICD-10 guidelines by the NCHS. For the purpose of this paper, we grouped reported causes of death into 4 main categories: (1) congenital malformations: ICD-10 codes Q00–Q99; (2) sudden unexpected death in infancy (SUDI): ICD-10 codes V01–Y89 and R00–R99; (3) perinatal conditions: ICD-10 codes P00–P96; and (4) all other causes: all other ICD-10 codes.
The SUDI grouping was chosen intentionally to minimize potential state-level differences in the attribution of unexplained deaths to sudden infant death syndrome (SIDS) versus “other unexplained causes” [7–9]. Some more disaggregated statistics for major causes of deaths (such as SIDS) were also computed as described further below.
Children born prior to 37 or after 42 weeks of gestation were excluded from this study. All other children born alive in the US between January 1, 2010, and December 31, 2012, including multiple births and children born with malformations (not reported in the NCHS dataset), were analyzed in this study.
In order to assess the extent to which state-level differences in infant mortality rates can be attributed to differences in maternal characteristics, we considered the following variables included in the original data file: mother’s age, educational attainment, smoking behavior, diabetes, chronic hypertension, and eclampsia. We divided maternal age into 5 categories (<20, 20–34, 35–39, 40–44, and >44 years) and used age 20–34 as the reference group in our multivariable analysis. Similarly, we divided maternal educational attainment into 4 categories: less than high school, high school or some college credit without degree, associate or bachelor’s degree, and master’s degree or doctorate. In response to a reviewer request, we also added controls for mother’s race: white, black, American Indian/Alaskan Native, and Asian/Pacific Islander. As for smoking, mothers reported the average number of cigarettes smoked per day during their first, second, and third trimesters. From this we constructed indicators for smoking (number of cigarettes per day > 0) or not for each trimester. We used indicators for previous diagnosis of diabetes, chronic hypertension, and eclampsia as provided in the dataset. All these variables were based on mother’s self-report in the hospital around the time of delivery and were reported on the birth certificate. In addition, we included controls for the following birth characteristics: birth weight category (<1,500, 1,500–1,999, 2,000–2,499, 2,500–2,999, 3,000–3,499, 3,500–3,999, 4,000–4,499, and >4,499 g), multiple birth (1 if singleton, 2 if twin, 3 if triplet, 4 if quadruplet, and 5 if quintuplet or higher), infant sex, and gestational age (indicators for 37 weeks, 38 weeks, 41 weeks, and 42 weeks of gestation) in our empirical models. Further details of all these variables are provided in S1 Table.
As a first step, we computed full-term infant mortality rates (FTIMRs) at the state level, and classified all US states into 5 groups: states with excellent FTIMR (FTIMR < 1.25—the European benchmark shown in Fig 1), states with good FTIMR (1.25 ≤ FTIMR < 1.75), states with average FTIMR (1.75 ≤ FTIMR < 2.25), states with fair FTIMR (2.25 ≤ FTIMR < 2.75), and finally states with poor FTIMR (FTIMR ≥ 2.75). The “excellent” group was chosen based on the FTIMRs observed in 6 European countries (Austria, Denmark, Finland, Norway, Sweden, and Switzerland), which ranged between 0.97 and 1.24, with a median FTIMR of 1.11. The remaining groups were defined by sequentially adding 0.5 deaths per 1,000 full-term live births (a 50% increase relative to the European average) to the cutoffs. In a second step, we decomposed mortality differences at the group level by cause of death. Third, we used multivariable regression models to assess the extent to which survival differences across states can be attributed to observable differences in maternal and birth characteristics. To do so, we first ran multivariable logistic models comparing infants born in the states with the highest mortality rates to infants born in the states with the lowest mortality rates. We estimated 3 separate models: a first model, where we did not adjust for any covariates; a second model, where we adjusted for maternal characteristics outlined in the covariates section above; and a third model (proposed by a reviewer), where we adjusted for maternal characteristics and birth characteristics (gestational age, infant sex, birth weight, and multiple birth). Model 2 was estimated to assess the extent to which state-level differences can be attributed to local variation in maternal characteristics such as age, education, race, and health status. Model 3 was estimated to assess the extent to which subsequent mortality differentials were explained by local variation in the prevalence of multiple births as well as differences in birth weight and the distribution of gestational age. In all 3 models, each observation corresponded to a child born full-term in the sample period. To assess the overall contribution of state-level characteristics to variation in FTIMR, we also estimated multilevel logistic models where we nested individual observations within states, and then estimated between-state variance in unconditional models as well as in models conditioning on maternal and birth characteristics. Lastly, we computed hypothetical mortality rates (which we refer to as “counterfactuals”) under the assumptions that (i) all US states achieved the overall FTIMR of the best-performing states (good FTIMR group) and (ii) all US states achieved the specific FTIMRs of the best-performing state in each cause-of-death category.
A total 10,175,481 children born full-term in the US between January 1, 2010, and December 31, 2012, were analyzed. FTIMR was 2.19 (95% CI 2.16, 2.22) per 1,000 full-term live births in the pooled sample. At the state level, estimated FTIMR ranged between 1.29 (95% CI 1.08, 1.53) in Connecticut and 3.77 (95% CI 3.39, 4.19) in Missouri. No state was classified as excellent in terms of their FTIMR; 10 states including Connecticut were classified as good, 17 as average, 11 as fair, and 13 as poor FTIMR (see Fig 2 and S2 Table for details).
Fig 3 compares early neonatal (death in the first 6 days after birth), late neonatal (death between 7 and 27 days after birth), and post-neonatal (death 28–364 days after birth) mortality rates across mortality groups. While only relatively minor differences were found with respect to early neonatal mortality, large absolute and relative differences were found for the post-neonatal period, with an average of 9.5 (95% CI 9.1, 9.9) deaths per 10,000 full-term births in states classified as having good FTIMR and a mortality rate of 20.9 (95% CI 20.1, 21.6) deaths per 10,000 full-term births in the states classified as having poor FTIMR.
Fig 4 summarizes the main causes of full-term infant mortality (FTIM). SUDI accounted for the largest proportion of deaths overall (43%), followed by congenital malformations (31%) and perinatal conditions (11%). The mortality risk due to congenital malformations increased from 5.6 deaths per 10,000 full-term live births in states with FTIMR < 2.75 to 8.4 deaths in states with poor FTIMR. The risk of SUDI was 5.6 in the states classified as having good FTIMR and 15.4 in the states classified as having poor FTIMR. Observed absolute mortality differences between FTIMR groups were smallest for perinatal conditions, with an estimated mortality rate of 2.1 in the states with good FTIMR and an estimated mortality of 2.8 in states with poor FTIMR.
In S1 and S2 Figs, we provide further details on the primary causes of congenital malformations. The 2 most common causes of deaths due to congenital malformation were Edwards syndrome and congenital malformations of the heart, which accounted for 10.9% and 14.6% of congenital malformation deaths, respectively.
In terms of the underlying causes of SUDI, 42.2% of SUDIs were due to SIDS (ICD-10: R95), followed by unknown and ill-defined causes (ICD-10: R99), which accounted for 20.6% of SUDIs, and accidental suffocation and strangulation (ICD-10: W75), which accounted for 16.1% of SUDIs. S3–S7 Figs provide further details on the spatial distribution of cause-specific SUDIs.
S8 Fig summarizes the relative importance of the 4 mortality groups in the neonatal, late neonatal, and post-neonatal periods. Congenital malformations accounted for 58.1% and 43.3% of overall mortality in the early neonatal and late neonatal periods, respectively. Perinatal conditions accounted for 31.7% and 22.8% of mortality in the same periods. In the post-neonatal period—which accounted for the majority of deaths overall (63.5%, as shown in Fig 3)—the large majority (60%) of deaths were due to SUDI.
Table 1 shows estimated OR for the group of states with poor FTIMR compared to the group of states with good FTIMR for the 4 main cause-of-death categories displayed in Fig 4. The table shows unadjusted OR estimates and OR estimates adjusted for the full set of covariates summarized in S1 Table. In unadjusted models, living in a state with poor FTIMR was associated with an increased odds of FTIM due to perinatal conditions of 35% (OR 1.35, 95% CI 1.17, 1.56) as well as an increased odds of death due to congenital malformations of 51% (1.51, 95% CI 1.24, 1.85). Risk differentials were largest for SUDI, with an estimated OR of 2.75 (95% CI 2.46, 3.07). When we adjusted for maternal age, education, race, and measures of health status, estimated risk differentials declined for all risk factors, with the largest declines for SUDI, where estimated OR fell from 2.75 in unadjusted models to 1.70 in models adjusting for both maternal and birth characteristics. In general, differences between models 2 (adjusting for maternal characteristics only) and 3 (adjusting for maternal characteristics and birth characteristics) were small and not statistically significant.
In Table 2, we show estimated state variability in mortality outcomes based on multilevel logistic models. State-level variation was highest for SUDI (estimated state-level variance 0.118, 95% CI 0.068, 0.168) and congenital malformations (0.061, 95% CI 0.028, 0.095). These state-level differences were reduced substantially for all causes when we controlled for differences in maternal and birth characteristics, with particularly large reductions for SUDI, where estimated state variability dropped to 0.034 (95% CI 0.014, 0.054) when both maternal and birth characteristics were included in the model.
Table 3 shows estimated annual FTIM for our 2 hypothetical scenarios. Under the assumption that all states would achieve the survival outcomes of the 10 states with the lowest mortality outcomes overall (good FTIMR group), infant mortality would decline by an estimated 2,023 (95% CI 1,717, 2,329) deaths each year. Under the more ambitious counterfactual that all states could achieve the cause-specific mortality rates of the best-performing state in each cause-of-death category, infant mortality among full-term births would be reduced by 4,003 deaths (95% CI 2,284, 5,587) each year. Under both hypothetical scenarios, only about 10% of the potential improvements were related to perinatal conditions or other causes. More than 75% of the excess burden of mortality in both scenarios was due to congenital malformations and SUDI.
The results presented in this paper show a large gap in the survival probabilities of full-term infants born in the US compared to European countries with low under-5 mortality rates. Pooling all available data between 2010 and 2012, we found that no single US state or territory achieved the full-term survival rates currently reported in leading European countries, with children born full-term in the 10 best-performing states facing about 50% higher risks of infant mortality, and children born in states with poor FTIMR facing almost 3 times the infant mortality risk of European countries with low infant mortality rates.
Given that survival rates among preterm infants in the US were found to be very similar to those of the same European countries (as illustrated in Fig 1), clinical care during or immediately after delivery likely does not explain much of the mortality gap observed. In the sample analyzed, perinatal conditions—where healthcare quality likely matters most—accounted only for about 11% of total infant mortality among full-term births. In terms of the big picture, the high burden of FTIM in the US seemed to be mostly due to SUDI and congenital malformations, which accounted for 42.9% and 31.1% of the total infant mortality burden among full-term children, respectively, and for almost 80% of excess deaths in our counterfactual analysis. From a policy perspective, deaths due to malformations are quite different from deaths classified as SUDI. Malformations are in practice hard, if not impossible, to prevent; in most cases, the only way to “prevent” malformation-related infant mortality is to increase screening and early termination. In terms of the overall magnitude, we found malformation-specific FTIMRs of less than 3 per 10,000 live births in some states, such as Vermont and New Jersey, and rates 3 times higher in quite a few states in the Mississippi delta and surrounding states (see S2 Fig for details). Globally, WHO estimates suggest that 330,000 children die annually during the neonatal period due to congenital malformations [10,11], which corresponds to a risk of approximately 2.5 deaths per 10,000. Taking these global estimates as a benchmark suggests that children in the US face about 3 times the risk of death due to malformation in other countries. In practice, the extent to which these differences reflect differences in screening and termination policies rather than differences in medical care across states and countries is not clear; further research investigating the reach and effectiveness of early screening programs across countries and states will be needed to better understand these current gaps.
With respect to actual health improvements, the area with the most obvious and ample room for increasing the chances of child survival is SUDIs. Given that the attribution of deaths to SIDS versus “other unexplained causes” was not obvious in many cases [7,8], we mostly focused on the larger SUDI category in this paper. More than 3,000 infants died in the US each year between 2010 and 2012 due to causes that were—as the name suggests—not expected under normal conditions. This is perhaps most immediately obvious when it comes to accidental suffocation or strangulation in bed. Over 600 infants die in the US each year due to suffocation in bed; new strategies to convey optimal sleeping arrangements to parents will need to be developed and tested to prevent these deaths.
SUDI mortality in the best-performing states of the US (California and New York) was less than 6 deaths per 10,000 births; rates were more than twice as high (>12) in 12 states, including Ohio, South Dakota, and Tennessee. A large fraction of these deaths were attributed to SIDS, which has previously been estimated to cause 6.4 deaths per 10,000 births [12]. Our results suggested SIDS incidence rates as low as 1.27 and 1.32 per 10,000 full-term live births in Nevada and New Mexico and as high as 13.33 and 8.75 in Arkansas and Mississippi. Evidence from European studies suggests that a large majority of SIDS deaths could historically be attributed to prone sleeping and maternal drug consumption [13]. Through active public health programs, the incidence of SIDS was lowered by 75% in Sweden [14] and Scotland [15]; general compliance with sleeping recommendations continues to be a challenge in the US, particularly among women with low socioeconomic status [16]. Empirically, a large proportion of the state-level differences in mortality due to both SIDS and the broader SUDI category could be attributed to state-level differences in maternal age and maternal education. As shown in the more detailed regression results in S4 Table, maternal characteristics were highly predictive of these mortality outcomes. We found that compared to children born to mothers with incomplete high school education, children of highly educated mothers (those with master’s degree or doctorate) had 74% lower odds of SUDI, and that the risk of SUDI almost linearly declined with maternal age (conditional on all other factors). This suggests that mortality in this category is strongly influenced by maternal behavior and the early home environment, both of which should at least in principle be modifiable through targeted information and behavioral change interventions.
Our analysis is not without limitations. First, we have relatively little information on children’s home environments, and thus cannot directly identify what is happening at children’s homes or compare underlying risk factors. Second, it is possible that state-level estimates that we present may be biased if people move before or after birth. Empirically, for 97% of the observations, state of birth is the same as state of residence, which means that these biases should be small if they exist. Third, as mentioned above, we do not have information on termination rates at the state level, which are likely to (at least partially) explain differences in birth outcomes observed. According to the latest estimates available, approximately 700,000 legally induced abortions occurred in 2012 in the US [17], which corresponds to about 20% of the annual sample analyzed in this study. While it seems likely that infant mortality rates would be higher without these terminations, our data do not allow us to directly quantify these differences. Last, it seems likely that some of the less common causes of death (particularly in the ICD-10 R and W categories) are miscoded or coded differentially across states. To reduce this type of measurement error, we grouped all SUDIs together for most of our analyses.
More than 7,000 children born alive at full-term in the US each year die within their first year of life. The results presented in this paper suggest that a substantial proportion of these deaths are preventable, with particularly large improvements possible for SUDI.
|
10.1371/journal.pcbi.1003767 | Quantitatively Characterizing the Ligand Binding Mechanisms of Choline Binding Protein Using Markov State Model Analysis | Protein-ligand recognition plays key roles in many biological processes. One of the most fascinating questions about protein-ligand recognition is to understand its underlying mechanism, which often results from a combination of induced fit and conformational selection. In this study, we have developed a three-pronged approach of Markov State Models, Molecular Dynamics simulations, and flux analysis to determine the contribution of each model. Using this approach, we have quantified the recognition mechanism of the choline binding protein (ChoX) to be ∼90% conformational selection dominant under experimental conditions. This is achieved by recovering all the necessary parameters for the flux analysis in combination with available experimental data. Our results also suggest that ChoX has several metastable conformational states, of which an apo-closed state is dominant, consistent with previous experimental findings. Our methodology holds great potential to be widely applied to understand recognition mechanisms underlining many fundamental biological processes.
| Molecular recognition plays important roles in numerous biological processes including gene regulation, cell signaling and enzymatic activity. It has been suggested that molecular recognition employs a variety of mechanisms, ranging from induced fit to conformational selection. In many realistic systems, conformational selection and induced fit are not mutually exclusive. An analytical flux analysis has been developed to determine the contribution of each model, but it is extremely challenging to obtain the necessary kinetic parameters for this flux analysis through experimental techniques. In this work, we have developed an approach integrating Markov State Models, molecular dynamics simulations, and flux analysis to tackle this problem. Using this approach, we have quantified the recognition mechanism of the choline binding protein to be ∼90% conformational selection dominant in the experimental conditions. Our methodology provides a way to quantify the molecular recognition mechanisms that are extremely difficult to be directly accessed by experiments. This opens up numerous possibilities for in silico design to fine tune the recognition event either to increase the degree of conformational selection or induced fit, so that new properties could be created to accommodate the needs of protein engineering, drug development and beyond.
| Protein-ligand recognition plays a key role in many aspects of biological processes, such as enzyme catalysis, substrate translocation, and drug therapy [1]. Current studies indicate two prevailing models to address the recognition process: the induced fit model (where ligand binding induces conformational changes of the protein) and the conformational selection model (where the ligand selects a pre-existing conformation of the protein to bind), both of which describe extreme situations (SI Fig. S1) [2]–[10]. Recent studies, however, suggest that many realistic systems show characteristics of both mechanisms [11]–[17].
A better understanding of the role of the two models may lead to an increased utilization of protein engineering techniques – for example we may fine tune their respective contributions to allow the creation of new properties [18]. In particular, augmenting the relative contribution of the induced fit mechanism might increase the binding specificity of a protein receptor. Direct applications of such protein engineering can also lead to better chemical sensors [19].
Hammes et al. have developed an analytical model based on flux analysis to determine the contribution of conformational selection and induced fit mechanism [20]. However, difficulties arise in obtaining the thermodynamic and kinetic parameters necessary for the flux analysis from experiments. For example, it is difficult for experiments to directly examine which conformation ligands choose to bind in the conformational selection model, or observe protein conformational changes upon the ligand binding in the induced fit model. Recent progress of NMR techniques such as paramagnetic relaxation enhancement and residual dipolar coupling enable the detection of metastable conformations of the apo protein in solution, and further provide dynamic information for transitions between these conformations [21]–[24]. However, it is still difficult to apply these techniques to monitor the dynamics of the ligand binding process. In any case, as long as one obtains necessary kinetic and thermodynamic parameters, the flux through each pathway can be quantified, allowing one to assign a percentage to the involvement of induced fit or conformational selection for a particular recognition process.
Quantifying the flux is difficult by direct Molecular Dynamics (MD) simulations as well. The current timescale of MD simulations, mostly on the order of tens to hundreds of nanoseconds, is far too short to witness many biological events which occur on the order of milliseconds to seconds [25]. Only if a specific protein-ligand recognition process occurs very quickly can direct MD simulations be efficient, as our previous study on the binding mechanism of L-Lysine-, L-Arginine-, L-Ornithine-binding protein (LAOBP) demonstrated [26]. For that particular system, we used a total of 13 µs MD simulations and Markov State Models (MSMs) to examine the binding events between arginine and LAOBP that occurs at a couple of microseconds.
In this work, we propose a novel approach to qualify the flux following conformational selection and induced fit model for a particular molecular recognition process. By combining the techniques of MSMs for apo protein dynamics, direct MD simulations of the protein-ligand binding and flux analysis, we offer a systematic method for finding the necessary kinetic parameters to quantitatively measure the portion of each flux through two binding pathways. Our application of such an approach in choline binding protein (ChoX) [27], a periplasmic binding protein (PBP) [28] from the ATP-binding cassette transporter ChoVWX, demonstrates that our method can explore the free energy landscape and successfully quantify the independent contribution of two concurrent binding mechanisms – induced fit and conformational selection – in a complex realistic protein-ligand system.
As the periplasmic component of the choline import system, ChoX binds choline or acetylcholine before transferring it to the transmembrane domain of ChoVWX. Currently, X-ray crystallography has characterized several structures of this protein, all of which are in the closed or semi-closed conformations, whether there is a ligand bound or not (Fig. 1) [27], [29]. In comparison to other PBPs, this unique property of ChoX – that it can apparently stay at its closed state without the help of the ligand – made the researchers raise the hypothesis that the binding mechanism of ChoX and its ligand follows the conformational selection model [30].
MSMs are a powerful approach to automatically identify metastable states from short MD simulations and calculate the equilibrium thermodynamic and kinetic properties [31]–[49]. It divides the protein conformational space into a number of non-overlapping metastable states such that the transitions within each state are fast but transitions across different states are slow. Time is coarse grained (with the smallest unit of τ, termed as the lag time) to ensure the model is Markovian so that the probability of transitioning from state i to j only depends on i but not any previously visited states. With the help of MSMs, one can extract long time dynamics from short simulations and directly obtain many useful parameters of thermodynamics and kinetics [50]–[60], which can be further utilized in the flux analysis. For example, Noé et al. have performed the first flux analysis based on the transition path theory and MSMs to investigate the major pathways for the folding of the WW domain [55]. More recently, Noé and coworkers have also used MSMs to study the mechanisms of protein-ligand association where protein does not undergo substantial conformational changes [61].
The flux analysis proposed by Hammes et al. [20] is an useful tool to calculate and compare flux in both induced fit and conformational selection pathways, allowing one to analyze the contribution of these two limiting mechanisms in a complicated binding event. To conduct the flux-based approach, many kinetic parameters are required – namely, the binding constants and transition rates between different metastable conformational states of the system.
The combination of MSMs and MD simulation allows us to obtain such parameters, which are difficult to be directly measured from experimental assays [20]. In our study, the three-pronged approach of MSMs, MD simulation, and flux analysis was successfully applied in ChoX binding event to quantify both limiting recognition mechanisms.
We used MD simulations to investigate the free energy landscape of apo ChoX. In particular, we have generated initial twenty 100-ns simulations, ten from apo-closed (PDB ID: 2RF1) and ten from apo-semiclosed (PDB ID: 2REJ) crystal structures [29]; and one hundred 50-ns additional simulations, seeded from random conformations of the previous twenty trajectories. In total, we collected 7 µs of apo simulations and constructed a MSM using the MSMBuilder package [33] (see Methods for the details of model construction). The implied timescale plots flatten at ∼15 ns, indicating that the model is Markovian with this or longer lag time (SI Fig. S2) [33]. We thus selected 20 ns as the lag time to construct our MSM. Since the macrostate-MSM underestimates the kinetics, we computed all the quantitative properties reported in this work such as equilibrium state populations and other kinetic properties based on the 500-state microstate-MSM. To better visualize the conformational dynamics of ChoX, we have also lumped the 500 microstates into 5 metastable states, denoted as S1 to S5 with descending populations (see the Methods section for details).
The projection of the free energy landscape of apo ChoX on the domain-domain opening and twisting angle is plotted in Fig. 2. It is clear that the most populated region is near (0°, 0°), corresponding to an apo-closed crystal structure. Our simulations demonstrate that such a conformation lies in the most dominant metastable state S1 with a population of ∼47%. This result is consistent with previous X-ray crystallography research, which noted that ChoX could exist in the closed conformation without the help of a ligand [29]. In addition to this metastable state, we found other states with different degrees of opening or twisting. Such metastable conformations were not discovered by X-ray crystallography, possibly due to the very compact crystal environment and strong contacts between unit cells present in the available apo structures of ChoX (SI Fig. S3) [29].
We also studied kinetics for the transitions between these metastable states. The mean first passage times (MFPTs) were calculated for each pair of states (SI Table S1), and these timescales are on the order of microseconds.
The structural features of the five metastable states are displayed in Fig. 3. The most populated state S1 is a closed conformation very similar to those discovered crystal structures. S2 is an open-and-twisted conformation with an essential hydrogen bond between N229 and G232 at the back of the hinge region. S3 is a closed conformation with a small opening at the side of the domain-domain interface, which is large enough to allow the diffusion of the ligand to the binding site (Fig. 3b). S4 is twisted to a very large degree. S5 is another closed structure similar to S1 with a different orientation of the helix containing residues 262–275. Further investigations show that hydrogen bonds may play an important role to stabilize these metastable conformations. One example from the metastable state S2 involves N229 and G232. When N229 was mutated to alanine or G232 was mutated to bulkier tyrosine to diminish the hydrogen bonds, fast transitions (within 50-ns) were observed from S2 to S1 (SI Fig. S4) compared to the wild type (∼2.07 µs), demonstrating the critical contribution of the hydrogen bond (N229-G232) to the metastability of S2.
We performed ten 50-ns simulations for each of the five metastable states by introducing ligands to the system. To increase the chance of observing a binding event within the length scale of MD simulations, we have added 20 ligands to the system (at a concentration of ∼0.069M). In each simulation, 20 ligands were randomly placed in the simulation box away from the binding site with the minimum distance of 17 Å and the protein conformations were randomly chosen from each metastable state.
For S1 with closed protein conformations (S1+L), we discovered two out of ten simulations where the ligand recognized the target and bound to it. In order to enhance the sampling, we have performed additional twenty 50-ns MD simulations and three of them were identified with binding events.
The pathways of the ligand binding to S1 can be mainly characterized as conformational selection, and these binding simulations achieved similar conformation compared to the X-ray bound structure with a RMSD as small as 1.6 Å of protein Cα atoms which are within 8 Å to the binding site (Fig. 4a). In addition, we examined the distances of the ligand choline to four essential residues at the binding site after the ligand binds to S1, and the values are similar to those from crystal structure. These results indicate that MD simulations have the capability to predict the bound state in silico.
In addition to these ligands that can directly bind to the closed conformation S1, we also found in other simulations that the ligand can interact with the conformation from the state S3 and, at the order of tens of nanoseconds induce the conformational change to the bound conformation S1L. Recall S3 was a closed and twisted state with a side-opening cavity ready for a ligand to insert. We also demonstrated that, from a distance analysis and an overlay of S3L with the holo crystal structure, the ligand stays close to Y119 and W205 (Fig. 5b). One trajectory was discovered with a transition from S3L to S1L, which suggests the possibility that the ligand can bind ChoX through an induced fit mechanism. Since only a single transition event was observed among ten simulations of S3+L, we have performed additional twenty 50-ns MD simulations of S3L complex to enhance the sampling. Among these twenty simulations, two of them displayed the transitions from S3L to S1L (SI Fig. S5).
For the remaining metastable states S2, S4 and S5, no ligands were found to bind to the correct binding site to form the complex: S2L, S4L or S5L (SI Fig. S6). However, in order to examine whether or not these protein-ligand complexes (S2L, S4L or S5L) if exist can induce the transitions to the bound state S1L, we have modeled these complexes using the AutoDock Vina [62], and initiate ten 50-ns MD simulations from each of these docked conformations. As shown in SI Fig. S7, none of these simulations contained any transitions to the bound state (S1L). These results indicate that the direct transitions from S2L, S3L and S5L to the bound state S1L are unlikely to occur.
From the MD simulations discussed above, we can obtain a rough picture of the hybrid mechanism of conformational selection and induced fit for ChoX. However, the great challenge is to quantify the percentage of each mechanism in complicated scenarios. In order to achieve this, we have applied the flux analysis theory [20], where the flux going through each pathway is utilized to quantify binding mechanisms. At first, the conformational selection pathway can be described as (SI Fig. S1):(1)where P1 represents the closed conformation S1 of ChoX, and other metastable conformations Pi (i = 2–5) can interconvert with P1:(2)In this work, we consider the conformational selection mechanism in a general context where the ligand selects to bind a certain metastable protein conformation including the ground state (P1 in this case).
On the other hand, the induced fit pathway can be described by a two-step process (SI Fig. S1), where the ligand first binds to any metastable conformation Pi other than the closed conformation P1:(3)And the binding will further induce the conformational change to reach the bound state:(4)The flux flowing through each of the above two pathways can be derived from the flux analysis theory as the following (see Methods for the details of the derivation):(5)(6)where FCS and FIF represent the flux through conformational selection and induced fit pathway respectively. and are the kinetic rate constant for ligand binding/unbinding to state i respectively; ki1 is the rate constant for the transition from state i to state 1; is the rate constant for the transition from the complex Si·L to the bound state S1·L; and [L]f is the free ligand concentration.
In this study, we have derived important parameters from MSMs and MD simulation that are missing in the flux analysis. Specifically, ki1 can be derived from the transition probability matrix. For MD simulations starting from state S2, S4 and S5, we didn't observe any successful binding events, therefore , , are all set to be zero. For S3, there exist multiple binding and unbinding events in our ten 50-ns MD simulations. Therefore, we have obtained and by computing the fraction remaining in the unbound state S3 as a function of time followed by fitting to Eq. 19 (SI Fig. S8). can be derived from MD simulations with ligands in a similar way to (Eq. 20). At last, we need to derive the values for (i = 2–5). Since the simulations of S2L, S4L and S5L didn't show any transitions from each state to S1L (SI Fig. S7), these values (, , ) are all set to be zero. For , we can obtain its value from twenty S3L MD simulations since the ligand binding further induced the conformational changes to the bound state in a fraction of these trajectories (Eq. 24).
We can then proceed to measure the percentage contribution of conformational selection and induced fit mechanism using equations (5) and (6). At the protein concentration fixed to 1 µM, Fig. 6 shows the contribution of conformational selection to the binding pathway depending on the ligand concentration. Conformational selection is dominant for a wide range of the ligand concentration, accounting for around 90% of the binding event at the concentration of choline in the lab conditions (µM scale) [63]. Therefore, we conclude that the binding mechanism of choline to ChoX is dominated by conformational selection.
In summary, we here propose a novel method that combines MSMs, MD simulations and flux analysis to quantify the binding mechanism of conformational selection and induced fit in complex binding events. In the case study of choline binding to ChoX, we were able to derive all the necessary parameters using MD simulations and MSMs. Based on these parameters, the percentage of each limiting binding mechanism could be quantitatively calculated as a function of ligand concentration. It would be difficult, using common experiments, to obtain these necessary parameters to elucidate these mechanisms. Finally, once the mechanism is quantified, one can further apply other techniques (e.g. in silico design) to the biological system to fine tune the binding event either to increase the degree of conformational selection or induced fit, so new properties of macromolecules could be explored and created to accommodate the needs of protein engineering and beyond.
We project the free energy landscape of ChoX onto two dihedral angles: the opening and twisting angles. The opening angle is defined as the angle between the normal vectors of the two planes formed by the center of mass of the following groups of Cα atoms (SI Fig. S9a):
Plane A: residues 31–114 & 234–316; 185–194; 159–166.
Plane B: residues 118–230; 185–194; 159–166.
The twisting angle planes are:
Plane C: residues 31–114 & 234–316; 185–194; 46–55.
Plane D: residues 118–230; 185–194; 46–55 [64].
From Principal Component Analysis (PCA) of the apo MD simulations, we found strong correlations between the opening angle and the first eigenvector (R2 = 0.77), as well as between the twisting angle and the second eigenvector (R2 = 0.53) (SI Fig. S9b). In this work, the degrees of apo-closed and holo-closed crystal structures were shifted to the (0°, 0°) point in corresponding settings.
The GROMACS 4.5.4 [65] software and Amber99sb force field [66] were used for all the MD simulations. The procedure is as follows: the protein (and ligand, when present) was solvated in a dodecahedron box with 14, 450 SPC water molecules [67] and enough counter ions to neutralize the system. The system was first minimized with a steepest descent algorithm, followed by a 200 ps MD simulation with position restraints for the heavy atoms. All the simulations were performed at NVT ensemble with 300K of temperature using V-rescale thermostat [68]. The cut-offs for both VDW and short-range electrostatic interactions were set to 10 Å and long-range electrostatic interactions were treated with the Particle-Mesh Ewald method [69]. The time-step was 2 fs and the neighbour list was updated every 10 steps. Water molecules were constrained by the SETTLE algorithm [70] and all protein bonds were constrained by the LINCS algorithm [71].
Using the MSMBuilder package [33], we first applied the k-centre algorithm to cluster all the conformations into 500 microstates based on the Cα atoms of residues in proximity to the binding site (i.e. within 8 Å of the ligand as defined by the holo crystal structure, PDB ID: 2REG). The generated microstates were small, with the average RMSD values to its central conformation in each state of about 1.9 Å. The transition probability matrix (T) was obtained by counting the number of transitions observed in the MD trajectories. We then examined the transition probability matrix, and removed two disconnected microstates from our model. The implied timescales (τk) obtained from the transition probability matrix T indicates the aggregated timescales for transitions between groups of microstates.(7)where μk is an eigenvalue of the transition matrix with the lag time τ.
We have examined the implied timescale plots for this 500-microstate model to select a lag time that ensures the model to be Markovian. As shown in SI Fig. S2, the implied timescale curves plateau at around 15 ns, indicating the model is Markovian with this or longer lag times. Thus we chose a lag time of 20 ns for our final MSM. In order to better visualize the conformational dynamics of apo ChoX, we have further lumped all the microstates into 5 metastable macrostates using the Robust Perron Cluster Analysis (PCCA+) algorithm [72].
The 500-state microstate-MSM was used to compute all the quantitative properties we report in this work. To obtain the populations of metastable states (P1, … P5) from the 500-state microstate-MSM, we simply sum over the equilibrium populations of all the microstates that belong to a certain metastable state: . To compute the MFPT between a pair of metastable state i and j from the microstate-MSM, we first set MFPTs of all the microstates that belong to the destination metastable state j to be zero: . We then computed MFPTs starting from any of the microstate that belong to the metastable state i: . Finally, we obtained the MFPTs from i to j by performing a weighted average over all the microstates that belong to i: .
The MFPT, determined by the following formula, calculates the average transition times between a pair of states [73].(8)where Pij is the transition probability from state i to state j, τ is the lag time of the transition probability matrix T, and MFPTjf is the mean first passage time of the state j to final state f. The boundary condition is MFPTff = 0.
In order to simulate the process of choline binding to the ChoX protein, we need to obtain the force field parameters for choline. We followed the same procedure as we previously published [74] to derive both bonded and non-bonded force field parameters of choline. Specifically, we have obtained the stretching, bending and torsion parameters by fitting against the Quantum Mechanics (QM) calculations performed using the Density Functional Theory with B3LYP/6-31G* in the Gaussian software [75]. Similar with previous studies [76], we have employed the Restrained ElectroStatic Potential (RESP) method to derive the partial charges from the QM calculations with HF/6-31G*. We have listed all the force field parameters in the format that is compatible with the GROMACS 4.5 software package in SI Text S1, .
The two limiting ligand-binding mechanisms: conformational selection and induced fit can be described by Eq. (1)–(4). Following the flux analysis theory developed by Hammes et al. [20], we can then derive the fractional flux passing through each pathway. We note that this flux analysis can be considered as a special case of transition path theory and yields consistent path fluxes for serial and parallel pathways [55], [77]. In particular, if one pathway is consisted of parallel reaction paths, its flux can be written as:(9)If one pathway contains serial segments, the flux is:(10)Now, let's consider the conformational selection pathway, which involves two steps. First, different protein metastable conformations (Si, i = 2–5) can all interconvert to the closed state (S1) at rates ki1 (Eq. 2). Therefore, the flux for this segment is: , where [Pi] is the concentration of the state Si. Next, the ligand can selectively bind to the closed protein conformation S1, to reach the bound state. For this part, the flux is: (Eq. 1). Therefore, the flux of conformational selection FCS can be derived as:(11)The detailed expression of FCS can be found in Eq. 5.
Following the similar procedure [20], we can also derive the flux for the induced fit pathway. In particular, there exist independent parallel pathways to reach the bound state, where the ligand can first bind to a certain metastable conformational state (Si, i = 2–5), and further induced the conformational change to the bound state. The flux of each of these pathways (Fi) can be written as:(12)where and denote to the rate for the ligand binding to state Si (i = 2–5) and the transition rate from the complex Si·L to the bound state S1·L respectively; when [L] exceeds. When [L] and [P] are comparable, [L]f is given by Eq. 13:(13)The overall flux for the induced fit pathway (FIF) can then be written as:(14)The fractional flux for conformational selection pathway is:(15)In order to obtain FCS%, we need to derive a series of parameters: (i = 2–5), (i = 1–5), (i = 2–5). are the transition rate constants from state i to state 1, which can simply be derived from MFPT [58]:(16)where MFPTi1 is the mean first passage time of state i to final state 1. The uncertainties of this set of rates are obtained from bootstrapping the MD dataset (containing N = 138 trajectories) with replacement for N times.
For , we use our MD simulations with ligand to derive their values. Based on Eq. (1) and (3), we can write the rate equation as:(17)In our simulations, the initial ligand concentration [L]0 is twenty times larger than the initial protein concentration [P]0: , therefore, . The forward reaction, which is only dependent on kon, can then be written as:(18)We can solve Eq. 18 to obtain [Pi]:(19)
We can then examine our ligand MD simulations initiated from different protein metastable conformations to obtain and . For MD simulations starting from state S2, S4 and S5, we didn't observe any binding events (for distances between the center of mass of ligand and center of mass of four critical residues, W43, W90, Y119, W205 [29], in the binding site to be less than 12 Å), therefore , , all equal zero (however we estimated their upper limits in Table 1). For S3, there exist multiple binding and unbinding events in our ten 50-ns MD simulations. We have thus computed the fraction remaining in the unbound state () as a function of time, and further fit Eq. 19 to it in order to obtain and (see SI Fig. S8 for the parameter fitting). We can then obtain . Thus, [P3·L] can be calculated by . For S1, we have observed multiple binding events; however none of the unbinding events from our MD simulations, due to the high stability of the bound state. We can then simplify Eq. 18 by only considering the forward reaction and can be written as:(20)We then obtain .
To further examine the robustness of the definition of the successful binding events, we have compared it with a more specific definition: the heavy atoms of the ligand form contact with atoms belonging to at least three critical residues (among W43, W90, Y119, W205 [29]) in the binding pocket. These two sets of rates only differ slightly (SI Table S2). Furthermore, we have compared the fraction of conformational selection computed based on two different definitions of ligand binding. As shown in SI Fig. S10, the results from two different definitions are also consistent.
In order to examine whether or not the protein undergoes any major conformational changes while collecting data for estimating the binding rates (e.g. , , and ), we have projected protein conformations in each binding MD simulation onto a pair of reaction coordinates (opening and twisting angles). As shown in SI Fig. S11, the protein remains in its initial metastable state during the whole course of all the binding MD simulations. These results confirm that our estimations of binding rates are clean.
In order to obtain , which is the dissociation constant for the ligand directly binding to S1, we have constructed a thermodynamic cycle to compute its value (SI Fig. S12):(21)where the free energy is directly related to : .
can be computed from the disassociation constant for the ligand binding by , where has been obtained experimentally as 2.7 µM [63]. is the free energy difference associated with the conformational transitions from different metastable states to the closed state S1:(22)where is the free energy difference between Si and S1, and fi denotes to the equilibrium population of Si:(23)From Eq. 21–23, we can compute the value of as 4.53 µM.
At last, we need to derive the values for (i = 2–5). As discussed before, there are no binding/unbinding events for S2, S4 and S5, and these values (, , ) are all set to be zero. For , we can obtain the value from the MD simulations initiated from the S3 state with ligand.(24)Eq. 24, helps us to calculate the forward transition rate .
For the ligand binding/unbinding (, and ) and transition from S3L to S1L (). We have performed the bootstrapping analysis to obtain the error bars. In particular, we bootstrapped the MD dataset (containing N MD simulation trajectories with N = 30, 9 and 20 for , /, and respectively) with replacement for N times.
For rates that are estimated to be zero with no observed transitions (e.g. , ), we have estimated the upper limit of these rates by assuming one binding/transition event occurs during our accumulated 500-ns simulation time. We also show that applying these upper limits of the rates to our flux analysis does not change qualitatively the conclusion of this paper: the fraction of the conformation selection slightly decreases from 93% to 82% at experimental concentrations (∼1 µM) as shown in SI Fig. S13.
We have listed all the necessary rate constants with uncertainties/upper limits in Table 1.
|
10.1371/journal.pntd.0007169 | The diversity, evolution and ecology of Salmonella in venomous snakes | Reptile-associated Salmonella bacteria are a major, but often neglected cause of both gastrointestinal and bloodstream infection in humans globally. The diversity of Salmonella enterica has not yet been determined in venomous snakes, however other ectothermic animals have been reported to carry a broad range of Salmonella bacteria. We investigated the prevalence and diversity of Salmonella in a collection of venomous snakes and non-venomous reptiles.
We used a combination of selective enrichment techniques to establish a unique dataset of reptilian isolates to study Salmonella enterica species-level evolution and ecology and used whole-genome sequencing to investigate the relatedness of phylogenetic groups. We observed that 91% of venomous snakes carried Salmonella, and found that a diverse range of serovars (n = 58) were carried by reptiles. The Salmonella serovars belonged to four of the six Salmonella enterica subspecies: diarizonae, enterica, houtanae and salamae. Subspecies enterica isolates were distributed among two distinct phylogenetic clusters, previously described as clade A (52%) and clade B (48%). We identified metabolic differences between S. diarizonae, S. enterica clade A and clade B involving growth on lactose, tartaric acid, dulcitol, myo-inositol and allantoin.
We present the first whole genome-based comparative study of the Salmonella bacteria that colonise venomous and non-venomous reptiles and shed new light on Salmonella evolution. Venomous snakes examined in this study carried a broad range of Salmonella, including serovars which have been associated with disease in humans such as S. Enteritidis. The findings raise the possibility that venomous snakes could be a reservoir for Salmonella serovars associated with human salmonellosis.
| Salmonella enterica is a remarkable bacterial species that causes Neglected Tropical Diseases globally. The burden of disease is greatest in some of the most poverty-afflicted regions of Africa, where salmonellosis frequently causes bloodstream infection with fatal consequences. The bacteria have the ability to colonise the gastrointestinal tract of a wide range of animals including reptiles. Direct or indirect contact between reptiles and humans can cause salmonellosis. In this study, we determined the prevalence and diversity of Salmonella in a collection of African venomous snakes for the first time. We showed that the majority of venomous snakes (91%) in our study carry Salmonella, and used bacterial genomics to assign two thirds of isolates to the S. enterica subspecies enterica which is associated with human salmonellosis. We identified two evolutionary groups of S. enterica subspecies enterica that display distinct metabolic profiles with infection-relevant carbon sources. Our findings could have a broad significance because venomous snakes can move freely around human dwellings in tropical regions of the world such as Africa, and could potentially shed contaminated faecal matter onto surfaces and into water supplies.
| Salmonella is a clinically relevant bacterial pathogen that poses a significant burden upon public health worldwide [1–4]. Two groups of Salmonella serovars have clinical relevance with distinct host-specificity and disease manifestations. Typhoidal Salmonella is restricted to human hosts and presents as a systemic infection, resulting in an estimated 223,000 fatalities per annum [5]. In contrast, nontyphoidal Salmonella typically manifests as a self-limiting gastrointestinal disease in otherwise healthy individuals around the world, causing an annual global disease burden of 93.8 million cases and 155,000 deaths [3]. Over the past two decades, an invasive form of nontyphoidal Salmonella (iNTS) has emerged as the most prevalent bacterial species to be isolated from the bloodstream of patients in sub-Saharan Africa [6]. Over 3.4 million cases and 680,000 deaths are estimated to occur worldwide each year as a result of iNTS [4].
The Salmonella genus contains two species; S. bongori and S. enterica. S. enterica is further divided into six subspecies; enterica (I), salamae (II), arizonae (IIIa), diarizonae (IIIb), houtanae (IV) and indica (VI) [7]. The subspecies are classified into approximately 2 600 serovars which are ecologically, phenotypically and genetically diverse [8]. Serovars which belong to S. enterica subspecies enterica cluster phylogenetically into two predominant clades (A and B) [9–11]. Here, we use the term Salmonella to refer to the S. enterica species and the S. enterica designation to refer to the S. enterica subspecies enterica alone, unless stated otherwise.
Biochemical properties such as carbon utilisation and anaerobic metabolism are often serovar-specific [12]. The ability of Salmonella to grow in a wide range of conditions reflects the adaptation of the bacteria to survive in the environment or in different hosts, as demonstrated by a recent study focused on genome-scale metabolic models for 410 Salmonella isolates spanning 64 serovars in 530 different growth conditions [13].
At the genus level, Salmonella has a broad host-range whilst individual serovars differ in host-specificity [14]. The majority of Salmonella infections in humans (99%) are caused by a small number of serovars belonging to the S. enterica subspecies [15]. Serovars which belong to non-enterica subspecies are associated with carriage in ectothermic animals such as reptiles and amphibians, but are rarely found in humans [14,16–18]. Carriage rates of non-enterica serovars in reptiles can be high. A study focused on snakes in a pet shop found that 81% of animals were carrying S. diarizonae [18]. Previous studies demonstrating the diverse range of Salmonella subspecies that colonise various reptilian species in different countries are summarised in Table 1.
Reptiles represent a significant reservoir for serovars of Salmonella that are associated with human disease. Over 60% of captive-bred reptiles between 1995 and 2006 in Denmark were reported to carry S. enterica subspecies enterica serovars [24]. About 6% of human salmonellosis cases were contracted from reptiles in the USA [25], and in South West England, 27.4% of Salmonella cases in children under five years old were linked to reptile exposure [26]. The latter study demonstrated that reptile-derived salmonellosis was more likely to cause bloodstream infection in humans than non-reptile-derived Salmonella [26]. Reptile-associated Salmonella is therefore considered to be a global threat to public health [27].
The majority of reptile-associated salmonellosis cases reported in humans are caused by Salmonella from non-venomous reptiles [27], probably because these animals are frequently kept as pets. Therefore, non-venomous reptiles have been the focus of numerous studies whilst the prevalence and diversity of Salmonella in venomous snakes has remained unknown. The recent inclusion of snakebite as a neglected tropical disease demonstrates that these reptiles frequently interact with humans in tropical and sub-tropical countries. The proximity of venomous snakes to humans may lead to contaminated faecal matter being shed on the surfaces and in water sources used for human homes and to irrigate salad crops [28–30]. Research to improve snakebite treatment at the Liverpool School of Tropical Medicine (LSTM) has resulted in the creation of the most extensive collection of venomous snakes in the UK (195). The LSTM herpetarium houses venomous snakes from a diverse range of species and geographical origins, representing an ideal source of samples to assess Salmonella in this under-studied group of reptiles.
The aims of this study were three-fold. Firstly, to determine the period prevalence of Salmonella in a collection of captive venomous snakes and investigate whether this group of reptiles could act as reservoirs for human salmonellosis. Secondly, to assess the serological and phylogenetic diversity of Salmonella amongst reptiles. Thirdly, to use the diversity of reptile-associated Salmonella to determine clade-specific differences that could reflect adaptation to survival in the environment or to different hosts. Here, we present the first whole genome-based comparative study of the Salmonella bacteria that colonise venomous and non-venomous reptiles.
The Salmonella isolates were derived from faecal samples from two collections of reptiles. One hundred and six faecal samples were collected from venomous snakes at LSTM from May 2015 to January 2017, with an emphasis on snakes originating from Africa (S1 Table), and investigated for the presence of Salmonella. All venomous snakes were housed in individual enclosures and fed with frozen mice. Sixty-nine of the samples (71%) were sourced from wild-caught snakes originating from: Togo, Nigeria, Cameroon, Egypt, Tanzania, Kenya, South Africa, and Uganda. A further 28 Salmonella isolates (29%) came from venomous snakes bred in captivity. The LSTM herpetarium is a UK Home Office licensed and inspected animal holding facility. A second collection of 27 Salmonella isolates from non-venomous reptiles and 1 Salmonella isolate from a venomous reptile were sourced from the veterinary diagnostics laboratory based at the University of Liverpool’s Leahurst campus (reptilian species described in S1 Table). These isolates were collected from June 2011 to July 2016 from specimens submitted as part of Salmonella surveillance for import/export, in addition to veterinary faecal samples and tissues from post mortem investigations. The provenance of the isolates is described in S1 Table. The majority of the non-venomous reptiles were sourced from a zoological collection, however two animals were privately owned and three were sourced from the Royal Society for the Prevention of Cruelty to Animals (RSPCA). The LSTM isolates are henceforth referred to as venomous snake isolates and the Leahurst isolates are referred to as non-venomous reptile isolates unless otherwise stated.
All media were prepared and used in accordance with the manufacturer’s guidelines unless otherwise stated. Salmonella was isolated using a modified version of the protocol described in the national Standard Operating Procedure for detection of Salmonella issued by Public Health England [31].
Faecal droppings were collected from reptiles and stored in 15 mL plastic centrifuge tubes at 4°C. Two different methods were used for the enrichment of Salmonella from faecal samples due to reagent availability at the time of isolation. S1 Table provides information on isolate specific methods. In enrichment method 1, faecal samples were added to 10 mL of buffered peptone water (Fluka Analytical, UK, 08105-500G-F) and incubated overnight at 37°C with shaking at 220 rpm. Following overnight incubation, 100 μL of the faeces mixture was added to 10 mL of Selenite Broth (19 g/L Selenite Broth Base, Merck, UK, 70153-500G and 4 g/L Sodium Hydrogen Selenite, Merck 1.06340-50G) and incubated overnight at 37°C with shaking at 220 rpm. In enrichment method 2, faecal samples were added to 10 mL of Buffered Peptone Water (Fluka Analytical, 08105-500G-F) supplemented with 10 μg/mL Novobiocin (Merck, N1628), and incubated overnight at 37°C with shaking at 220 rpm. Following overnight incubation, 100 μL of the faeces mixture was added to 10 mL of Rappaport-Vassilliadis Medium (Lab M, UK, LAB086) and incubated for 24 hours at 42°C with shaking at 220 rpm.
Following enrichment, 10 μL of overnight broth was spread onto Xylose Lysine Deoxycholate (XLD) (Oxoid, UK, CM0469) agar plates which were incubated overnight at 37°C. Putative Salmonella colonies were selected by black appearance on XLD plates and confirmed by pink and white colony formation on Brilliant Green Agar (Merck, 70134-500G) supplemented with 0.35 g/L Mandelic Acid (Merck, M2101) and 1 g/L Sodium Sulfacetamide (Merck, S8647).
To identify S. enterica species, colony PCR of the Salmonella specific ttr locus, which is required for tetrathionate respiration [32], was performed. PCR reagents included MyTaq Red Mix 1x (Bioline, UK, BIO-25043), ttr-4 reverse primer (5'-AGCTCAGACCAAAAGTGACCATC-3') and ttr-6 forward primer (5'-CTCACCAGGAGATTACAACATGG-3') on colonies suspected to be Salmonella. PCR reaction conditions were as follows: 95°C 2 min, 35 x (95°C 15 s, 60°C 30 s, 72°C 10 s), 72°C 5 min. PCR products were visualised using agarose gel (3.5%) (Bioline, BIO-41025) electrophoresis in TAE buffer. Midori Green DNA stain (3 μL/100 mL) (Nippon Genetics, Germany, MG 04) was used to visualise DNA bands under UV light. Throughout the isolation procedure, S. enterica serovar Typhimurium (S. Typhimurium) strain LT2 [33,34] was used as a positive control, and Escherichia coli MG1655 [35] was used as a negative control (S1 Table).
All non-venomous reptile isolates, one venomous reptile isolate and 87 of 97 venomous snake isolates were sent for whole-genome sequencing. Isolates were sent to either MicrobesNG, UK or the Earlham Institute, UK for whole-genome sequencing on the Illumina HiSeq platform (Illumina, California, USA). Isolates which were sequenced by MicrobesNG were prepared for sequencing in accordance with the company’s preparation protocol for single colony-derived bacterial cultures (http://www.microbesng.uk).
Isolates which were sequenced by the Earlham Institute were prepared by inoculating a single colony of Salmonella into a FluidX 2D Sequencing Tube (FluidX Ltd, UK) containing 100 μL of Lysogeny Broth (LB, Lennox) and incubating overnight at 37°C, with shaking at 220 rpm. LB was made using 10 g/L Bacto Tryptone (BD Biosciences, UK, 211705), 5 g/L Bacto Yeast Extract (BD, 212750) and 5 g/L Sodium Chloride (Merck, S3014-1kg). Following overnight growth, the FluidX 2D Tubes were placed in a 95°C oven for 20 minutes to heat-kill the isolates.
DNA extractions and Illumina library preparations were conducted using automated robots at MicrobesNG or the Earlham Institute. At the Earlham Institute, the Illumina Nextera XT DNA Library Prep Kit (Illumina, FC-131-1096) was used for library preparation. High throughput sequencing was performed using an Illumina HiSeq 4000 sequencing machine to generate 150 bp paired-end reads. Sequencing was multiplexed with 768 unique barcode combinations per sequencing lane. The insert size was approximately 180 bp, and the median depth of coverage was 30x.
At MicrobesNG (https://microbesng.uk), genomic DNA libraries were prepared using the Nextera XT Library Prep Kit (Illumina, FC-131-1096) with two nanograms of DNA used as input and double the elongation time that was described by the manufacturer. Libraries were sequenced on the Illumina HiSeq 2500 using a 250 bp protocol.
The Salmonella In Silico Typing Resource (SISTR) v1.0.2 was used for serovar prediction [36]. Enterobase [37] was used to assign a Multi Locus Sequence Type (MLST) to each isolate, based on sequence conservation of seven housekeeping genes [37]. Where available, reference isolates representing previously sequenced Salmonella isolates for all subspecies and serovars identified were included in the analysis. Reference sequence assemblies were downloaded from the National Center for Biotechnology Information (NCBI). Accession numbers are available in S2 Table.
Fastqc v0.11.5 (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/) and multiqc v1.0 (http://multiqc.info) were used to assess read quality. Kraken v0.10.5-beta [38] was run to ensure reads were free from contamination using the MiniKraken 8gb database and a Salmonella abundance cut-off of 70%. Trimmomatic v0.36 [39] was then used on the paired end reads to trim low-quality regions using a sliding widow of 4:200. ILLUMINACLIP was used to remove adapter sequences.
Genomes were assembled using SPADES v3.90 [40]. QUAST v4.6.3 [41] was used to assess the quality of assemblies, the results of which can be found in S3 Table. Assemblies which comprised of greater than 500 contiguous sequences were deemed too fragmented for downstream analysis. All assemblies which passed QC were annotated using Prokka v1.12 [42]. Roary v3.11.0 [43] was used to generate a core genome alignment. SNP-sites v2.3.3 [44] was used to extract SNPs. A maximum likelihood tree was built from the core genome SNP alignment of all isolates using RAxML-NG v0.4.1 BETA [45] with the general time reversible GTR model and gamma distribution for site specific variation and 100 bootstrap replicates to assess support. The tree was rooted using the Salmonella species S. bongori. Interactive Tree Of Life v4.2 [46] was used for tree visualisation. We confirmed that there was no bias in phylogenetic signal between the two different sequencing platforms used by assessing clustering patterns within the phylogenies. S1 Table contains details of the sequencing facility. Monophyletic clustering of isolates was used to assign subspecies to newly sequenced Salmonella from venomous and non-venomous reptilian hosts. The level of association between venom status and phylogenetic clade was determined using odds ratios and χ2 statistics using the OpenEpi website (http://www.openepi.com).
Genes involved in the utilisation of each carbon source were identified using KEGG [47] and relevant literature [9,32,48–59] (see S4 Table). Genes involved in the uptake of carbon sources were prioritised. Sequences were downloaded using the online tool SalComMac [60], which allows the download of fasta sequences of the genes in S. Typhimurium strain 4/74. In the case of lac genes, the sequences were taken from the E. coli reference sequence MG1655. The sequences can be found in S1 Text.
The software tool MEGABLAST v2.2.17 [61] was used to perform a BLAST search of genes in the reptile-derived genomes against a custom-made database of genes diagnostic of Salmonella Pathogenicity Islands and genes involved in carbon utilisation. To confirm all MEGABLAST results, the short reads were mapped against each gene using BWA v0.7.10 [62] and SAMtools v0.1.19 [63]. The resulting bam files were manually assessed for gene presence and absence using Integrative Genomics Viewer v2.4.15 [64]. The results were plotted against the maximum likelihood phylogeny using Interactive Tree Of Life v4.2 [46].
Differential carbon source utilisation of 39 reptile-derived Salmonella isolates from S. diarizonae, S. enterica clade A and S. enterica clade B was assessed. Filter-sterilised carbon sugar solutions were added into M9 (Merck, M6030-1kg) agar at concentrations detailed in S5 Table. Isolated colonies were transferred from LB agar plates onto M9 carbon source plates using a sterile 48-pronged replica plate stamp and incubated at 37°C under aerobic conditions. An LB control plate was used to validate successful bacterial transfer and all experiments were performed in duplicate. If no growth was seen under aerobic conditions for a particular carbon source, the procedure was repeated under anaerobic conditions (approx. 0.35% oxygen) with 20 mM Trimethylamine N-oxide dihydrate (TMAO) (Merck, 92277) as a terminal electron acceptor. Anaerobic conditions were achieved by incubating plates in an anaerobic jar with 3x AnaeroGen 2.5 L sachets (Thermo Scientific, UK, AN0025A) to generate anaerobic gas. Oxygen levels were measured using SP-PSt7-NAU Sensor Spots and the Microx 4 oxygen detection system (PreSens, Regensburg, Germany). Salmonella growth was determined at 18, 90 and 162 hours in aerobic growth conditions and at 162 hours in anaerobic growth conditions. A sub-set of growth positive isolates were assessed for single colony formation to validate the results of the replica plating.
Antimicrobial susceptibility was determined using a modified version of the European Committee on Antimicrobial Susceptibility Testing (EUCAST) disk diffusion method [65] using Mueller Hinton (Lab M, LAB039) agar plates and a DISKMASTER dispenser (Mast Group, UK, MDD64). Inhibition zone diameters were measured and compared to EUCAST zone diameter breakpoints for Enterobacteriaceae [66]. Isolates were first tested with six commonly used antibiotics (Ampicillin 10 μg, Chloramphenicol 30 μg, Nalidixic Acid 30 μg, Tetracycline 30 μg, Ceftriaxone 30 μg, and Trimethoprim/Sulfamethoxazole 25 μg) and were then tested with five additional antibiotics (Meropenem 10 μg, Gentamicin 10 μg, Amoxicillin/Clavulanic Acid 30 μg, Azithromycin 15 μg, and Ciprofloxacin 5 μg) if any resistance was seen to the primary antibiotics (all disks from Mast Group). If resistance was observed phenotypically, then the presence of antimicrobial resistance genes were investigated using the Resfinder software tool [67]. Antimicrobial resistance was defined as resistance to one antimicrobial to which isolates would normally be susceptible [68]. Multidrug resistance was defined as an isolate which showed resistance to three or more antimicrobials to which it would normally be susceptible [68].
Salmonella carriage is well documented amongst reptiles (Table 1), however, to our knowledge no published study reports the incidence of Salmonella in venomous snakes. The period prevalence of Salmonella was assessed in a collection of 106 venomous snakes housed at the LSTM venom unit between May 2015 and January 2017. A remarkably high proportion (91%; 97/106) of the faecal samples contained Salmonella (S1 Table), which should be seen in the context of the significant carriage rate of Salmonella by other non-venomous reptiles described in the literature [27]. Variable rates of Salmonella carriage have been observed in collections of reptiles (Table 1), and the large proportion of venomous snakes carrying Salmonella in our study sits at the higher end of the reported spectrum. Our findings pose important public health considerations for individuals who work with venomous snakes housed in captivity, which may previously have been overlooked.
To assess diversity, 87 venomous snake-derived Salmonella isolates, 27 non-venomous reptile-derived Salmonella isolates and one venomous reptile-derived Salmonella isolate were whole-genome sequenced. In silico serotyping revealed 58 different Salmonella serovars (Fig 1). A wide range of serovars was found in each of the venomous and non-venomous snake collections. Given that many of the serovars identified have a very broad host specificity, we suggest that the presence of particular serovars is not linked to the venom status of the reptile.
Similar levels of Salmonella carriage were seen in wild-caught and captive-bred reptiles. It is likely that the difference in serovar distribution reflects the sourcing of the reptiles from two independent housing facilities, and represents a limitation of our study. Nevertheless, an unprecedented level of Salmonella diversity was identified amongst both venomous and non-venomous reptiles. Following whole-genome sequencing, multi-locus sequence typing (MLST) was used for sub-serovar genetic characterisation of Salmonella [37]. In all cases, isolates falling within the same serovar had identical sequence types (S1 Table), reflecting the intra-serovar homogeneity of the Salmonella isolated in this study.
The most common serovar to be identified amongst the venomous snake isolates was S. Souhanina (n = 12). All S. Souhanina isolates clustered locally on the phylogeny (Fig 2) falling within a 5 SNP cluster characteristic of a clonal expansion event [70]. Four of these isolates were found in captive-bred reptiles, whilst 11 isolates came from venomous snakes which originated in Cameroon, Uganda, Tanzania, Nigeria, Togo and Egypt. The close phylogenetic relationship between the S. Souhanina isolates that belong to the same MLST type (ST488) from imported animals with a range of origins and from captive animals suggests that local Salmonella transmission may be occurring. Local transmission of near-identical salmonellae could occur between snakes or as a result of a single contaminated food source such as frozen mice [71,72].
Although our data suggest that S. Souhanina was locally transmitted within the herpetarium, a single source of Salmonella would not explain the wide variety of serovars and MLST types carried by this collection of venomous snakes. Significant Salmonella diversity was reported in a study that involved 166 faecal samples from wild-caught reptiles in Spain, identifying 27 unique serovars [73]. An assessment of Salmonella diversity in wildlife in New South Wales, Australia identified 20 unique serovars amongst 60 wild-reptiles [74]. We speculate that the majority of the diversity of Salmonella identified here originated from wild-caught reptiles and reflect their varied habitats.
Underpinning our strategy for sampling Salmonella was the assumption that venomous snakes can carry and shed Salmonella for long periods of time. The longitudinal shedding of Salmonella has been reported in 12 captive non-venomous snakes from 7 different species. Over 10 consecutive weeks, 58% of the snakes shed Salmonella intermittently [75]. Chronic Salmonella carriage has been reported in many other animals, including laying hens which shed the bacteria continually for up to 10 weeks [76]. To assess the continuity of Salmonella shedding from venomous snakes in this study we collected three faecal samples from a Western Green Mamba from Togo over a three-month period between 31st October 2016 and 31st January 2017. All three faecal samples contained Salmonella which belonged to sequence type ST488, showing that individual snakes have the capacity to shed the same sequence type of Salmonella over a 90-day period in this study.
We propose that the majority of the reptile-derived Salmonella described in this study were acquired by reptiles prior to captivity, whilst some isolates were transmitted locally within the herpetarium.
Because the majority of venomous snakes examined in this study were of African origin or belonged to a species of snake native to the African continent (Fig 2), we compared the Salmonella serovars isolated from all reptiles in this study with those most frequently associated with human disease in Africa. The Salmonella serovar distribution has been reported by the WHO global foodborne infections network data bank based on data from quality assured laboratories in Cameroon, Senegal and Tunisia [69] (Fig 1). Eleven snake-derived isolates belonged to serovars commonly pathogenic in humans. This finding prompted us to determine the proportion of all venomous snakes and non-venomous reptiles that carried antimicrobial resistant Salmonella (Table 2). In Salmonella collected from venomous snakes, 4.1% of isolates (4/97) were resistant to at least one antimicrobial and two isolates were multidrug resistant (Table 2). Three resistant isolates from venomous snakes belonged to the serovar Enteritidis and were closely related to the global S. Enteritidis epidemic clade which causes human disease in Africa [77]. These findings demonstrate that venomous snakes are capable of carrying and shedding Salmonella that have the potential to cause disease in humans.
Here we have shown that venomous snakes can shed Salmonella. The vast diversity of Salmonella has long been acknowledged in the literature [37]. To study the diversity of reptile-associated Salmonella from an evolutionary perspective, we obtained 87 high quality whole-genome sequences for a phylogenetic comparison that involved 24 contextual Salmonella genomes (methods). The 87 genomes represented 60 isolates from venomous snakes, 26 Salmonella isolates from non-venomous reptiles and 1 Salmonella isolate from a venomous reptile. Following a comprehensive comparative genomic analysis, we identified a total of 405,231 core genome SNPs that differentiated the 87 isolates, and were used to infer a maximum likelihood phylogeny (Fig 2). SNPs are a valuable marker of genetic diversity [70], and the identification of hundreds of thousands of core-genome SNPs reflects a high level of genetic diversity among the reptile associated Salmonella isolates. The collection of reptile-derived Salmonella represented most of the known diversity of the Salmonella genus [7], spanning four of the six Salmonella enterica subspecies: diarizonae, enterica, houtanae and salamae. Reptile-derived S. enterica subspecies enterica isolates were approximately equally distributed among two distinct phylogenetic clusters, known as clade A (58%) and clade B (48%) [9–11] (Fig 2). No significant association was found between venom status and phylogenetic group (OR = 1.1, CI = 0.3–3.0, χ2 = 0.02, P = 0.4).
The unique collection of diverse Salmonella isolates was used to determine the phenotypic and genotypic conservation of infection-relevant properties and genomic elements. Whilst the reptile-associated Salmonella belonged to five evolutionary groups, the majority of isolates were classified as S. diarizonae or S. enterica. The clustering of S. enterica into two clades (A and B) has previously been inferred phylogenetically based on the alignment of 92 core loci [9,10]. The biological significance of S. enterica clade A and clade B has been established as the two clades differ in host specificity, virulence-associated genes and metabolic properties such as carbon utilisation [11]. The genome sequences were used to expand upon pre-existing knowledge and determine phenotypic and genotypic conservation of metabolic and virulence factors across S. diarizonae and S. enterica (clades A and B).
Although the majority of Salmonella serovars of public health significance belong to clade A, certain clade B serovars such as Salmonella Panama have been associated with invasive disease [78,79]. The clade B S. enterica generally carry a combination of two Salmonella genomic islands. The Salmonella Pathogenicity Island-18 encodes an intracellularly expressed pore forming hemolysin hlyE and the cytolethal distending toxin islet which includes the gene cdtB [2,9]. It has been suggested that the two islands are associated with invasive disease, as previously they had only been identified in S. enterica serovar Typhi and Paratyphi A, which cause bloodstream infections [2,9]. The combination of hlyE and cdtB genes were present in all S. diarizonae and S. enterica clade B isolates in this study, but absent from all but one S. enterica clade A isolate (14L-2174). We propose that the significant proportion of reptiles which carried S. enterica clade B could partially explain the increased likelihood of reptile-associated salmonellosis involving invasive disease, compared to non-reptile-acquired salmonellosis [26].
To assess metabolic differences that distinguish S. enterica clade A, S. enterica clade B and S. diarizonae, we phenotypically screened 39 reptile isolates for the ability to catabolise a number of infection-relevant carbon sources [48,77,80,81] (S4 Table and S5 Table). A summary of the results for phenotypic carbon utilisation and the presence of genes associated with the cognate metabolic pathway is shown in Fig 3.
In general, the genotype accurately reflected phenotype in terms of carbon source utilisation; however, this was not always the case (Fig 3). Discrepancies between phenotypic growth and genotype suggests that mechanisms of Salmonella metabolism remain to be elucidated. For example, S. diarizonae isolate LSS-18 grew well on myo-inositol as a sole carbon source (Fig 3) but showed zero percent homology with any of the iol genes from the well-characterised Salmonella strain 4/74. The ability to utilise lactose was a property of most S. diarizonae isolates, consistent with previous reports that 85% of S. diarizonae are Lac+ [82]. It is estimated that less than 1% of all Salmonella ferment lactose due to the loss of the lac operon from the S. enterica subspecies [83]. It was interesting to discover that one non-venomous snake isolate (13L-2837) which belongs to S. enterica clade B was capable of utilising lactose as a sole carbon source. Isolate 13L-2337 belongs to the serovar S. Johannesburg and to our knowledge this is the first published occurrence of a Lac+ S. Johannesburg isolate. The 13L-2837 pan-genome had zero percent homology to the lac genes from reference strain E. coli MG1655 (sequence in S1 Text) (results in Fig 3), suggesting an alternative method for lactose utilisation. The 13L-2837 S. Johannesburg isolate also lacked the ability to grow on dulcitol, despite possessing all of the relevant gat genes, raising the possibility of an inverse relationship between the ability of Salmonella to utilise dulcitol and lactose as a sole carbon source. These findings require further investigation which is beyond the scope of the current study.
The majority of S. enterica clade A and clade B isolates utilised dulcitol, whereas dulcitol was rarely used as a sole carbon source by S. diarizonae. These findings are consistent with a study of Salmonella derived from Australian sleepy lizards, which demonstrated that dulcitol utilisation was observed in almost all S. enterica and S. salamae isolates but only 10% of S. diarizonae isolates [84]. Over 50% of the S. enterica clade A isolates lacked the gatY gene but grew well on dulcitol as a sole carbon source, suggesting that GatY is not required for dulcitol catabolism. A variety of repertoires of dulcitol catabolic genes have been described across Salmonella, with individual serovars carrying one of two gat gene clusters [59]. Both of these clusters carry the gatY gene. Our findings may indicate that a third gat gene cluster is carried by some Salmonella serovars.
In the majority of cases, allantoin was only utilised as a sole carbon source by S. enterica clade A isolates, consistent with a previous report that described an association of clade A with the allantoin catabolism island [9]. The majority of clade B isolates lacked the allantoin catabolism island and thus were unable to utilise allantoin as a sole carbon source. However, we identified one clade B isolate as an exception, isolate 11L-2351 which was sampled from a non-venomous reptile. This isolate belongs to the serovar Montevideo, which is frequently associated with outbreaks of human salmonellosis [85–87]. In reptiles, the end product of the purine catabolic pathway is not allantoin, but uric acid [9]. The consequent absence of allantoin from the snake gastrointestinal tract could explain why a substantial number of S. enterica clade B were found in snakes.
It is possible that the gain and loss of allantoin catabolic genes is relevant to host specificity. A relationship between the pseudogenization of the allantoin metabolic genes and niche adaptation has also been proposed for the invasive nontyphoidal Salmonella (iNTS) reference isolate for S. Typhimurium: D23580 [49,88]. Compared with S. Typhimurium isolate 4/74, which shows a broad host range, D23580 is unable to utilise allantoin as a sole carbon source, consistent with the adaptation of invasive Salmonella in Africa towards non-allantoin producing hosts [49,88]. Furthermore, accumulation of pseudogenes in the allantoin degradation pathway has been reported in host-restricted Salmonella serovars which cause invasive disease, suggesting that the ability to grow on allantoin is a marker of a switch from enteric to invasive disease [89]. These findings may reflect the clinical observation that snake-acquired salmonellosis is frequently an invasive disease that commonly results in hospitalisation, compared to disease caused by Salmonella derived from allantoin-producing hosts.
Although reptiles are known to harbour a diverse range of Salmonella bacteria, until now Salmonella carriage has not been examined in many key reptilian species. Here, we have shown that venomous snakes harbour and shed a wide variety of Salmonella serovars that represent much of the spectrum of the Salmonella genus and are phylogenetically distributed in a similar way to Salmonella found in non-venomous reptiles. We demonstrated that venomous snakes can carry and excrete Salmonella serovars which cause human disease. One of the Salmonella isolates was resistant to first-line antimicrobial agents. It is possible that venomous snakes represent a previously uncharacterised reservoir for Salmonella both in captive settings and in the wider environment. Further study is required to investigate the relationship between clinical cases and reptile-derived Salmonella in tropical regions inhabited by venomous reptiles such as Africa. We believe that our study provides a good baseline for this future work.
Reptiles are an ideal population of animals for the study of genus-level evolution of Salmonella because they carry phylogenetically diverse isolates that belong to the majority of Salmonella subspecies. By demonstrating the phenotypic and genotypic conservation of metabolic properties across three phylogenetic groups of Salmonella we have shed new light on the evolution of Salmonella serotypes.
|
10.1371/journal.pgen.1005804 | Pedigree- and SNP-Associated Genetics and Recent Environment are the Major Contributors to Anthropometric and Cardiometabolic Trait Variation | Genome-wide association studies have successfully identified thousands of loci for a range of human complex traits and diseases. The proportion of phenotypic variance explained by significant associations is, however, limited. Given the same dense SNP panels, mixed model analyses capture a greater proportion of phenotypic variance than single SNP analyses but the total is generally still less than the genetic variance estimated from pedigree studies. Combining information from pedigree relationships and SNPs, we examined 16 complex anthropometric and cardiometabolic traits in a Scottish family-based cohort comprising up to 20,000 individuals genotyped for ~520,000 common autosomal SNPs. The inclusion of related individuals provides the opportunity to also estimate the genetic variance associated with pedigree as well as the effects of common family environment. Trait variation was partitioned into SNP-associated and pedigree-associated genetic variation, shared nuclear family environment, shared couple (partner) environment and shared full-sibling environment. Results demonstrate that trait heritabilities vary widely but, on average across traits, SNP-associated and pedigree-associated genetic effects each explain around half the genetic variance. For most traits the recently-shared environment of couples is also significant, accounting for ~11% of the phenotypic variance on average. On the other hand, the environment shared largely in the past by members of a nuclear family or by full-siblings, has a more limited impact. Our findings point to appropriate models to use in future studies as pedigree-associated genetic effects and couple environmental effects have seldom been taken into account in genotype-based analyses. Appropriate description of the trait variation could help understand causes of intra-individual variation and in the detection of contributing loci and environmental factors.
| Unravelling overall trait architecture of complex traits and diseases is important for phenotype prediction and disease prevention and correct modelling of the trait will further aid discovery of causative loci. Here we take advantage of genome-wide data and a large family-based study to examine the role of common genetic variants, pedigree-associated genetic variants, shared family environment, shared couple environment and shared sibling environment on 16 anthropometric and cardiometabolic traits. By analysing up to ~20,000 Scottish individuals, we find that common genetic variants, pedigree-associated genetic variants and recently-shared environment of couples are the most important contributors to variation in these traits, while past family and sibling environment have a limited impact. Further studies on the pedigree-associated genetic variation and the shared couple environment effect are needed, as little research has been devoted to them so far.
| Phenotypic variation for a quantitative trait is attributable to the summed effects of genetic and environmental influences together with any covariances and interactions. The proportion of phenotypic variance contributed by genetic variation is termed the heritability (h2) [1]. The heritability scales the influence of genetic and environmental factors on phenotypic variation. This provides us with insights into the genetic and environmental architecture of human complex traits and our potential ability to dissect out loci associated with trait variation and is also useful for the prediction of heritable disease risk [2,3]. As a consequence, such knowledge is of potential value for clinical diagnosis, therapy, prevention and prognosis [4]. Therefore, obtaining unbiased estimates of variation caused by different factors and the heritability of traits relevant to health and disease processes is important.
A classic approach to gauging the heritability in humans is by comparing the observed phenotypic similarity to the expected genetic resemblance between relatives inferred from family pedigrees [5]. This method evaluates the pedigree based heritability (hped2) indirectly without requiring information on the inheritance of individual loci and thus, is quite practical and still widely-used in twin, family and other pedigree studies [6,7]. Note that, hped2 is often considered to be an estimate of the true heritability h2. Genome-wide association studies (GWAS), on the contrary, identify causal loci through their association with recorded genetic markers and then aggregate the proportion of variance explained by statistically-significant variants [8,9], which is sometimes referred to as the “GWAS heritability” (hGWAS2). Each approach has its limitations and drawbacks. Pedigree studies require genealogical information from known relatives to deduce their expected genetic resemblance and hped2 may be biased due to the factors shared among relatives (including dominance, epistasis, common environment, genetic-by-environment correlation and genetic-by- environment interaction) if such effects are present and the available pedigree structure does not allow these to be accounted for in the analysis [10–12]. Although GWAS have been very successful at discovering novel loci for a range of polygenic disease and complex traits, they have been less successful at capturing the full extent of known trait genetic variance [11,12]. This is probably because of their failure to detect particular types of variants such as common variants with small effects, rare variants, copy number variants and structural variants, as a consequence of inadequate sample size, genotyping platform design and analyses used, together with the stringent statistical tests applied [10,13,14]. As a result, there usually is a substantial gap between the estimates of hped2 and hGWAS2, often termed the “missing heritability” [11,15].
Recently, Yang et al. [16,17] have championed an approach, known as GREML [18], to estimate the amount of trait variance explained by SNPs. The estimation of the SNP (or genomic) heritability (hg2), which refers to the additive genetic effects captured by genotyped SNPs, utilises a matrix comprising realised genetic relationships inferred from genomic marker data originally gathered for GWAS (known as genomic relationship matrix or GRM) [16,17]. The hg2 estimate from this approach, when estimated using unrelated individuals, lies between the hped2 and hGWAS2 estimates, and has been considered as a lower limit for the former and an upper limit for the latter [11,12]. As an example, for height, hGWAS2,hg2 and hped2 from three different studies are 0.10, 0.45 and 0.80 respectively [5,8,17]. This suggests that a substantial proportion of the genetic contribution to trait variation is SNP-associated and hence contributes to hg2 but not all this variation is detected by current GWAS, probably due to a combination of insufficient sample size and stringent significant thresholds employed. The difference between hg2 and hped2 may be largely due to trait associated variants not in linkage disequilibrium (LD) with genotyped SNPs, such as rare variants, copy number variants (CNV) and other structural variants as mentioned above. Variation associated with such effects is captured by hped2 due to strong LD in relatives [19].
Recent studies have started dissecting the heritable component of variation and other components shared among relatives by studying more complex populations made-up of both unrelated individuals and extended pedigrees [11,12,19]. For instance, Zaitlen et al. [12] have demonstrated that simultaneously including in a GREML analysis a GRM and a modified GRM (in which entries smaller than a certain threshold in the GRM are set to zero) can be used to jointly estimate SNP-associated and total heritabilities in the presence of relatives. We also note that shared environment may be an important contributor to heritability inflation when close relatives are included in analysis.
In this study, we use data from a single homogeneous cohort consisting of approximately 20,000 adults with varying degrees of relationships sampled from Scotland. The individuals have data on over 520,000 SNPs distributed across the autosomes. The dense marker information together with extended genealogical information allows us to partition the phenotypic variance and explore the genetic and environmental effects shared among related individuals (both biological relatives and couples).
We analyse eight anthropometric traits, comprising height, weight, fat, body mass index (BMI), hips, waist, waist-to-hips ratio (WHR) and a body shape index (ABSI) [20] and eight cardiometabolic traits, comprising levels of creatinine, urea, total cholesterol (TC) and high density lipoprotein (HDL) in serum, level of glucose in blood, systolic blood pressure (SBP), diastolic blood pressure (DBP) and heart rate (HR).
In our work, we implement alternative models to estimate effects that might contribute to the variation in the 16 traits analysed. Results show that, with these data, we can separate total genetic variation into SNP-associated and pedigree-associated genetic influences. We also observe that past family environment and shared full-sibling environment generally have a limited impact on trait variation, whereas the effect in couples of living in the current (shared) environment is always important in our data.
We conducted variance component analyses to dissect the phenotypic variation for traits recorded in the Generation Scotland: Scottish Family Health Study (GS:SFHS) cohort [21] into genetic and environmental factors. Analyses utilised a mixed-model approach implemented in a restricted maximum likelihood (REML) framework using the GCTA software [16]. The population was divided into two tranches of approximately equal size and genotyped in two stages. All initial analyses were performed with the first 10,000 genotyped individuals, (named GS10K). GS10K comprised small nuclear families (largely two parents and two offspring) together with unrelated individuals, although inevitably there were second degree and more distant relationships included. The second tranche completed the genotyping of the rest of the population (another 10,000 individuals) including further relatives in incomplete families (e.g. missing samples from parents and additional siblings, as well as other relationships), resulting particularly in a proportional increase in the number of second and third degree relationships (Table 1). To confirm results obtained from GS10K, some of the analyses were repeated in the whole 20,000 individual sample (named GS20K).
We first explored the extent to which estimates of hg2 were inflated by the inclusion of relatives. We subsequently analysed our data allowing trait variation to be potentially influenced by both genetic and environmental effects. We assumed that the genetic effects comprised additive genetic effects associated with genotyped SNPs (hg2) and additional additive genetic effects associated with pedigree but not with genotyped SNPs (hkin2), and we assumed that the environmental effects potentially comprised nuclear family effects (ef2) common to both parents and offspring, full-sibling effects (es2) common to just siblings and couple effects (ec2) common to just the members of a couple (Fig 1). The total heritability, termed hgkin2 in this manuscript, referred to as hIBS>t*2 in Zaitlen et al. [12] and comparable to hped2 from traditional pedigree studies, was estimated as the sum of hg2 and hkin2 for each model. To allow estimation of the influence of each effect, we generated five design matrices: GRMg, GRMkin, ERMFamily, ERMSib and ERMCouple respectively, where GRM refers to genomic relationship matrices and ERM refers to environmental relationship matrices.
For brevity, we named different alternative models using abbreviations according to first subscript letter of the effects examined. We coded ‘G’ for GRMg, ‘K’ for GRMkin, ‘F’ for ERMFamily, ‘S’ for ERMSib and ‘C’ for ERMCouple–e.g. model ‘GKC’ = GRMg + GRMkin + ERMCouple. All models included a residual matrix (allowing effects specific to an individual that were not shared with any other member of the population).
We identified the most appropriate model for each trait by a stepwise model selection process via removing non-significant components from the full model based on a Wald test of their estimated effect and a likelihood ratio test (LRT), and we estimated the effects of significant factors using the selected models in GS10K. We repeated the model selection and corresponding variance component analyses in GS20K to identify differences resulting from analysing a more complex population structure, encompassing a larger proportion of close relationships.
More details about traits, matrices and models are given in Material and Methods and S1 Table and S2 Table. In the main manuscript, we only list results for the final models identified by the model selection procedure and the full model, but a comprehensive list of estimates obtained for the different effects for each trait and each model is available in S3 Table and S4 Table.
Model robustness and the effectiveness of the model selection were tested using simulated data based on GS10K.
We conducted a simulation study using real genotype and pedigree information from GS10K to evaluate the robustness of our models. To make computation feasible, we mainly focused on data simulated under the simplest and most complex models (models ‘G’, ‘K’, ‘F’, ‘S’, ‘C’, ‘GK’, ‘GF’ and ‘GKFSC’) and those representing the commonest conclusions of model selection in analyses of the real GS10K data (models ‘GF’, ‘GFS’, ‘GKC’ and ‘GKSC’). S5 Table shows the simulated and observed values for each parameter as well as the model we used for analyses in different scenarios.
In the first scenario, we examined the performance of our models (models ‘G’, ‘K’, ‘F’, ‘S’ and ‘C’) when simulated phenotypes were only contributed by one of the five corresponding effects plus residual variation. Under these models (S5 Table), the mean of overall estimates per parameter was very close to its simulated value, indicating that our design matrices GRMg, GRMkin, ERMFamily, ERMCouple and ERMSib worked well in simple models and were able to capture their corresponding effects even when the simulated variance associated with an effect was low (≤ 3%).
In the second scenario, we evaluated the performance of our models (models ‘GK’ and ‘GF’) when the simulated phenotypes were determined by SNP-associated genetic effects and one of the familial effects (either pedigree-associated genetics or nuclear family environment) plus residual variation. Results (S5 Table) indicate that, in cohort with familial structure, failure to account for or inaccurate modelling of familial effects (i.e. when models used were inconsistent with phenotypic contributors) would result in upward bias for hg2 in the presence of relatives. However, this upward bias due to the confounding familial factors could be eliminated by either excluding nominally related individuals or using the appropriate models for analysis. The former method removes the ability to estimate the familial effects as well as reducing the sample size, whereas using the appropriate models, estimates obtained were very close to their parameter settings and gave a good idea of the magnitude and approximate values of SNP and familial effects as well as the total proportion of variance explained by additive genetics (hgkin2=hg2+hkin2), despite the fact that the means of estimates of hg2,hkin2 and ef2 were usually significantly different from the original parameter settings.
In the third scenario, we inspected the performance of the full model ‘GKFSC’ and models selected from analyses of real phenotypes in GS10K other than ‘GF’ (models ‘GFS’, ‘GKC’ and ‘GKSC’). Results (S5 Table) demonstrate that all models were robust in terms of the mean of overall estimates per parameter being either unbiased or very close to original settings.
Fig 2 summarizes the main results from these simulations, showing the overall performance of our design matrices from simple models to complex models. The median of estimates for each component was unbiased across simple and complex models, however, the estimates for hkin2,ef2 and ec2 were quite variable in the full model, probably due to limitations imposed by the data structure. All of the above verify the robustness of our models.
Although we confirmed that our models were robust (S5 Table and Fig 2), the potentially high correlation between ERMFamily matrix and combined ERMCouple and GRMkin matrices may make it challenging to jointly estimate hkin2,ef2 and ec2 accurately in our sample as the standard errors for those parameter estimates obtained from the full model were high (S4 Table). Thus the most challenging part of our study may be to precisely dissect pedigree-associated genetic effects, shared nuclear family environment and shared couple environment. Therefore, we performed model selection using simulated data to test our model selection procedure where simulated phenotypes were contributed by moderate SNP-associated genetic effects and low sibling environmental effects plus a) moderate nuclear family environmental effects but low pedigree-associated genetic effects and couple environmental effects; b) low nuclear family environmental effects but moderate pedigree-associated genetic effects and couple environmental effects; or c) moderate nuclear family environmental effects, pedigree-associated genetic effects and couple environmental effects. All scenarios included residual variation.
S6 Table shows the parameter settings and the summary of model selection procedure performance for these scenarios. We expected that our model selection procedure was able to identify SNP genetics (GRMg) and nuclear family environment (ERMFamily) or SNP and pedigree genetics (GRMkin) and couple environment (ERMCouple) or SNP and pedigree genetics and nuclear family and couple environment accordingly, since they were the major factors in each corresponding scenario.
As results demonstrated, in all situations our model selection procedure generally (≥80%) selected the appropriate model which contains all major components of phenotypic variation. The remaining times in the first two of these scenarios, pedigree-associated genetic effects or those plus shared couple environment were selected instead of nuclear family environmental effects or vice versa, and in the remaining two replicates in the third of these scenarios we missed pedigree-associated genetic effects. In addition, our model selection never fully detected all minor contributions to the phenotype in the first two of these scenarios when the minor effects were too small (e.g. effects contribute to ≤5% of the phenotypic variance).
Both issues identified above (~20% chance of selecting inappropriate models and failure to identify all minor effects) are likely to have been due to limitations in the data structure of GS10K, which provides too few of the appropriate relationships for corresponding effects (pedigree-associated genetics, nuclear family, sibling and couple environment) to resolve correlations between parameters and detect minor effects. These limitations have been greatly ameliorated in the GS20K data.
We also conducted variance component analyses using the final selected model for each replicate (S6 Table). For those replicates that had appropriate models after model selection, the estimates of factors that remained in the models were usually close to, and not significantly different from, their simulated values, indicating that the results from selected models were reliable. More details about simulation study can be found in S1 Text, S5 Table and S6 Table.
In the first analyses of the real data, we looked for evidence of familial effects (either pedigree-associated genetics or nuclear family environment) in our cohort. As shown by simulation (S5 Table), if there were any familial effects, we should obtain inflated estimates of hg2 when we conducted variance component analyses using model ‘G’ in the presence of relatives, compared to the estimates of hg2 given from the unrelated subpopulation. GS10K consists of nearly 10,000 genotyped individuals with multiple degrees of relationship, which allows us to explore the impact of familial effects on hg2 estimation in this cohort.
Table 1 shows the population structure of genotyped individuals in GS10K. The degree of relationship between two individuals was identified according to an approximate range of the expected pair-wise relatedness (r), which was from 0.5i-0.5 to 0.5i+0.5 for ith degree relatives (e.g. pairs of individuals with relatedness from 0.354 to 0.707 were considered as 1st degree relatives).
With these criteria, GS10K consisted of more than 3,500 pairs of 1st degree relatives, around 450 pairs of 2nd and 500 pairs of 3rd degree relatives, but the majority of pairs of individuals (over 99.9%) were genetically unrelated (more distant than 5th degree relatives, r ≤ 0.022). In total, there were around 6,600 unrelated individuals (defined using the criteria described above) in GS10K.
We estimated hg2 for each trait using model ‘G’ for subpopulations of GS10K made-up of individuals with different degrees of relatedness (using the upper bound of the expected relatedness of each category as GRM cut-off points in GCTA). Fig 3 shows how hg2 estimates for height, BMI and HDL changed as we progressively included more closely related individuals in the relationship matrix. Results for the remaining traits are shown in S3 Table.
In general, hg2 estimates were stable as we gradually added more closely related individuals in the analyses until the inclusion of 1st degree relatives that resulted in inflation of the estimates (Fig 3 and S3 Table). Based on our results, hg2 was overestimated only when 1st degree relatives were included. For glucose and DBP, the hg2 estimates did not appear inflated after 1st degree relatives were included, suggesting that these traits were not affected by familial effects (S3 Table).
The increase in hg2 estimates resulting from the inclusion of 1st degree relatives provided evidence of familial variation in our cohort. However, it is not clear whether these familial effects are due to pedigree-associated genetic effects or shared nuclear family environment or both because either of them has the ability to inflate hg2 estimates (this was also observed in the simulation data: S5 Table: scenario ii). Therefore, we attempted to tease out this familial variance from the total phenotypic variance and dissect the familial variation as well as the remaining trait variation further using the full model ‘GKFSC’ and the stepwise selection procedure to define a final model containing the most important effects contributing to trait variation.
Table 2 shows the results for final models selected from stepwise model selection strategies and for the proportions of total phenotypic variance explained by different effects using final models, as well as for those obtained using the full model.
The mean estimates for hg2,hkin2,ef2,es2 and ec2 across all traits in the full model were 0.18, 0.22, 0.03, 0.03 and 0.11, respectively. However, the majority of estimates for parameters other than hg2 obtained using the full model were not significantly different from zero according to either the Wald test or LRT performed and had large standard errors in general. These results suggest that the full model ‘GKFSC’ may suffer from the inclusion of correlated factors, as foreseen in the simulation study, probably due to a low number of different types of pairwise relationship in GS10K.
Therefore, we utilised a model selection procedure designed to provide more precise estimates of the parameters retained in a more robust and parsimonious final model, where the least significant effects are removed from the model. More details about the selection procedure are given in Material and Methods. We have demonstrated the effectiveness of our model selection procedure by simulation in the previous section and S6 Table.
As shown in Table 2, SNP-associated genetic effects (represented by GRMg) were retained in the final models for all 16 traits, indicating that all traits examined here are heritable. Regarding variation associated with families, pedigree-associated genetic effects (represented by GRMkin) and nuclear family environmental effects (represented by ERMFamily) were retained in the final models for 10 and 4 out of 16 traits respectively. However, in GS10K, the data structure did not allow for both familial effects to be retained together in the final models for any trait. Additionally, the final models for glucose and DBP included neither GRMkin nor ERMFamily, which is consistent with the previous conclusion derived from S3 Table, suggesting that familial effects may be limited for these traits.
The additional environmental influences of couple environmental effects (represented by ERMCouple) were retained in the final models for 12 out of 16 traits and sibling environmental effects (represented by ERMSib) only remained for creatinine and TC.
Although the final model varied between traits, the model ‘GKC’ was most often selected (9 out of 16 traits) in the model selection procedure in GS10K. Therefore, this suggests that the common environment shared by couples, SNP-associated and pedigree-associated genetic effects are important for the control of a large proportion of the human complex traits we examined, while the shared family and full-sibling environment have a more limited impact
SNP-associated genetic effects (GRMg) in the final models provided estimates of hg2 ranging between 0.10 and 0.30 with a mean of 0.19 for the 15 traits, excepting height for which nearly half of its phenotypic variation (0.47) was SNP-associated.
For the 10 traits that retained pedigree-associated genetic effects (GRMkin) in the final models, the estimates of hkin2 ranged from 0.13 to 0.36 with a mean of 0.26, except for creatinine for which nearly half of its phenotypic variation (0.45) was pedigree-associated. For the 10 traits that retained both GRMg and GRMkin in the final models, the estimates of hkin2 accounted for 56% of the total heritability (hgkin2=hg2+hkin2).
Regarding nuclear family environmental effects, the estimates of ef2 for 4 traits that retained ERMFamily in the final models were of 18% for anthropometric and of 10% for cardiometabolic traits.
Creatinine and TC were the only two traits for which the common sibling environment (ERMSib) was kept in the final models, and es2 contributed 7% and 12% of their phenotypic variance respectively.
For those 12 traits that demonstrated evidence of couple effects (i.e. retained ERMCouple in the final models), ec2 accounted for 13.5% of the phenotypic variance on average (of 15% for anthropometric traits and of 11% for cardiometabolic traits).
Compared to the results from the full model in Table 2, using the selected final models provided similar but more precise (i.e. with smaller standard errors) parameter estimates. Therefore, whereas the full models gave a general picture of the important components in the architecture of the traits, the final selected models provided a parsimonious model with more precise estimates of the most important effects.
We added an extra 10,000 genotyped and phenotyped individuals from the same population, providing 20,000 individuals in total, in order to confirm and build upon the results of the model selection in a more complex data set. The difference in sample sizes and numbers of different relationships between GS10K and GS20K is shown in Table 1. The extra 10,000 genotyped individuals in GS20K consisted mainly of the relatives of those already genotyped in GS10K, which substantially increased the proportion of 2nd and 3rd degree and sibling relationships in GS20K. We repeated the model selection procedure and corresponding variance component analyses using selected models in GS20K to identify changes resulting from the increased complexity and sample size of the population.
Results for model selection and variance component analyses using the final selected model as well as the full model are shown in Table 3. In general, the parameter estimates obtained from the full model in GS20K were similar to those obtained from the full model in GS10K but the number of non-significant estimates were much lower due to smaller standard errors. Note that standard errors of estimates are not only reduced using GS20K, but, unlike results from GS10K in Table 2, are also similar between full and reduced models, suggesting the change is due to improved structure of the data to separate effects as well as increased sample size.
The final models selected from model selection in GS20K were generally similar to those in GS10K, but, owing to the presence of more nuclear family members and siblings in GS20K, we now had better power to detect the past environmental effects (either nuclear family environment or sibling environment), although the estimated effects were usually small. Moreover, due to an increased number and higher proportion of 2nd and 3rd degree relatives, we had better resolution for familial effects in GS20K. Pedigree-associated genetics and nuclear family environment were now separable and the data structure in GS20K can provide sufficient evidence for both types of familial effects. For weight, urea, TC and HR, familial effects switched from nuclear family environment in GS10K to pedigree-associated genetics or pedigree-associated genetics plus nuclear family environment in GS20K. However, as in GS10K (Table 2 and S3 Table), there was still no evidence of either genetic or environmental familial effects for glucose and DBP in GS20K. The results from final selected models in GS20K are summarized in Fig 4.
The heritability estimate is nearly 90%, 60% and 60% for height, creatinine and HDL respectively, and for the remaining anthropometric and cardiometabolic traits, it ranges from 30%-50% and 20–30% for the two types of trait, respectively (Fig 4B). Although the proportion of genetic variance explained by SNP-associated and pedigree-associated genetic effects varies across traits, each genetic effect explains around 50% of the genetic variance on average (Fig 4C). In GS20K, the most commonly selected model was ‘GKSC’ (10 out of 16 times, Fig 4A and Table 3). SNP-associated genetic effects, pedigree-associated genetic effects, sibling environment and couple environment appeared in the final models for 16, 14, 12 and 16 out of 16 times respectively and the means of estimates for hg2,hkin2,es2 and ec2 for traits which retained corresponding matrices (GRMg, GRMkin, ERMSib and ERMCouple respectively) in the final models were of 0.20, 0.23, 0.05 and 0.11 respectively (Fig 4A and Table 3). For the nuclear family environment, the mean of estimates for ef2 for 4 traits which retained ERMFamily in final models was of 0.04 (Fig 4A and Table 3). On average across traits, our environmental matrices and the final selected models retained through our model selection procedure could explain ~16% and ~56% of the total phenotypic variance respectively (Fig 4B).
The major change in GS20K compared to GS10K is the significant evidence of effects of the sibling environment, particularly for cardiometabolic traits, resulting from the higher proportion of sibling relationships in GS20K (more than 12 times compared to GS10K, Table 1). However, the sibling effects were only 5% on average and were still relatively low compared to genetic effects and couple environment. Therefore, despite the change in population structure in GS20K, the major components for anthropometric and cardiometabolic traits were SNP-associated and pedigree-associated genetic effects and couple environment as they were in GS10K (Table 2).
The aim of this study was to better understand the architecture of human complex traits by dissecting phenotypic variation into SNP-associated additive genetic variation (hg2), pedigree-associated genetic variation (hkin2) and environmental influences of common environment shared by nuclear family members (ef2), full-siblings (es2) and couples (ec2). We generated five design matrices GRMg, GRMkin, ERMFamily, ERMSib and ERMCouple to describe the five effects and we examined 16 human complex traits using genome-wide genotype data and genealogical information in the Generation Scotland: Scottish Family Health study (GS:SFHS) comprising samples from up to 20,000 individuals.
The results of these analyses suggest that SNP-associated genetic effects, pedigree-associated genetic effects and current environment shared by couples were the major contributors to phenotypic variation for anthropometric and cardiometabolic traits. Past environmental influences, such as shared sibling environment or nuclear family environment, made relatively small or undetectable contributions to trait variation (Table 2 and Table 3). The relative importance of a couple or spousal effect for most traits was also noted by Liu et al. [22], in analyses based only on pedigree relationships, although they did not find a significant spousal effect for cholesterol, HDL or glucose for which a significant couple effect was detected in this study.
Considering the low number of non-zero off-diagonal entries in ERMCouple (1,283 or 1,767 pairs in GS10K or GS20K), the signal of couple effects was quite strong. We did observe significant phenotypic correlation between couple pairs for almost all traits in our data (S7 Table). For some traits this presumably represents current shared environment due to cohabitation, such as living habits and diet. For traits related to obesity, it is reasonable that current environmental effects are more important than past environmental effects since traits like BMI, fat, HDL and blood pressure are potentially influenced by recent food intake, exercise and medical treatment.
It should be noted that in our sample participants have an average age of ~50 years and individuals currently sharing a common household environment will largely be couples, whereas most individuals involved in sibling and parent-offspring relationships will no longer be cohabiting at the point when the data were recorded. It has been previously reported in obesity studies that common childhood environment only affects individuals in their mid-childhood but the influence does not last past adolescence [23,24]. Therefore, although the impacts of nuclear family or sibling environmental effects on the 16 traits we examined were relatively small, family and sibling environmental effects could be more important in younger cohorts and might be of greater importance for other complex traits and diseases where long-term environment may have an influence on a phenotype that is relatively stable over time.
For some traits, the most obvious example being height, couple effects may also, in part or completely, reflect assortative mating. A study by Keller et al. has shown that h2 estimate for height would be 13% higher with assortative mating than it would have been under random mating [23]. If there was assortative mating for any of the traits which retained ERMCouple in final models but we modelled the couple correlation as an environmental effect, we would expect to obtain biased ec2 estimates. Moreover, modelling assortative mating as an environmental effect removes variance from the residual (“error”) variance. We therefore might obtain an inflated hg2 estimate if we have not taken assortative matting into account and reduce the residual variance as a consequence of modelling assortative matting as an environmental effect. In addition, assortative mating will have consequences for our interpretation of GWAS results as the combined effect of detected loci on the trait variance will be greater than the sum of the effects of the individual loci due to the positive correlations between loci. However, except for height, where the phenotype will be largely fixed by the time of marriage, for most traits it is difficult to determine whether assortative mating and/or shared environment are responsible for observed phenotypic correlations between couples.
Shared sibling environment was undetected for most of the traits in GS10K (Table 2), whereas there was significant evidence of it for many traits in GS20K (Table 3), indicating that the detection power of sibling environment benefits from the increase in number and proportion of sibling relationships (Table 1). Sibling effects, where detected, explained 5%, on average, of the trait variation. Estimated sibling effects may be inflated by non-additive genetics, (i.e. dominance and epistasis). As sibling effects only capture a fraction of the non-additive variation, the actual variation contributed by non-additive genetics might potentially be large and would merit further study.
Our analyses split the genetic variation approximately equally on average across traits between that which was associated with SNPs (hg2) and that which was associated with pedigree (hkin2). A plausible interpretation for the division of genetic effects into hg2 and hkin2 is that hg2 is able to explain the genetic variation attributed by common variants inherited from distant ancestors that are in LD at the population level and are well captured due to association with genotyped SNPs [12]. On the other hand, hkin2 accounts for the genetic variation due to rare variants, CNVs and other structural variation, etc. that cluster in specific families and are captured due to strong linkage in high-order pedigrees but are not in population-wide LD with common SNPs.
We compared hg2 and hgkin2 (calculated as hg2+hkin2) estimates obtained in final models from model selection in GS20K to two relevant publications from Zaitlen et al. [12] and Vattikuti et al. [19] that also explored the influence of including relatives on h2 estimation in family-based studies and compared hgkin2 estimates obtained in final models in GS20K to published twin studies [6,24–31]. Comparisons are shown in Table 4.
When comparing with two family-based GREML studies (Table 4), our hg2 and hgkin2 estimates from final models are generally higher than published relevant results, except for the hg2 estimate for SBP and the hgkin2 estimates for glucose and SBP. When comparing with twin studies (Table 4), our hgkin2 estimates for all anthropometric traits, urea, TC and HDL given by final selected models in GS20K are reasonably close to reported hped2 estimates, which suggests little missing heritability. Hence, our results provide no evidence that heritabilities given by previous twin studies were inflated for these traits. For glucose, SBP, DBP and HR, however, our hgkin2 estimates are significantly lower than previously published estimates of hped2, whereas for creatinine, hgkin2 is significantly larger.
To validate the analytical approach used in this study and to evaluate model robustness, we conducted a detailed simulation study using real genotype and pedigree information obtained from GS10K. The simulation results confirmed that our models were generally robust (S5 Table). However, the inevitable correlations between our design matrices can, under some circumstances, make it challenging to partition variance for correlated factors in variance component analyses and accurately discriminate between competing models in model selection. Nonetheless, any influence of inaccurately partitioning variance among correlated matrices was relatively limited and our models were always able to provide us with a good idea of the magnitude of corresponding effects as the mean estimate for each parameter was always very close the simulated settings when the model used for analysis matched the simulated sources of trait variation.
The effectiveness of the model selection procedure was also validated using the simulated data with the model selection procedure often (≥80%) resulting in models containing all major phenotype components (S6 Table). However, due to the limited number of appropriate relationships in GS10K to resolve correlations between matrices and to detect factors with small effects, our model selection procedure may omit minor effects (contributing 5% or less of the trait variance, for example). In addition, the procedure may sometimes identify incorrect models (not being able to distinguish familial effects as mentioned in the simulation study and S6 Table) and this might be the case for weight, urea, TC and HR in Table 2. However, with sufficient data from higher order pedigree relationships, as was the case in GS20K, the impact of covariances between design matrices in first order relatives (parent-offspring, siblings and couples) are mitigated and further components of variance became separable (Table 3).
To sum up, we provide evidence that for the traits we have analysed, heritabilities are divided approximately evenly between pedigree-associated and SNP-associated genetic effects. This is the case even when, as here, we have taken care to consider various models of environmental covariation of first-degree relatives (including couples). It appears that confounding factors like dominance, shared full-sibling environment and the past rearing environment seem to have relatively small contribution to phenotypic variation for these traits in our population. We find that current shared environment of couples is able to account for another ~11% on average of the phenotypic variation of human complex traits. This has been seldom mentioned in previous heritability studies but we note that as an effect that inflates the covariance between nominally unrelated individuals, it should not substantially bias or inflate hped2 and hgkin2. It should be taken into account that couple effects may also be present in cohorts of unrelated individuals which may often include couples but ignore any correlation between them. Therefore, it might bias hg2 from genotype-based studies which do not account for such couple effects and could have an impact on GWAS studies.
Overall, our work shows that SNP-associated genetic effects, pedigree-associated genetic effects and current shared couple environmental effects are three fundamental components of phenotypic variation for traits related to anthropometrics and cardiometabolism and current shared environmental effects have more impact than past shared environmental effects. This also has implications for models to be used in further studies of the architecture of complex traits including utilising the appropriate models for GWAS and related analyses and for personalised disease risk prediction.
The data were obtained from the Generation Scotland: Scottish Family Health Study (GS:SFHS). Ethical approval for the study was given by the NHS Tayside committee on research ethics (reference 05/s1401/89) and participants provided written consent. Governance of the study, including public engagement, protocol development and access arrangements, was overseen by an independent advisory board, established by the Scottish government
Our dataset came from the Generation Scotland Scottish Family Health Study (GS:SFHS) project (http://www.generationscotland.org), which was collected by a cross-disciplinary collaboration of Scottish medical schools and the National Health Service (NHS) from Feb 2006 to Mar 2011 [21,32].
Data for 16 complex traits were used. These were 8 anthropometric traits: height, weight, fat, body mass index (BMI=WeightHeight2), hips, waist, waist-to-hips ratio (WHR) and a body shape index (ABSI =Waist Circumference×Height5/6Weight2/3) [20] and 8 cardiometabolic traits: levels of creatinine, urea, total cholesterol (TC) and high density lipoprotein (HDL) in serum and glucose in blood after a four hour fast period, systolic blood pressure (SBP), diastolic blood pressure (DBP) and heart rate (HR). None of the traits was adjusted for medication or fasting status. We explored the phenotypic distributions of these traits and conducted natural logarithm transformations for them, except for height, sodium and fat, to obtain approximate normal distributions. We set phenotypes with values greater or smaller than the mean ± 4 standard deviations (after adjusting for sex, age and age2) to missing.
Data also contained the information of sex, age, clinics where the phenotypes were measured and Scottish Index of Multiple Deprivation (SIMD, an environmental ranking based on living areas, [33]). A descriptive analysis can be seen in S1 Table.
The first set of analyses presented in the manuscript are based on a data set of nearly 10,000 individuals from GS:SFHS (GS10K). These have multiple degrees of kinships, including 5,061 family members from 1,612 nuclear or extended families, and were genotyped with the Illumina OMNiExpress chip (707,686 SNPs). We conducted data quality control in Plink v1.07 [34] and GenABEL v1.7–6 [35]. SNPs with a minor allele frequency (MAF) < 0.05, a Hardy-Weinberg Equilibrium’s (HWE) p-value <10−6 and a missingness > 2% were excluded. Duplicate samples, gender discrepancies and individuals with more than 5% missingness were also removed. After the quality control we kept 9,863 individuals genotyped for 550,796 common SNPs over the 22 autosomes.
An extended dataset (GS20K) was used to validate the results obtained with GS10K and evaluate the effect of including further close relationships in our data. The extra 10,000 individuals were genotyped with the same chip and quality control was performed using the same criteria as in the GS10K. After quality control, GS20K consisted of 20,032 individuals, 18,293 of whom came from 6,578 nuclear or extended families, and 519,729 common SNPs across the 22 autosomes.
A comparison of the difference in relationships between GS10K and GS20K can be seen in Table 1.
Our model allows trait variation to be influenced by the genetic effects associated with SNPs (hg2) and pedigree (hkin2) and the environmental effects shared by families (ef2), couples (ec2) and full-siblings (es2), (Fig 1). To estimate the influence of each effect, we generated five design matrices: GRMg, GRMkin, ERMFamily, ERMSib and ERMCouple.
A genomic relationship matrix (GRM) contains estimated genomic relatedness between pairs of individuals calculated from identity-by-state marker relationships as in Yang et al. [16,17].
Each off-diagonal entry in the GRM represents the realised genomic relationship between a pair of individuals:
1N∑i=1N(xji−2pi)(xki−2pi)2pi(1−pi)
where, pi is the minor allele frequency (MAF) for SNP i, xji or xki is the allelic dose for individual j or k at locus i (x = 2 if the individual carries two rare alleles, x = 1 if the individual is heterozygous, x = 0 if the individual carries two common alleles) and N is the total number of SNPs.
Each entry on the diagonal represents the inbreeding coefficient calculated as:
1+1N∑i=1Nxji2−(1+2pi)xji+2pi22pi(1−pi)
We used GCTA [16] to generate GRMg and obtained GRMkin by modification of GRMg in R [36]. Their definitions are identical to matrices KIBS and KIBS>t in Zaitlen et al. [12] respectively.
GRMg: a GRM estimated using all common SNPs, and designed to capture the additive genetic variance explained by common SNPs in the population sample.
GRMkin: a modified GRM calculated as in Zaitlen et al. [12] designed to estimate the extra genetic effects associated with pedigree, the variance explained by shared genetic factors in close relatives. GRMkin was created by setting to 0 all entries in GRMg smaller than 0.025.
The number of entries different from 0 in each of the matrices is shown in Table 1.
An environmental relationship matrix (ERM) is a covariance matrix designed to capture the variance due to common environmental effects shared among a specified group of individuals.
The ERM coefficient for each pair of individuals is 1 in if they share a particular environment, e.g., living in the same area or coming from the same family; otherwise, it is 0. Each entry on the diagonal is 1.
We generated 3 different ERMs in R [36]: ERMCouple, ERMSib and ERMFamily.
ERMCouple: ERMCouple was designed to capture the common environmental effects shared between a couple. The ERM coefficient of two individuals was 1 if they were identified as a couple, defined as a pair of individuals with at least one offspring within GS:SFHS. Each entry on the diagonal was 1.
ERMSib: ERMSib was designed to capture the common environmental effects shared between full-siblings. The ERM coefficient of two individuals was 1 if they were identified as full-siblings. Each diagonal entry was 1.
ERMFamily: ERMFamily was designed to capture the common environmental effects shared within each nuclear family comprising parents and offspring. The ERM coefficient of two individuals was 1 if they were identified as a parent-offspring pair, full-siblings or a couple. The ERM coefficient of two individuals was 1 if they were identified as nuclear family members, including parent-offspring, couple and full-sibling relationships. Each diagonal entry was 1.
The number of entries different from 0 in each of the environmental matrices is shown in Table 1. Details about model and matrices we defined can be seen in Fig 1.
We used the genomic and environmental matrices described above to partition the phenotypic variance observed for the traits using a mixed model in a restricted maximum likelihood (REML) framework. The analyses were implemented in GCTA [16]. The equations used to evaluate each model were the subsets of the full model:
y=Xβ+gg+gkin+ef+es+ec+ε,with
V=GRMgσg2+GRMkinσkin2+ERMFamilyσef2+ERMSibσes2+ERMCoupleσec2+Iσε2
where y is an n × 1 vector of observed phenotypes with n being the sample size (number of individuals), and V the total phenotypic variance matrix, β is an m × 1 vector of fixed effects with m being the total level of covariates and X its design matrix with dimension n × m, gg is an n × 1 vector of the total additive genetic effects of the individuals captured by genotyped SNPs with gg∼N(0,GRMgσg2), gkin is an n × 1 vector of the extra genetic effects associated with the pedigree for relatives with gkin∼N(0,GRMkinσkin2), ef, es and ec are n × 1 vectors representing the common environmental effects shared by nuclear family members, full-siblings and couples with ef∼N(0,ERMFamilyσef2), es∼N(0,ERMSibσes2) and ec∼N(0,ERMCoupleσec2) and ε is an n × 1 vector of residuals. We fitted a range of models including different combinations of effects, and named them using abbreviations according to the effects used. We used the codes ‘G’ for GRMg, ‘K’ for GRMkin, ‘F’ for ERMFamily, ‘S’ for ERMSib and ‘C’ for ERMCouple –e.g. ‘GKC’ = GRMg + GRMkin + ERMCouple, and the proportion of total phenotypic variance captured by each matrix was termed hg2,hkin2,ef2,es2 and ec2 accordingly. All models include a residual matrix and the total heritability hgkin2 is always the sum of hg2+hkin2 for any model.
There were 31 different models from all the possible combinations of the five matrices. The abbreviations for each model and the formulae to estimate each term in each model are listed in S2 Table. The results for each model are listed in S4 Table.
In addition to the matrices described (including the residual matrix), we always included the fixed effects of sex, age, age2, sex-by-age interaction, clinic, standardised SIMD and SIMD2 and the first 20 eigenvectors of GRMg (to ameliorate problems associated with data structure).
We conducted a stepwise model selection to find the most appropriate genetic and environmental model for each trait and dissect the phenotypic variation into its components (SNP-associated additive genetic variance, pedigree-associated genetic effects shared among relatives and common environmental effects shared among the specified groups including nuclear family members, couples and full-siblings).
The stepwise selection began with the full model ‘GKFSC’, where all matrices were fitted together. We performed a Wald test and a log-likelihood ratio test (LRT, using a mixture distribution of χdf=02 and χdf=12 with a probability of 0.5 [16]) for each component and removed the component, if any, that was non-significant for both tests at α = 5% level and had the highest p-value for the Wald test. We repeated this process until all the remaining components were significant for at least one test. We did not correct for the limited number of traits analysed so error rates in this procedure should be considered to be on a per trait basis.
In order to evaluate the robustness of our models and the performance of our stepwise model selection, we conducted a simulation study. We simulated, based on the real genotypic information and the real pedigree, different sets of phenotypes for each of the 9,863 individuals in GS10K.
For simulating the genetic effects, we used a similar approach to Zaitlen et al. [12] by dividing the genome into two: even and odd chromosomes, and randomly selecting 550 SNPs from even and odd chromosomes (approximately 1 from each 500 SNPs), representing the observed causal loci that were in LD with the SNPs (SNP-associated genetic effects) and the unobserved genetic variants that were not in LD with the SNP array (pedigree-associated genetic effects) separately. In a later step, only even chromosomes were used to generate GRMg and GRMkin. Each locus was assigned an effect size driven from exponential distribution as in Fisher [37] and the summed effects for even and odd chromosome SNPs were designed to explain hg2 and hkin2 of the trait variance respectively.
For environmental factors, we simulated a sibling environmental effect, a couple environmental effect and two nuclear family environmental effects (youth and adulthood environments) for each individual. The corresponding effect sizes for sibling, couple and nuclear family environmental effects were derived from N(0,es2), N(0,ec2) and N(0,ef2) accordingly and were the same among full-siblings, between couples and among nuclear family members.
In addition, we simulated a random residual effect for each individual, the residuals were derived from N(0,ee2) where ee2 represents the proportion of variance remaining in each of the scenarios. For each scenario, each component (hg2,hkin2,ec2,es2,ef2) was given a proportion of the variance explained and ee2 was 1−hg2−hkin2−ec2−es2−ef2. The final phenotypes would be the sum of these genetic and environmental effects and residuals, and the expected mean and variance of simulated phenotypes were 0 and 1, respectively. More details about how we simulated phenotypes can be found in S1 Text.
We evaluated the robustness of our models under situations where phenotypes were contributed by i) one of the five effects, ii) SNP-associated genetic effects and one of the familial effects (either pedigree-associated genetic effects or nuclear family environmental effects) and iii) SNP-associated genetic effects, familial effects and other environmental effects. All scenarios included residuals and 50 to 100 replicates were analysed for each scenario. The results of simulations were evaluated using a Z-test, which tested whether the mean estimate for each parameter deviated significantly from its simulated value. Note, it was too time consuming to explore all the possible combinations of models and simulated phenotypes, therefore, we mainly focused on the models that were selected in model selection procedure for the real phenotypes in GS10K (Table 2) as well as the fundamental models of our study. More details about the parameter settings for these scenarios can be found in S5 Table.
ERMFamily posited a relationship between siblings, parents-offspring and couples is somewhat confounded with the addition of GRMkin and ERMCouple, making separation and estimation of these effects (ef2, hkin2 and ec2) challenging, as confirmed by the results from analysis of real phenotypes in GS10K (Table 2). Hence, we evaluated the effectiveness of our model selection procedure under situations where phenotypes were contributed by moderate SNP-associated genetic effects and low sibling environmental effects plus a) moderate nuclear family environmental effects but low pedigree-associated genetic effects and couple environmental effects, b) low nuclear family environmental effects but moderate pedigree-associated genetic effects and couple environmental effects and c) moderate nuclear family environmental effects, pedigree-associated genetic effects and couple environmental effects. All scenarios included residuals. More details about the parameter settings for these scenarios can be found in S6 Table. We conducted the model selection procedure for each replicate to see whether the final model selected matched the simulated phenotypic components for these scenarios (Note: we ran 10 replicates for each scenario here). In addition, variance component analyses were performed using final selected models for these replicates to see whether the estimates of parameters were close to their simulated values.
|
10.1371/journal.pgen.1006369 | Mutations in TSPEAR, Encoding a Regulator of Notch Signaling, Affect Tooth and Hair Follicle Morphogenesis | Despite recent advances in our understanding of the pathogenesis of ectodermal dysplasias (EDs), the molecular basis of many of these disorders remains unknown. In the present study, we aimed at elucidating the genetic basis of a new form of ED featuring facial dysmorphism, scalp hypotrichosis and hypodontia. Using whole exome sequencing, we identified 2 frameshift and 2 missense mutations in TSPEAR segregating with the disease phenotype in 3 families. TSPEAR encodes the thrombospondin-type laminin G domain and EAR repeats (TSPEAR) protein, whose function is poorly understood. TSPEAR knock-down resulted in altered expression of genes known to be regulated by NOTCH and to be involved in murine hair and tooth development. Pathway analysis confirmed that down-regulation of TSPEAR in keratinocytes is likely to affect Notch signaling. Accordingly, using a luciferase-based reporter assay, we showed that TSPEAR knock-down is associated with decreased Notch signaling. In addition, NOTCH1 protein expression was reduced in patient scalp skin. Moreover, TSPEAR silencing in mouse hair follicle organ cultures was found to induce apoptosis in follicular epithelial cells, resulting in decreased hair bulb diameter. Collectively, these observations indicate that TSPEAR plays a critical, previously unrecognized role in human tooth and hair follicle morphogenesis through regulation of the Notch signaling pathway.
| Ectodermal dysplasias refer to a large group of inherited disorders characterized by developmental defects in tissues of ectodermal origin. The study of these conditions has been instrumental in the discovery of biological pathways involved in the regulation of epithelial tissue morphogenesis. In this report, through the delineation of the molecular basis of a novel form of autosomal recessive ectodermal dysplasia, we identified a new key player in ectodermal development. We detected a number of mutations in TSPEAR co-segregating with abnormal hair and tooth development in three families. TSPEAR encodes the thrombospondin-type laminin G domain and EAR repeats (TSPEAR) protein, whose function is poorly understood. TSPEAR was found to be strongly expressed in murine hair and tooth. Using a reporter assay, we showed that it regulates Notch activity. Accordingly, NOTCH1 expression was altered in patient skin, and NOTCH1, as well as many of its known targets, was down-regulated in TSPEAR deficient keratinocytes. Moreover, Tspear silencing in mouse hair follicle organ cultures was found to induce apoptosis in follicular epithelial cells, resulting in decreased hair bulb diameter. Collectively, these observations indicate that TSPEAR plays a critical, previously unrecognized role in human tooth and hair follicle morphogenesis through regulation of the Notch pathway. As such, these new data are likely to lead to further investigations aimed at characterizing the role of Notch signaling pathway in other forms of ectodermal dysplasias as well as acquired hair and tooth pathologies.
| Ectodermal dysplasias refer to a large clinically and genetically heterogeneous group of disorders characterized by developmental defects affecting tissues of ectodermal origin [1]. These conditions therefore feature various combinations of cutaneous, nail, hair, dental or limb anomalies which demarcate the various subtypes of ED [1]. Over the past few years, the molecular basis of many of these diseases has been deciphered, leading to the identification of a number of signaling pathways responsible for regulating ectodermal tissue ontogenesis. Among these regulatory systems, the ectodysplasin/EDAR signaling pathway, which regulates NFkappaB activity and is critically involved in murine tooth and hair development [2], is the best known and was shown to be involved in the pathogenesis of various clinical forms of hypohidrotic ectodermal dysplasia [3,4]. Additional regulatory factors which were found to be involved in the pathogenesis of ectodermal dysplasias include p63, DLX3, MSX1 and WNT proteins [5–8]. Finally, structural proteins, such as connexins and desmosomal proteins, have also been implicated in the pathogenesis of a number of ectodermal dysplasias [9,10].
Although the concomitant presence of hair and tooth abnormalities is not unusual among ectodermal dysplasias, as seen in hypohidrotic ectodermal dysplasia (MIM305100) [11], neonatal ichthyosis-sclerosing cholangitis syndrome (MIM607626) [12], p63 syndromes [13] and alopecia-neurological defects-endocrinopathy (MIM612079) [14], it is rarely seen in the absence of other ectodermal or visceral defects. In the present study, we aimed at identifying the molecular basis of a novel form of ectodermal dysplasia combining scalp hypotrichosis and hypodontia.
We studied two consanguineous families of Arab Moslem origin and one Jewish Ashkenazi family, comprising together a total of 5 patients (Fig 1A). All affected individuals displayed hypodontia (Fig 1B) as well as various degrees of scalp hypotrichosis more prominent on the anterior part of the scalp (Fig 1C). The severity of the phenotype, including the degree of alopecia and hypondontia was mild in family C patient. Patients also shared subtle dysmorphic features including a long oval face, square chin, down slanting of palpebral fissures, low insertion of columella and thick lips. We also identified in one patient (family A, IV-4) body hypertrichosis over the chest while in other patients of family A and B, body hair was missing or sparse. Follicular accentuation was most marked over bony prominences (Fig 1D). No visceral or neurological additional features were identified. Audiometry performed in family A and family C patients was normal (clinical information on all families is summarized in S1 Table). A skin biopsy obtained from patient IV-4 (Family A) scalp demonstrated paucity of mature hair follicles (Fig 1E). Scanning electron microscopy of patient hair samples showed abnormal structure of the follicular cuticle (Fig 1F).
After having excluded by direct sequencing pathogenic mutations in the coding sequences of WNT10A and TP63, which have been associated with a phenotype reminiscent of that displayed by the patients [13,15], DNA samples extracted from individuals IV-4, III-7, IV-3 and III-5 of family A and individuals II-1, I-1 and I-2 of family C, were subjected to whole exome sequencing. Data were filtered as detailed in Materials and Methods and scrutinized for mutations in any single gene common to both families.
Using this approach, we identified three mutations in TSPEAR encoding thrombospondin-type laminin G domain and EAR repeats, a member of the EAR family of proteins [16]. These proteins feature EAR domains, which are likely to mediate protein-protein interactions [16]. All affected individuals of family A were found to carry a homozygous missense sequence variation, c.1726G>T, as well as a homozygous single base pair deletion in TSPEAR, c.1728delC (Fig 2A), whereas individual II-1 of family C carried two heterozygous missense mutations: c.1852T>A and c.1915G>A (Fig 2A). Individual III-1 of family B was subsequently found by direct sequencing of TSPEAR coding sequences to carry c.1726G>T in a heterozygous state as well as to be compound heterozygous for two heterozygous deletions, c.1728delC and c.454_457delCTGG (Fig 2A).
Mutations c.1728delC and c.454_457delCTGG are both predicted to result in premature termination of protein translation (p.K577Sfs*36; p.L152Wfs*28). Mutations c.1852T>A and c.1915G>A are expected to result in two amino acid substitutions, p.Y618N and p.D639N, respectively, affecting two highly conserved residues (Conseq scores 9 and 9, respectively; range 1–9) located in two EAR domains of the protein (Fig 2B). Both c.1852T>A and c.1915G>A were foreseen to be pathogenic by two prediction software (Polyphen2 scores 1 and 1, respectively; range 0–1; SIFT scores 0 and 0.04, respectively; range 1–0). Sequence variation c.1726G>T is likely to be in linkage disequilibrium with c.1728delC. c.1726G>T is predicted to result in a single amino acid substitution (p.V576F) whose significance is unclear given conflicting results of prediction software (damaging according to PolyPhen and tolerated according to SIFT). In addition, given the predicted effect of the adjacent frameshift c.1728delC, the functional consequence of p.V576F is likely to be marginal.
Co-segregation of all four mutations with the disease phenotype was then confirmed by PCR-RFLP (Fig 1A; see experimental details in Materials and Methods). Using the same assays, mutations c.1852T>A, c.1915G>A, c.454_457delCTGG and c.1728delC were excluded from a panel of 476, 415, 294 and 309 population-matched healthy individuals respectively. We then ascertained the ESP, NCBI, UCSC, HGMD, ExAc, 1000 genomes and Ensembl databases for the presence of each of the 4 mutations. Among these, mutation c.1915G>A was present in a heterozygous state in 0.7% of a panel of control individuals (n = 35,626), suggesting that it may be associated with a common phenotype in the general population such as hypondontia whose prevalence ranges between 2% and8% [17–19]. Mutation c.454_457delCTGG was absent in all public databases while mutation c.1852T>A was present in a heterozygous state in 2 individuals out of 60,136 tested. In addition, mutation c.1728delC was present in a heterozygous state in 3 individuals out of 60,032 tested.
Given the phenotype displayed by patients carrying biallellic mutations in TSPEAR and because the role of TSPEAR in cutaneous tissues is unknown [16,20], we hypothesized that TSPEAR may be involved in the regulation of human tooth and hair follicle morphogenesis. To explore this hypothesis, we used microarray analysis to compare whole exome expression profiles of primary human keratinocytes transfected with control or TSPEAR-specific siRNA (see experimental details in Materials and Methods and S1 Fig). Pathway analysis of the data (fully available in S2 Table) revealed down-regulation of NOTCH1 as well as abnormal expression of numerous Notch signaling pathway-associated genes (Fig 3A and S3 Table). Quantitative RT-PCR was used to validate these observations (Fig 3B). Interestingly, immunostaining of a skin biopsy obtained from patient IV-4, family A, also revealed decreased NOTCH1 expression in the epidermis (Fig 3C and 3D), thus supporting the hypothesis that TSPEAR mutations exert a loss of function effect mediated through NOTCH1.
To ascertain the possibility that TSPEAR regulates ectodermal ontogenesis by modulating Notch signaling, we co-transfected HaCaT cells seeded on DLL1-coated plate with a Notch luciferase reporter construct and with a TSPEAR-specific siRNA or a control siRNA. Luciferase activity in TSPEAR down-regulated cells was significantly decreased as compared with control cells, supporting a role for TSPEAR in the regulation of Notch signaling (Fig 3E).
NOTCH1 has been associated with the regulation of dental epithelial stem cells differentiation [21] and TSPEAR was found to be expressed in the enamel organ (S2 Fig). In addition, NOTCH1 is essential for normal hair follicle postnatal development [22–26]. To investigate the role of TSPEAR in hair follicles, we obtained skin biopsies from transgenic K14/H2B/GFP mice which express green fluorescent protein (GFP) in hair follicle epithelium (see experimental details in Materials and Methods). Tspear was found to be expressed in murine hair matrix keratinocytes, outer root sheath, inner root sheath, hair shaft and the hair follicle infundibulum (Fig 4A). We then down-regulated Tspear expression using specific siRNAs. siRNA-mediated down-regulation of Tspear in mouse skin organ cultures (Fig 4B) resulted in reduced hair bulb diameter (Fig 4C–4F). This correlated with hair growth arrest as attested by decreased hair follicle pigmentation (Fig 4G–4I) (HF pigmentation is closely linked to the growth phase of the hair cycle (anagen) [27], and markedly elevated apoptotic activity both in the hair bulb (Fig 4J–4L) and infundibular (Fig 4M–4O) hair follicle compartments, as measured by the TUNEL assay. Tspear knock-down also resulted in decreased Notch1 expression in murine hair follicle organ cultures (Fig 4P).
In the present report, we studied a novel form of ectodermal dysplasia characterized by oligodontia, alopecia and facial dysmorphism and caused by mutations in TSPEAR. The physiological functions of TSPEAR are essentially unknown to date. It belongs to a family of proteins featuring EAR domains, which are predicted to form beta-propeller structures likely to mediate protein-protein interactions [16]. Some of these proteins have been found to be associated with various neurological conditions [16]. Mutation c.1728delC in TSPEAR has been reported to cause congenital sensorineural deafness in a single family and to result in inhibition of TSPEAR secretion [28]. However, this mutation was identified in the present study in a homozygous state in 2 different patients with hypotrichosis, hypodontia and normal hearing (Fig 1A). This observation coupled with the fact that we identified three other mutations in TPSEAR in additional patients with ectodermal dysplasia and normal hearing suggests the possibility that deafness in this previous single family [28] may have been due to co-inheritance of additional genetic variants. In contrast, an association study recently demonstrated the presence of genome-wide significant variations within the TSPEAR gene locus for sheep fiber diameter[29], which is in line with the reduced hair bulb diameter caused by TSPEAR down-regulation in mouse hair follicles (Fig 4F).
Although the detailed mechanism of action of TSPEAR during tooth and hair follicle morphogenesis remains to be fully delineated, our current data support the possibility that TSPEAR regulates Notch signaling, a key biological pathway previously shown to affect the development of many ectodermal tissues [23,30,31]. Although decreased Notch expression due to nicastrin mutations has been associated with perifollicular inflammation [32], overt inflammatory manifestations were not observed in patients carrying TSPEAR mutations possibly due to the fact nicastrin may affect additional targets beyond Notch and/or Notch signaling may be affected to a lesser degree by TSPEAR mutations as compared with nicastrin mutations. Supporting this possibility is the presence of follicular accentuation in our patients (Fig 1D). Interestingly, a number of known targets of NOTCH which were found to be affected by TSPEAR silencing in the current study (Fig 3B), have previously been associated with disorders featuring abnormal hair and tooth development including the oculo-dento-digital dysplasia syndrome (MIM257850), the odonto-onycho-dermal dysplasia syndrome (MIM257980), the p63 syndromes (MIM604292) and the tricho-dento-osseous syndrome (MIM190320) caused by mutations in GJA1, WNT10A, TP63 and DLX3, respectively [13,15,33–35]. Collectively, these data demarcate a group of inherited disorders sharing both phenotypic and pathophysiological features.
All affected and healthy family members or their legal guardian provided written and informed consent according to a protocol approved by our institutional review board and by the Israel National Committee for Human Genetic Studies in adherence with the Helsinki principles.
Genomic DNA was extracted from peripheral blood leukocytes using the 5 Prime ArchivePure DNA Blood Kit (5 Prime Inc., Gaithersburg, USA) or from OG-500 saliva collection kit (DNA Genotek Inc., Ottawa, Canada) according to the manufacturer's instructions.
Exome sequencing of individuals IV-3, IV-4, III-5, III-7 from family A, I-1, I-2 and II-1 from family C was performed by Otogenetics corporation using in-solution hybridization with Agilent AV5 + UTR Exome (71Mb) version 4.0 (Agilent, Santa Clara, USA) followed by massively parallel sequencing (Illumina HiSeq2000) with 100-bp paired-end reads. Reads were aligned to the Genome Reference Consortium Human Build 37 (GRCh37/hg19) using Burrows-Wheeler Aligner (BWA)[36].
Duplicate reads, resulting from PCR clonality or optical duplicates, and reads mapping to multiple locations were excluded from downstream analysis. Reads mapping to a region of known or detected insertions or deletions were re-aligned to minimize alignment errors. Single-nucleotide substitutions and small insertion deletions were identified and quality filtered using the Genome Analysis Tool Kit (GATK) [37]. Rare variants were annotated using ANNOVAR[38] and identified by filtering the data from dbSNP138, the 1000 Genomes Project, the Exome Variant Server, and an in-house database of sequenced individuals. Variants were classified by predicted protein effects using Polyphen2 [39] and SIFT [40]. S4 Table summarizes exome sequencing details.
Genomic DNA was PCR-amplified using oligonucleotide primer pairs spanning the entire coding sequence as well as intron–exon boundaries of WNT10A, TP63 and TSPEAR (S5 Table) and Taq polymerase (Qiagen, Hilden, Germany). Cycling conditions were as follows: 94°C, 2min; 94°C, 40 sec; 61°C, 40 sec; 72°C 50 sec, for 3 cycles, 94°C, 40 sec; 59°C, 40 sec; 72°C 50 sec, for 3 cycles, 94°C, 40 sec; 57°C, 40 sec; 72°C 50 sec, for 34 cycles. Gel-purified (QIAquick gel extraction kit, QIAGEN, Hilden, Germany) amplicons were subjected to bidirectional DNA sequencing with the BigDye terminator system on an ABI Prism 3100 sequencer (Applied Biosystems, NY, USA).
To screen for the c.1728delC mutation (families A and B), we PCR-amplified a 148 bp fragment with Taq polymerase (Qiagen, Hilden, Germany) and the following primers 5`- CTCCGTCATCTACGAGCTGAACGTGACCGCGCAGGCCTTTTT-3`and 5`- GATGAGCCTAACGGGGATTCC-3`. The mutation creates a recognition site for endonuclease MseI (New England Biolabs, Frankfurt, Germany). To screen for the c.454_457delCTGG mutation (family B), we PCR-amplified a 770 bp fragment, with Taq polymerase (Qiagen, Hilden, Germany) and the following primers 5`- TCTCACCACCTGTGCTCATC-3`and 5`- CACCTGTTCTCGCCAATGTC -3`. The mutation creates a recognition site for endonuclease BglI (New England Biolabs, Frankfurt, Germany). To screen for the c.1852T>A mutation (family C), we PCR-amplified a 221 bp fragment, with Taq polymerase (Qiagen, Hilden, Germany) and the following primers 5`- GTAGCTTCTGGCCAATCCCC-3`and 5`- GAAGCAAG GCTCTGGGAGG-3`. The mutation creates a recognition site for endonuclease MseI (New England Biolabs, Frankfurt, Germany). To screen for the c.1915G>A mutation (family C), we PCR-amplified a 220 bp fragment, with Taq polymerase (Qiagen, Hilden, Germany) and the following primers 5`- GGATGGAAGAGGCTCAGATG-3`and 5`- AGATGAGG TAGGCACCAGCCGTGGTGCTGAAGGCCTCCGAAT-3`. The mutation creates a recognition site for endonuclease EcoRI (New England Biolabs, Frankfurt, Germany). PCR cycling conditions were as follows: 94°C, 2min; 94°C, 40 sec; 61°C, 40 sec; 72°C 50 sec, for 3 cycles, 94°C, 40 sec; 59°C, 40 sec; 72°C 50 sec, for 3 cycles, 94°C, 40 sec; 57°C, 40 sec; 72°C 50 sec, for 34 cycles. PCR products were incubated with the appropriate enzyme at 37°C for 16 hours followed by 20 min of inactivation at 65°C. The digested PCR products were electrophoresed in ethidium bromide-stained 3% agarose gels.
For quantitative real-time PCR (qRT-PCR), cDNA was synthesized from 1000 ng of total RNA using qScript kit (Quanta Biosciences, Gaithersburg, MD). cDNA PCR amplification for TSPEAR and GAPDH (as a control) was performed with TaqMan SNP expression assays # Hs00376562_m1 and Hs02758991_g1 respectively (Applied Biosystems, Forster, CA, USA) according to the manufacturer protocol. cDNA PCR amplification for other genes was carried out with the PerfeCTa SYBR Green FastMix (Quanta Biosciences, Gaithersburg, USA) on a StepOnePlus system (Applied Biosystems, Waltham, USA) with gene-specific intron-crossing oligonucleotide pairs (S6 Table). Cycling conditions were as follows: 95°C, 20 sec and then 95°C, 3 sec; 60°C, 30 sec for 40 cycles. Each sample was analyzed in triplicates. For each set of primers, standard curves were obtained with serially diluted cDNAs. Results were normalized to GAPDH mRNA levels. qRT-PCR results were analyzed by t-test statistical analysis.
Keratinocytes (KCs) cell cultures were established from skin biopsies after written informed consent had been obtained as previously described [41]. Primary KCs were maintained in KC Growth Medium (KGM) supplemented with 0.4% bovine pituitary extract, 0.1% human epidermal growth factor (hEGF), 0.1% insulin, 0.1% hydrocortisone and 0.1% gentamicin/amphotericin B. HaCaT cells were kindly provided by Dr. Dina Ron (Technion, Haifa, Israel). The cells were maintained in MEM media supplemented with 10% fetal calf serum, 1% L-glutamine, 1% streptomycin and 1% amphotericin (Biological Industries, Beit-Haemek, Israel).
KCs were cultured in 6 well culture plates at 37°C in 5% CO2 in a humidified incubator and were harvested at 60% confluence. To down regulate TSPEAR expression, we used human TSPEAR small interference RNAs (siRNA) (Santa Cruz; sc-62060) (5`-CCUUCUCGGUGAACAGUAUtt-3`, 5`-CAUUGCCGC CACCUAUUUAtt-3`and 5`-CACUCCUGACCUUUCGUAAtt-3`). As control siRNA, we used Stealth RNAi Negative Control Duplex (Invitrogen, Carlsbad, CA). Twenty five pmol of siRNAs were transfected into KCs using Lipofectamine RNAiMax (Invitrogen). The transfection medium was replaced after 6 hours with high calcium (1.4mM)-containing KGM.
Total RNA (200 ng) was reverse transcribed and cRNA prepared using TargetAmp-Nano Labeling Kit (Epicentre Biotechnologies, Madison, WI) according to the manufacturer's protocol. One and a half μg of biotinylated cRNA was hybridized to HumanHT-12 v4 Expression BeadChip (encompassing more than 47,000 transcript targets), washed, and scanned on a BeadArray 500GX Reader using Illumina BeadScan image data acquisition software (version 2.3.0.13). Quality control and quantile normalization of the microarray data was done by BeadStudio 3.0 software (Illumina). The scanning data of the three biological repeats (total of 12 data sets) were exported to JMP genomic Software (SAS, Cary, NC), log transformed and non-expressed genes (detection p-value<0.01), transcripts with low expression (log2 value < 6.5) or with low variation across all samples (variation < 0.05) were removed from the analysis. The data was analyzed using two-way ANOVA and differently expressed genes (DEGs) were defined as transcripts that were statistically significant at corrected p-value ≤0.05 using the False Discovery Rate (FDR) with at least 0.75 delta differences. Pathway analysis to identify statistically significant functional categories in the data set was performed using Ingenuity Pathway analysis (IPA 8.0, QIAGEN Redwood City, www.qiagen.com/ingenuity).
For immunofluorescence analysis of skin biopsies, 5 μm paraffin-embedded sections were kept overnight at 37°C and de-paraffinized using xylene/ethanol. Antigen retrieval was done with 0.01M citrate buffer, pH 6.0 (Invitrogen, Carlsbad, CA) in a microwave for 25 min. Sections were blocked with 2% bovine serum albumin (BSA) in phosphate-buffered saline (PBS) for 30 min at room temperature. Primary antibodies used: rabbit anti-TSPEAR primary antibody (Abcam, Cambridge, MA, USA, 1:200 dilution); goat anti-NOTCH primary antibody (Santa Cruz, Dallas, TX, USA, 1:75 dilution). Both antibodies were diluted in 2% BSA PBS and incubated overnight at 4°C. Rhodamine Red-X goat anti rabbit IgG (H+L) (Life Technologies/Invitrogen) and Alexa Fluor 568 donkey anti goat IgG (H+L) (Thermo Fisher Scientific) were used as a secondary antibody and were diluted 1:200 with 2% BSA in PBS followed by incubation for 45 min at room temperature. Coverslips were mounted in DAPI Fluoromount-G (Southern Biotechnologies, Birmingham, AL). Negative controls consisted of slides processed similarly while omitting the primary antibody. As a positive control for TSPEAR staining, we used normal placenta tissue[42]. Specimens were examined using either a Nikon 50I microscope connected to DS-RI1 digital camera or a Zeiss LSM700 confocal microscope for fluorescence image acquisition.
HaCaT cells were seeded on hDLL1 (R&D Systems, Minneapolis, MN) coated 24 wells plate (50,000 cells/well). Twenty four hours after seeding, cells were transfected with a Notch response element-containing luciferase reporter construct, kindly obtained from Dr. David Sprinzak (Biochemistry Department, The George S. Wise Faculty of Life Sciences, Tel Aviv University) as well as a Renilla expression vector, and control siRNA (Stealth™ RNAi Negative Control Duplex Invitrogen, Carlsbad, CA) or TSPEAR specific siRNA (sc-91435; Santa Cruz Biotechnology, Santa Cruz, CA) using Lipofectamine2000 (Invitrogen, Carlsbad, CA). Forty eight hours after transfection, luciferase activity was read using a dual luciferase assay (Promega, Madison, USA). Luciferase activity was normalized to Renilla luciferase.
K14-Cre and H2B-GFP loxP mice were purchased from The Jackson Laboratory. K14-Cre mice contain a human keratin 14 promoter directing expression of Cre recombinase, while H2B-GFP mice have a fusion H2B histone with a C-terminally attached eGFP. K14-Cre and H2B-GFP loxP mice were crossed, heterozygous littermates interbred, and resulting pups genotyped. Homozygous K14 H2B-GFP+/+ mice were selected for and used as breeding pairs. All mice were kept at the University of California, San Diego (UCSD) animal facilities, and all animal experiments were approved by the UCSD Institutional Animal Care and Use Committees and were conducted in accordance with the Guideline for the Care and Use of Laboratory Animals.
Dorsal skin was isolated from K14-H2B-GFP mice and placed dermal side down into individual sterile petri dishes containing 4 ml warmed supplemented William’s E. Media (WEM) as previously described[43], with the orientation of the hair parallel to the longitudinal axis. Thin dorsal tissue strips were then sliced off, gently abrading to remove any loose hair.
Tissue strips were transfected with mouse Tspear siRNA (sc-270602; Santa Cruz Biotechnology, Santa Cruz, CA) or control siRNA (sc-36869; Santa Cruz Biotechnology). All reagents required for transfection were obtained from Santa Cruz Biotechnology (siRNA transfection reagent, sc-29528; siRNA transfection medium, sc-36868). Transfection was performed as previously described [43] and following 7 hours of transfection, tissue strips were maintained in 6-well plate with 2 ml supplemented WEM for an additional 24 hours. Following 24 hours, two tissue strips from each treatment group were used for average hair bulb measurement. In detail, two hundred μl of 2% low melt agarose solution (Agarose II; Midsci, MO, USA) dissolved in sterile Dulbecco’s phosphate buffered saline (DPBS) were used to coat the bottom of 6-well glass bottom plate (MatTek, Ashland, MA). Dorsal tissue strips were positioned parallel to the bottom of the plate, and embedded in agar. One ml of WEM was gently added to the wells. The 6-well plate was then placed in a pre-warmed incubation chamber (37°C, 5% CO2) of a Zeiss Axio Observer.Z1 microscope. Axiovision software (4.8.2 SP3) was used to select and mark multiple non-overlapping fields of view on each tissue strip covering a significant area of each strip. Fluorescent images were captured using an automatic exposure time and an excitation wavelength of 470 nm. Data were analyzed in a single blinded manner. Raw data collected were given an alpha-numeric cipher and were subsequently analyzed by a blinded investigator unaware of the conditions tested or grouping. In each hair follicle, three horizontal measurements were done in the area of hair bulb and proximal hair shaft. The average of those three measurements was calculated for each hair follicle from the two different treatment group and was defined as average hair bulb diameter. Three independent experiments were done with 3 different mice. In each mouse, 2 skin samples from each of the two treatment groups (Tspear-siRNA vs. control-siRNA) were used for average hair bulb measurement.
In addition, 3 strips from each treatment group were collected 24 hours following transfection for RNA isolation (RNeasy kit, Qiagen, Valencia, CA) and validation of Tspear silencing. For reverse transcription, we used an iScript cDNA synthesis kit (Bio-Rad, Hercules, CA) with 1 μg of RNA as starting material. The resulting cDNA was diluted 1:10 with nuclease free water and 5 μl of the diluted solution was used as template for subsequent qPCR reactions.
Quantitative real-time PCR was performed on an Applied Biosystems 7300 using TaqMan Universal PCR Master Mix (Applied Biosystems, Carlsbad, CA), template cDNA, and TaqMan primers. TaqMan primers were ordered from Life technologies; Tspear (Mm00455327_m1), and Gapdh (Mm99999915_g1) which served to normalize data (Life Technologies/Invitrogen, Grand Island, NY). To quantify Notch1 expression, we used 2x SYBR Green qPCR Master Mix (Biotool, Houston,TX) and Primer bank ID 13177625a1 Notch1 primers while Rplp0 served as an internal control (Forward primer 5’-GAGATTCGGGATATGCTGTTGG-3’ and Reverse primer 5’-CGGGTCCTAGACCAGTGTTCT-3’).
Finally, tissue strips from each treatment group were frozen in liquid nitrogen 48 hours after transfection for immunohistochemistry studies. Seven μm-thick cryosections were prepared and stored at −80°C until use. For Tspear protein expression, cryosections were first air-dried for 10 min and then fixed in acetone at -20°C for another 10 min. After air drying, the slides were washed three times for 5 min in PBS. Following 20 min incubation of cryosections with 2% goat serum in PBS, cryosections were incubated overnight at 4°C with the rabbit anti-Tspear monoclonal antibody (Abcam, Cambridge, MA) at 1:50 dilution with 2% goat serum in PBS solution. This was followed by incubation with Alexa Fluor® 568 Goat Anti-Rabbit secondary Antibody (Life Technologies/Invitrogen) for 45 min at RT in 1:200 dilution with 2% goat serum in PBS solution. Incubation steps were interspersed with three washes, 5 min each, with PBS. Then sections were embedded and counterstained with DAPI for the identification of cell nuclei.
For Masson-Fontana histochemistry, cryosections were air dried and fixed in ethanol-acetic acid. The sections were washed in Tris-buffered saline (TBS) and distilled water several times. Cryosections were treated with ammoniacal silver solution (Thermo Fisher Scientific, Carlsbad, CA) for 40 min at 56°C in the dark. After washing in distilled water, the sections were treated with 5% aqueous sodium thiosulphate (Sigma-Aldrich, St. Louis, MO) for 1 min. Then, the sections were washed in running tap water for 3 min and were counterstained with haematoxylin for 45 seconds. After washing in distilled water, sections were dehydrated and mounted in Eukitt (Sigma-Aldrich, St. Louis, MO). Tspear immunoreactivity and melanin content by Masson-Fontana histochemistry were compared between test and control sections by quantitative histomorphometry as previously described [43,44] using NIH IMAGE software (NIH, Bethesda, MD, USA).
For apoptosis detection, a kit for DeadEnd Fluorometric TUNEL system analysis (Promega, Madison, WI, USA) was used. Briefly, cryosections were air dried and fixed in 4% formaldehyde in PBS for 25 minutes. Following several washing steps in PBS, permeabilization step with 0.2% Triton in PBS was conducted for 5 min. Following several washing steps in PBS, sections were equilibrated with equilibration buffer for 5–10 min and then incubated with TdT reaction mix for 60 minutes at 37°C. Following stop reaction step and washing steps with PBS, sections were embedded and counterstained with DAPI for the identification of cell nuclei.
The URLs for data presented herein are as follows:
1000 genomes project, http://www.1000genomes.org/
ConSurf, http://consurftest.tau.ac.il/
dbSNP, http://www.ncbi.nlm.nih.gov/SNP/
Exome Variant Server (http://evs.gs.washington.edu/EVS/)
GenBank, http://www.ncbi.nlm.nih.gov/Genbank/
NHLBI Grand Opportunity Exome Sequencing Project, https://esp.gs.washington.edu/drupal/
Online Mendelian Inheritance in Man (OMIM), http://www.omim.org
PolyPhen-2, http://genetics.bwh.harvard.edu/pph2/
SIFT, http://sift.jcvi.org/
UCSC Genome Browser, http://genome.ucsc.edu/
Human Gene Mutation Database, http://www.hgmd.cf.ac.uk/ac/index.php
ANNOVAR, http://annovar.openbioinformatics.org/en/latest/
Burrows-Wheeler Aligner, http://bio-bwa.sourceforge.net/
GATK, https://www.broadinstitute.org/gatk/
Ingenuity Pathway Analysis (IPA), www.qiagen.com/ingenuity
|
10.1371/journal.pgen.1006650 | SRC-2-mediated coactivation of anti-tumorigenic target genes suppresses MYC-induced liver cancer | Hepatocellular carcinoma (HCC) is the fifth most common solid tumor in the world and the third leading cause of cancer-associated deaths. A Sleeping Beauty-mediated transposon mutagenesis screen previously identified mutations that cooperate with MYC to accelerate liver tumorigenesis. This revealed a tumor suppressor role for Steroid Receptor Coactivator 2/Nuclear Receptor Coactivator 2 (Src-2/Ncoa2) in liver cancer. In contrast, SRC-2 promotes survival and metastasis in prostate cancer cells, suggesting a tissue-specific and context-dependent role for SRC-2 in tumorigenesis. To determine if genetic loss of SRC-2 is sufficient to accelerate MYC-mediated liver tumorigenesis, we bred Src-2-/- mice with a MYC-induced liver tumor model and observed a significant increase in liver tumor burden. RNA sequencing of liver tumors and in vivo chromatin immunoprecipitation assays revealed a set of direct target genes that are bound by SRC-2 and exhibit downregulated expression in Src-2-/- liver tumors. We demonstrate that activation of SHP (Small Heterodimer Partner), DKK4 (Dickkopf-4), and CADM4 (Cell Adhesion Molecule 4) by SRC-2 suppresses tumorigenesis in vitro and in vivo. These studies suggest that SRC-2 may exhibit oncogenic or tumor suppressor activity depending on the target genes and nuclear receptors that are expressed in distinct tissues and illuminate the mechanisms of tumor suppression by SRC-2 in liver.
| Liver cancer is the third leading cause of cancer-associated deaths worldwide with limited responses to targeted therapies. An unbiased forward genetic screen previously revealed a tumor suppressor role for the Steroid Receptor Coactivator 2 (Src-2) in liver cancer driven by the MYC oncogene. Yet, SRC-2 has been shown to promote survival and metastasis in prostate cancer cells, suggesting a tissue-specific and context-dependent role for SRC-2 in tumorigenesis. Through the use of mice lacking SRC-2, we provide unequivocal evidence that this protein restrains MYC-induced liver tumorigenesis, and we have begun to identify key downstream SRC-2 target genes that mediate this effect. This work provides important new insights into the mechanism of tumor suppression by SRC-2 in MYC-induced liver cancer. Our study also suggests that SRC-2 may exhibit oncogenic or tumor suppressor activity depending on the target genes and nuclear receptors that are expressed in distinct tissues.
| Hepatocellular carcinoma (HCC) is the fifth most common solid tumor and the third leading cause of cancer-related deaths, resulting in approximately 700,000 deaths per year worldwide [1]. Liver tumorigenesis occurs in settings of chronic inflammation, cirrhosis, or glycogen storage disease [2, 3]. Previous studies have described genomic alterations in human HCC, with recurrent loss of the TP53 and RB tumor suppressor genes, and amplification or overexpression of the MYC oncogene in 40–60% of HCCs [4–6]. Despite this wealth of data, the critical genes and pathways that contribute to HCC development are incompletely understood. A better understanding of the mechanisms underlying HCC initiation and progression may accelerate the development of novel therapeutic strategies.
Complementary to large-scale genome sequencing studies, forward genetic mutagenesis screens in mice provide an unbiased approach to study the significance of gene mutations in tumorigenesis [7–12]. Previously, we utilized the Sleeping Beauty (SB) DNA transposon system to identify mutations that cooperate with MYC to accelerate liver tumorigenesis in mice. This led to the identification of Steroid Receptor Coactivator 2 (SRC-2, also known as NCOA2, TIF2, GRIP1) as a novel gene that functions to restrain MYC-induced liver cancer [13]. SRC-2 encodes a potent transcriptional coactivator that cooperates with nuclear receptors (NRs) to control multiple physiological processes including glucose homeostasis, energy metabolism, and reproduction [14–22]. Mice with whole-body or liver-specific deletion of Src-2 develop glycogen storage disease Type 1 (Von Gierke’s disease), and exhibit decreased expression of the SRC-2 target Glucose 6 phosphatase (G6pc) [15]. Moreover, a significant fraction of patients with Von Gierke’s disease develop hepatic adenomas and are susceptible to developing HCC [23]. Several lines of evidence from our previous study supported a cell-autonomous tumor suppressor role for SRC-2 in liver tumorigenesis [13]. First, recurrent transposon insertions in SB-induced liver tumors resulted in decreased mRNA expression of Src-2 and one of its characterized targets G6pc. Second, inhibition of Src-2 using shRNAs promoted tumor formation by mouse hepatoblasts in immunocompromised mice. Third, deletion of Src-2 predisposed mice to diethylnitrosamine (DEN)-induced liver tumorigenesis. Finally, we observed decreased expression of SRC-2 (NCOA2) in human HCC samples. Consistent with these findings, depletion of SRC-2 in human breast cancer cells stimulated cell proliferation by modulating estrogen-regulated genes [24]. Nevertheless, multiple observations suggest that further functional studies of SRC-2 are needed to establish whether this protein is a bona fide tumor suppressor in liver cancer. For example, copy number gains of SRC-2 are frequent in liver cancer [25, 26], although this is likely due to the proximity of this gene to the MYC gene on chromosome 8q. Furthermore, a recent study demonstrated that SRC-2 promotes lipogenesis and enhanced cell survival and metastasis in prostate cancer [27], suggesting a tissue-specific and context-dependent role for SRC-2 in tumorigenesis.
To definitively test the tumor suppressor activity of SRC-2 in MYC-mediated liver tumorigenesis in vivo and to further investigate the mechanism(s) through which this coactivator inhibits liver tumorigenesis, we examined the consequences of genetic deletion of Src-2 in a MYC-induced liver cancer model. Indeed, liver tumor burden was significantly increased in Src-2-/- mice. RNA sequencing (RNAseq) and in vivo chromatin immunoprecipitation assays revealed a set of direct SRC-2 target genes in liver. Inhibition of SRC-2 or select SRC-2 target genes accelerated proliferation of human liver cancer cells in vitro and tumorigenesis in vivo, while overexpression of SRC-2 targets, or SRC-2 itself, resulted in tumor suppressive effects. These findings provide important new insights into the mechanism of tumor suppression by SRC-2 in MYC-induced liver cancer.
To determine whether SRC-2 suppresses MYC-mediated liver cancer, we employed a mouse model of MYC-induced liver cancer previously utilized in a SB mutagenesis screen [28]. Mice harboring a MYC transgene under the control of a doxycycline-regulatable promoter (tet-o-MYC) were crossed with mice expressing tet-transactivator protein (tTA) driven by the liver-activator protein (LAP) promoter. Removal of doxycycline leads to MYC induction in the liver and development of tumors that resemble human hepatocellular cancer. We bred this model to Src-2+/- mice and generated tet-o-MYC; LAPtTA animals harboring wild type, heterozygous, or homozygous null alleles of Src-2 (S1A Fig) [29]. Loss of SRC-2 was confirmed by western blotting with tumor lysates from Src-2+/+ and Src-2-/- animals (S1B Fig). Doxycycline was withdrawn at 6 weeks, and mice were monitored for early-developing tumors (Fig 1A). All animals were euthanized and dissected at 15 weeks of age (9 weeks after MYC induction). Histologic analysis confirmed that tumors arising in these animals resembled human hepatocellular cancer (Fig 1B) and, consistent with prior reports, Src-2-/- mice exhibited an accumulation of glycogen and lipid droplets in non-neoplastic hepatocytes and in liver tumors (S2 Fig) [15]. Notably, Src-2-/- mice exhibited a significant enhancement of liver tumor burden compared to Src-2+/+ animals (Fig 1C and 1D, p<0.0295). Therefore, genetic inactivation of Src-2 is sufficient to accelerate MYC-mediated liver tumorigenesis.
To investigate the mechanisms through which SRC-2 suppresses liver tumorigenesis, we used RNA-Seq to assess global gene expression in liver tumor nodules from Src-2+/+ and Src-2-/- animals. We identified 865 differentially expressed genes between wild type and knockout tumors. DAVID Gene Ontology analysis identified biological processes enriched in Src-2-/- liver tumors (Fig 2A and 2C). Downregulated genes included regulators of fatty acid and glucose metabolism, and cell adhesion. Upregulated genes included mediators of growth factor signaling and inflammation. Key genes from each of these categories were validated using quantitative real-time PCR (qRT-PCR) (Fig 2B and 2D). Thus, Src-2 may function to restrain HCC by regulating multiple biological pathways relevant to tumorigenesis.
To distinguish direct versus indirect SRC-2 target genes, we overlapped the list of genes that were downregulated in Src-2-/- liver tumors with genes that were bound by SRC-2 in genome-wide chromatin immunoprecipitation (ChIP) Seq analysis of murine liver [17] (Fig 2E). We identified 47 genes that were bound by SRC-2 and downregulated in Src-2-/- liver tumors (S1 Table). To identify clinically relevant candidate genes, we used data from a previously described gene expression analysis of human liver tumors and paired adjacent normal tissue [30] to assess expression of 23 of these genes that were downregulated by at least 2-fold in Src-2-/- liver tumors and were expressed in the human dataset. Of these, 19/23 genes were downregulated in human HCC samples (S3 Fig). We selected four putative downstream targets of SRC-2 for further study: Small Heterodimer Partner (Shp), Dickopff 4 (Dkk4), Cell Adhesion Molecule 4 (Cadm4), and Thyroid hormone responsive (Thrsp). These genes were selected because they were downregulated in Src-2-/- tumors (our RNA-Seq analysis) and in human HCCs, they harbored mutations in human cancers (S2 Table, S3 Table), and they were directly bound by SRC-2. Indeed, we confirmed using qRT-PCR that expression of three out of four of these genes (Shp, Dkk4, and Cadm4) was significantly downregulated in an independent set of Src-2-/- liver tumors (S4 Fig), and identified SRC-2 ChIP-seq peaks in the proximal promoter and/or enhancer regions of each gene (Fig 2F). Although Thrsp was not significantly downregulated in the independent tumors, it was downregulated in a cohort of 91 HCC tumors relative to paired normal adjacent tissue [30] and we therefore included it in selected functional studies.
In our ChIP-Seq analysis, we also found that SRC-2 bound to the proximal promoter of Vegfc, Fgf1, and Masp1, and that mRNA expression was upregulated in Src-2-/- liver tumors (S5 Fig). Vegfc and Fgf1 encode growth factors that promote cell growth and survival [31, 32]. Masp1 is a key component of the complement cascade, which has also been implicated in promoting tumorigenesis [33, 34]. Although activation of gene targets is thought to serve as the primary function of this nuclear receptor coactivator, SRC-2 was previously reported to cooperate with NRs including Glucocorticoid Receptor and Estrogen Receptor to mediate transcriptional repression [35, 36]. Therefore, we speculate that SRC-2 might also repress downstream target genes that promote growth and proliferation. Future studies are warranted to assess SRC-2-mediated gene repression in the context of liver tumorigenesis.
To functionally validate SRC-2 target genes as putative tumor suppressors, we next performed loss-of-function experiments in human HCC cells. HepG2 and Huh7 were chosen for these studies since these cell lines are widely used for functional analysis of genes in HCC and they express MYC at levels comparable to liver tumors in Src-2-/-; tet-o-MYC; LAPtTA mice (S6 Fig). DKK4 and CADM4 were expressed at high levels in HepG2 cells, and SHP was highly expressed in Huh7 cells, allowing examination of the consequences of their inhibition in either of these cell lines. THRSP was not expressed in Huh7 or HepG2 cells, precluding analysis of THRSP loss of function in these cells.
SHP encodes an orphan nuclear receptor that lacks a conserved DNA binding domain and physically interacts with nuclear receptors and transcriptional factors to facilitate transcriptional repression [37, 38]. In the liver, SHP transcriptionally represses CYP7A1 to regulate bile acid biosynthesis. Loss of Shp in mice results in abnormal accumulation of bile acids and liver tumor development [39, 40]. To determine whether SHP inhibition promotes proliferation and tumorigenesis in human cells, we utilized shRNAs to suppress SHP in Huh7 cells. qRT-PCR confirmed inhibition of SHP mRNA using two independent shRNAs (Fig 3A). Cells with stable inhibition of SHP grew significantly faster than control cells (Fig 3B). Moreover, SHP depletion accelerated tumor formation of Huh7 cells in immunocompromised mice (Fig 3C). Although we detected an increase in Cyp7a1 in Src-/- tumors (S7A Fig), CYP7A1 was not expressed in human HCC cells. Taken together, our data provide evidence that SHP is a downstream target of SRC-2 that inhibits liver tumorigenesis.
A previous study demonstrated that SHP suppressed proliferation by transcriptionally repressing Cyclin D1 (Ccnd1) expression and that Shp-/- liver tumors exhibited increased Ccnd1 expression [40]. However, it was also reported that CCND1 levels were unaffected in livers of mice overexpressing SHP [41]. Notably, we failed to observe a significant change in CCND1 mRNA or protein in Huh7 cells after SHP knockdown (S7B and S7C Fig). Similarly, we failed to detect a difference in CCND1 mRNA in Huh7 xenograft tumors lacking SHP (S7D Fig). These findings suggest that in addition to its known effects on bile acid homeostasis, SHP suppresses liver tumorigenesis by regulating tumor cell proliferation through a mechanism that is independent of CCND1.
We next utilized shRNAs to inhibit expression of DKK4 and CADM4 in HepG2 cells. DKK4 belongs to the Dickopff (DKK) family of secreted glycoproteins and negatively regulates Wnt signaling [42]. qRT-PCR and western blotting confirmed a reduction in mRNA and protein, respectively (Fig 3D). DKK4 shRNA-1 and shRNA-3 cells grew significantly faster than control cells (Fig 3E). Moreover, depletion of DKK4 enhanced tumorigenesis in vivo (Fig 3F). Similarly, inhibition of CADM4, which encodes a cell adhesion molecule that belongs to the immunoglobulin superfamily [43], significantly increased cell proliferation and tumorigenesis in vivo (Fig 3G–3I). Thus, multiple SRC-2 target genes, including SHP, CADM4, and DKK4, exhibit tumor suppressor activity in human HCC cells.
We next determined whether SRC-2 overexpression is sufficient to suppress tumorigenesis in human liver cancer cells. Huh7 cells were infected with an SRC-2-expressing or an eGFP control lentivirus, and overexpression of SRC-2 was confirmed by quantitative RT-PCR and western blotting (Fig 4A and 4B). Upregulation of SRC-2 and its target SHP (Fig 4B) were associated with a concomitant decrease in cell proliferation and tumorigenesis in immunocompromised mice (Fig 4C and 4D). Thus, as in mouse, SRC-2 restrains liver tumorigenesis in human HCC cells. Notably, although DKK4 transcript levels increased by 4-fold upon SRC-2 overexpression, DKK4 protein levels were only modestly affected, suggesting the existence of post-transcriptional mechanisms that control DKK4 expression independently of SRC-2 in these cells.
To validate the ability of SRC-2 targets to suppress tumorigenesis, we next overexpressed individual target genes in human HCC cells using lentivirus and assessed tumor development in vivo. Complementary to the loss-of-function experiments (Fig 3), SHP, DKK4, or CADM4 overexpression significantly reduced tumor formation in immunocompromised mice (Fig 5A–5C and 5E–5G). We also observed reduced tumor formation upon enforced expression of THRSP (Fig 5D and 5H), which encodes an acidic protein that responds robustly to thyroid hormone stimulus [44] that has not been previously linked to liver cancer.
We previously demonstrated that inhibition of Src-2 promoted tumor formation of murine hepatoblasts in immunocompromised mice [13]. To expand these findings to human liver cancer cells, we performed SRC-2 loss of function studies in HepG2 and Huh7 cells. As expected, SRC-2 inhibition in HepG2 cells resulted in decreased expression of SRC-2 and its targets SHP and CADM4, and significantly increased cell proliferation and tumorigenesis in vivo (Fig 6A–6D). Similarly, inhibition of SRC-2 in Huh 7 cells resulted in decreased SHP and DKK4 expression, and a concomitant increase in cell proliferation (S8A and S8B Fig). We next sought to determine whether any of the SRC-2 targets alone or in combination were sufficient to rescue the enhanced cell proliferation and tumor burden resulting from SRC-2 knockdown. Rescue experiments were performed in HepG2 cells because three of the four putative SRC-2 target genes (SHP, CADM4, and DKK4) were expressed in these cells. Indeed, enforced expression of SHP, CADM4, DKK4, and THSRP in combination significantly reduced proliferation and tumor burden (Fig 6C and 6D, S9A Fig). Moreover, individual overexpression of CADM4 and SHP were sufficient to suppress the increase in cell proliferation and tumorigenesis of SRC-2 knockdown cells (Fig 6C and 6D, S9A Fig). In contrast, overexpression of either DKK4 or THRSP alone significantly impacted rates of cell proliferation but not tumor burden (S9A–S9C Fig). These data provide convincing evidence that SHP and CADM4 function as important anti-tumorigenic SRC-2 target genes in human liver cancer cells. Our data also suggest that DKK4 and THRSP may not be targets of SRC-2 in HepG2 cells, and thus may be dysregulated in liver cancer cells through additional SRC-2-independent mechanisms.
Finally, we sought to identify the putative nuclear receptors that cooperate with SRC-2 to activate transcription of target gene expression and suppress proliferation and tumorigenesis. We screened the promoter regions of DKK4, THRSP, CADM4, and SHP for nuclear receptor binding motifs using NHRscan, a computational predictor of nuclear hormone receptor binding sites [45]. We then assessed whether the NR binding motifs overlapped with SRC-2 ChIP-Seq peaks identified in this study. This analysis revealed that the promoter regions of Dkk4 and Thrsp both contained Thyroid Receptor (TR) binding motifs, denoted as Everted Repeat 6 (ER6) (S10A and S10B Fig) [46]. Recently, ChIP-Seq analysis identified a TR peak upstream of the Thrsp promoter, although TR binding to Dkk4 was not verified in this study [47]. However, TR is known to inhibit liver tumorigenesis through transcriptional activation of DKK4 [48]. NHRscan analysis also uncovered several Direct Repeats (DR) that overlap with SRC-2 binding regions upstream of CADM4 (S10C Fig). Previous studies showed that RAR heterodimerizes with RXRA and preferentially binds to DR-rich regions in the genome [49, 50]. Finally, NR binding motif analysis revealed that the SHP promoter harbors HNF4A and FXR binding motifs overlapping with SRC-2 ChIP-Seq peaks (S10D Fig). Indeed, genome-wide ChIP-Seq analysis in mouse liver identified a FXR peak that overlaps with the SRC-2 ChIP Seq peak [51].
To determine whether SRC-2 cooperates with FXR in activating SHP expression in human liver cancer cells, we performed transactivation assays with a luciferase reporter construct harboring the proximal promoter of SHP, and a truncated reporter construct harboring a deletion that encompasses the FXR binding site. FXR was previously shown to activate the human SHP (NR0B2) promoter [52]. FXR was expressed in Huh7 cells infected with an eGFP control or SRC-2 lentivirus. Overexpression of SRC-2 and FXR increased SHP reporter activity by approximately 9-fold in Huh7 cells compared to cells expressing FXR alone (S11A–S11C Fig). Interestingly, while the truncated reporter construct was significantly less active, it was also measurably stimulated by SRC-2 expression. These findings suggest that SRC-2 can interact with other factors that transactivate the SHP promoter. These data provide additional evidence that SRC-2 directly induces SHP expression. Future studies are warranted to dissect additional SRC-nuclear receptor interactions in liver cancer and in different tumorigenic contexts.
Recently, large-scale studies have identified multiple types of recurrent genomic alterations of SRC-2 in human HCC, including missense mutations and amplifications [25, 26]. Notably, SRC-2 and MYC are both located on the short arm of chromosome 8. MYC is amplified in 40–60% of human HCCs and a number of studies have previously documented 8q gains in a significant fraction of liver cancers [53–55]. Thus, it is possible that SRC-2 copy number gains may occur simply due to a passenger effect associated with MYC amplification and may not functionally contribute to tumorigenesis. In support of this concept, Kaplan-Meier analysis revealed that survival of HCC patients with SRC-2 amplification or mRNA upregulation was not significantly different than survival of patients lacking these alterations (S12 Fig). In contrast, we previously showed that low expression of SRC-2 in tumors is strongly associated with poor survival in HCC patients [13, 56] and HCC patients harboring SRC-2 missense mutations similarly exhibit poorer overall survival (S12 Fig). Taken together, these studies point to a tumor suppressor role for SRC-2 in HCC. Nevertheless, in light of recent evidence indicating that SRC-2 has oncogenic activity in prostate cancer [27], a direct demonstration of the tumor suppressor activity of SRC-2 in liver cancer, and a better understanding of the underlying mechanisms, would provide important insight into the role of SRC-2 in HCC. Through the use of Src-2-/- mice, we have now provided unequivocal evidence that this protein restrains MYC-mediated liver tumorigenesis in vivo and we have begun to identify key downstream SRC-2 target genes that mediate this effect.
The orphan nuclear receptor SHP represents one such direct SRC-2 target gene with strong anti-tumorigenic activity. SHP has been extensively studied for its role in liver bile acid homeostasis and as a transcriptional repressor of other NRs. Mice lacking Shp accumulate bile acids due to de-repression of the SHP target Cyp7a1 and develop HCC [39, 40]. SHP is also downregulated in liver cancer and low expression of SHP is associated with poor survival of HCC patients [57]. Accordingly, our data demonstrate that SHP inhibition accelerates tumor formation by human HCC cells in mice. Although we detected an increase in Cyp7a1 in Src-/- tumors, we did not detect expression of CYP7A1 in human HCC cells, nor did we detect a difference in expression of another putative SHP target, Ccnd1 (S7 Fig). These data suggest that SHP represses hepatic tumorigenesis through mechanisms that are independent of these genes. Importantly, overexpression of SHP alone was sufficient to reverse the tumor enhancing effect of SRC-2 knockdown in HepG2 cells (Fig 6C and 6D, S9A Fig). In light of these findings, future studies are warranted further characterize SHP targets that control proliferation and metabolism in liver cancer and other tumor types. It will also be worthwhile to investigate whether treatment with NR agonists that are known to induce SHP expression may inhibit liver tumorigenesis. These studies may impact our understanding and treatment of additional types of cancers as SHP was recently found to be downregulated in lung tumors and low expression was associated with poor survival of stage I non-small cell lung cancer patients [58].
DKK4 was also identified as a novel anti-tumorigenic SRC-2 target gene in this study. DKK4 encodes a secreted glycoprotein that competes with Wnt ligand binding to LRP5/6 to attenuate canonical Wnt signaling [59]. Dysregulation of the Wnt pathway is a key molecular lesion in liver cancer. More than 60% of liver tumors exhibit an accumulation of β–catenin, a hallmark of activated Wnt signaling. Recent findings demonstrated that DKK4 overexpression suppressed migration, invasion, and tumor formation of human hepatoma cells in mice [42, 48]. Consistent with these data, our findings revealed that DKK4 suppresses tumorigenesis of human HCC cells in vivo, whereas shRNA-mediated inhibition of DKK4 accelerated tumorigenesis (Fig 3D–3F, Fig 5B and 5F). Collectively, these data suggest that DKK4 may be an important downstream component of the SRC-2-regulated gene expression network that inhibits liver tumorigenesis and uncovers functional antagonism between SRC-2 and the Wnt signaling pathway. However, it is important to note that DKK4 mRNA and protein levels did not correlate in the SRC-2 gain-of-function (Fig 4) and loss of function studies (Fig 6), and that overexpression of DKK4 alone was insufficient to rescue the tumor burden of SRC-2 knockdown in HepG2 cells (S9 Fig). Thus, there are likely additional mechanisms independent of SRC-2 that control DKK4 expression in liver cancer cells.
We also demonstrated that two additional genes without a prior known role in liver cancer, CADM4 and THRSP, have strong anti-tumorigenic activity in this tumor type. Consistent with these results, expression of CADM4, which encodes a member of the immunoglobulin superfamily of proteins, is reduced in multiple tumor types and suppresses tumor formation of prostate, renal and colon cancer cells in immunocompromised mice [60, 61]. Moreover, overexpression of CADM4 was sufficient to reverse tumor acceleration by SRC-2 knockdown in HepG2 cells (Fig 6C and 6D, S9A Fig). THRSP encodes a key modulator of lipogenesis and is expressed in lipogenic tissues such as liver, breast, and adipose tissue [62]. Although Thrsp did not exhibit consistent downregulation in Src-/- tumors, a previous gene expression analysis [30] revealed that THRSP was significantly downregulated in a cohort of 91 HCC tumors relative to paired normal adjacent tissue (S3 Fig). Moreover, THRSP inhibited growth and induced cell death of human breast cancer cells [63]. Although THRSP was not expressed in either of the human liver cancer cell lines we tested, enforced expression of THRSP significantly reduced tumor burden of Huh7 cells in vivo (Fig 5D and 5H). These findings set the stage for further study of the roles of CADM4 and THRSP in HCC pathogenesis.
In summary, these results firmly establish the potent anti-tumorigenic activity of SRC-2 in human and mouse liver cancer and begin to dissect the SRC-2-regulated gene expression network that mediates these effects. Furthermore, these studies provide insight into the molecular mechanisms through which this transcriptional coactivator may limit tumorigenesis in some tissues and promote oncogenesis in others. In the prostate, SRC-2 amplification coactivates androgen receptor-mediated gene transcription to promote prostate lipogenesis, tumor progression, and metastasis [27]. In liver, SRC-2 cooperates with multiple nuclear receptors, several of which are documented tumor suppressors, including Thyroid Receptor (TR), Estrogen Receptor (ER), Hepatocyte Nuclear Factor 4 alpha (HNF4A), Retinoid X Receptor alpha (RXRA), Farnesoid X Receptor (FXR), and Retinoic Acid Receptor alpha (RARA) [15, 17–20, 24, 64, 65] to coactivate a distinct program of target genes resulting in tumor suppression. Recently, a small molecule that stimulated SRC transcriptional activity was developed and shown to promote cell death in breast cancer cells [66]. Determining whether small molecule-mediated activation of SRC-2 can attenuate liver tumorigenesis represents an exciting area for future investigation.
Mice were monitored closely throughout all experimental protocols to minimize discomfort, distress, or pain. Signs of pain and distress include disheveled fur, decreased feeding, significant weight loss (>20% body mass), limited movement, or abnormal gait. If any of these signs were detected, the animal was removed from the study immediately and euthanized. All sacrificed animals were euthanized with CO2. The animals were placed in a clear chamber and 100% CO2 was introduced. Animals were left in the container until clinical death ensured. To ensure death prior to disposal, cervical dislocation was performed while the animal was still under CO2 narcosis. All methods were performed in accordance with the recommendations of the Panel on Euthanasia of the American Veterinary Medical Association and protocols approved by the UT Southwestern Institutional Animal Care and Use Committee (protocol # 2011–0119).
HepG2 and Huh7 cells were cultured in Dulbecco’s Modified Eagle Medium (GIBCO) supplemented with 10% FBS (Invitrogen) and 1% Penicillin/Streptomycin (Invitrogen). Huh7 and HepG2 cells were a gift from Hao Zhu (UT Southwestern Medical Center).
The Institutional Animal Care and Use Committeee (IACUC) of UT Southwestern Medical Center approved all procedures involving mice. Src-2-/-mice were obtained from Pierre Chambon and maintained on a mixed C57BL/6J and 129sV background [29]. LAPtTA and tet-O-MYC mice were obtained from Dean Felsher and maintained on a FVB/NJ background [28]. Simultaneously, Src-2+/- mice were bred with tet-o-MYC and LAPtTA mice to generate Src-2+/-; tet-o-MYC and Src-2+/-; LAPtTA mice, respectively. In the final cross, Src-2+/-; LAPtTA females were bred with Src-2+/-; tet-o-MYC males to obtain tet-o-MYC; LAPtTA mice with all 3 alleles of Src-2 (WT, heterozygous, or homozygous null). The MYC transgene is on chromosome Y, precluding analysis of females.
The following plasmids were used: TRC shRNA for SHP (UT Southwestern core facility, Jerry Shay laboratory V2LHS_239330, V2LHS_72556); TRC shRNA constructs for DKK4 (GE Dharmacon RHS4430-200191360 V2LHS_197942 RHS4430-200173366—V2LHS_204025); GIPZ shRNA for CADM4 (GE Dharmacon V3LHS_375253, V3LHS_375254); GIPZ shRNA for SRC-2 (GE Dharmacon V2LHS_199063, V2LHS_357381). pLJM1-EGFP was a gift from David Sabatini (Addgene plasmid # 19319), and pLX304, which also harbors a V5 tag, was a gift from David Root (Addgene plasmid # 25890). A human SHP plasmid was a gift from Steven Kliewer (UT Southwestern Medical Center); pCMX-FXR and hSHP-LUC plasmids were a gift from David Mangelsdorf (UT Southwestern Medical Center). A SHPΔ215-569-LUC deletion mutant construct lacking the FXR response element was generated by PCR amplification as previously described [52]. The pHRL-SV40 Renilla reporter plasmid was a gift from Joshua Mendell (UT Southwestern Medical Center).
Whole liver was dissected from euthanized mice, washed, and placed in ice-cold PBS. At the time of dissection, we captured images of both the dorsal and ventral sides of the intact liver, and estimated the mean percent tumor burden for each mouse using NIH Image J software. We measured the surface area of the liver tumors and the total surface area (including normal liver and all tumors). For percent tumor burden calculation, we divided the surface area of the liver tumors by the total surface area (including normal liver and tumors) and then multiplied by 100. For histological analysis, tissues were fixed in 10% formalin, embedded in paraffin, and sectioned. Hematoxylin and eosin (H &E) and Periodic acid-Schiff (PAS) staining were performed on normal liver and liver tumor tissues at the Pathology Core, UT Southwestern Medical Center.
Total RNA was isolated from liver tumors and normal tissues using Trizol (Invitrogen) followed by additional cleanup and DNase digestion using the RNeasy Mini Kit (Qiagen). Total RNA was isolated from cells using only the RNeasy Mini Kit (Qiagen). For qRT-PCR of mRNA, cDNA synthesis was performed with 1 μg RNA for reverse transcription using Superscript III First Strand synthesis kit (Invitrogen). mRNA expression was assessed using quantitative real-time PCR with a 2X SYBR Green Master Mix (R&D Systems). mRNA levels were normalized to β-actin mRNA expression, with gene expression levels measured using a standard curve for each set of primers crossing exon-exon junctions for each gene. All PCR assays were performed in triplicate. PCR primers are shown in S4 Table.
RNA sequencing was performed in the McDermott Center Sequencing Core at UTSW Medical Center. RNA was extracted from tet-o-MYC; LAPtTA; Src-2+/+ and tet-o-MYC; LAPtTA; Src-2-/- liver tumors. Four μg of total DNAse treated RNA was prepared with the TruSeq Stranded Total RNA LT Sample Prep Kit from Illumina. Poly-A RNA was purified and fragmented before strand specific cDNA synthesis. cDNA was A-tailed and indexed adapters were ligated. Samples were PCR amplified and purified with AmpureXP beads and validated on the Agilent 2100 Bioanalyzer. Samples were quantified by Qubit (Invitrogen) prior to normalization and pooling. Sequencing was performed on an Illumina Hiseq 2500 to generate 51-bp single-end reads. Reads were trimmed to remove low-quality regions in the ends. Trimmed reads were mapped to the mouse genome (mm10) using TopHat v2.0.1227 guided by iGenomes GTF file (https://ccb.jhu.edu/software/tophat/igenomes.shtml). Alignments with mapping quality less than 10 were discarded. Expression abundance estimation and differential expression analysis were carried out using Cufflinks/Cuffdiff (v2.1.1) software. Genes with the nominal p-value cutoff of 0.05 were considered significantly differentially expressed between the tet-o-MYC; LAPtTA; Src-2+/+ and tet-o-MYC; LAPtTA; Src-2-/- liver tumors if the genes were also downregulated in human HCCs, harbored mutations in human cancers, and were directly bound by SRC-2 (based on ChIP-Seq data in mouse liver).
Gene Ontology analysis was performed using the DAVID Functional Annotation tool (http://david.abcc.ncifcrf.gov/) on differentially expressed genes between the tet-o-MYC; LAPtTA; Src-2+/+ and tet-o-MYC; LAPtTA; Src-2-/- liver tumors to identify biological processes specifically enriched in the Src-2-/- group. Biological processes were assessed for statistical significance (p <0.05).
ChIP-Seq for SRC-2 (at CT4) was performed by Active Motif, Inc. (Carlsbad, CA) as previously described with no additional filtering [17]. Briefly, mouse liver samples were submerged in PBS containing 1% formaldehyde, cut into small (~1 mm3) pieces with a razor blade and incubated at room temperature for 15 minutes. Fixation was stopped by the addition of 0.125 M glycine (final concentration). The tissue pieces were then treated with a TissueTearer and finally spun down and washed twice in PBS. Chromatin was isolated by the addition of lysis buffer, followed by disruption with a Dounce homogenizer. Lysates were sonicated and the DNA was sheared to an average length of 300–500 bp. Genomic DNA (Input) was prepared by treating aliquots of chromatin with RNase, Proteinase K and heated for reverse-crosslinking, followed by ethanol precipitation. Pellets were resuspended and the resulting DNA was quantified on a NanoDrop spectrophotometer. An aliquot of chromatin (30 μg) was precleared with protein A agarose beads (Invitrogen). Genomic DNA regions of interest were isolated using 4 μg of antibody. Complexes were washed, eluted from the beads with SDS buffer, and subjected to RNase and proteinase K treatment. Crosslinking was reversed by incubation overnight at 65°C, and ChIP DNA purified by phenol-chloroform extraction and ethanol precipitation. Illumina sequencing libraries were prepared from the ChIP and input DNAs by the standard consecutive enzymatic steps of end-polishing, dA-addition, and adaptor ligation. After a final PCR amplification step, the resulting DNA libraries were quantified and sequenced on Illumina NextSeq 500 (75 nt reads, single-end).
The sequences identified were mapped to the mouse genome (NCBI37/UCSC mm9) using BOWTIE function in Galaxy. Only the sequences uniquely mapped with no more than 2 mismatches were kept and used as valid reads. PCR duplicates were also removed. Peak calling was carried out by MACS (version 1.4.2 20120305) in Galaxy/Cistrome (options—mfold 10, 30—pvalue 1x10-5), on each ChIP-Seq file against the matching input file. To account for the different sequencing depths between samples, the signal files generated from MACS were normalized to sequencing depth [67]. The peak summits were used as the binding site centers, and the normalized signal files were used as the binding strength for further analysis. Assigning peaks to a given gene was performed with the Genomic Regions Enrichment of Association Tool (version 3.0.0) using the basal plus extension setting [68].
Cells and tissues were lysed in RIPA buffer and then homogenized using a Bioruptor sonicator (Diagenode). Proteins were quantified using the Bicinchoninic Acid (BCA) assay (Thermo Scientific) and subject to separation by using NuPage Bis-Tris gels (Invitrogen) for electrophoresis. The proteins were subsequently transferred to a nitrocellulose membrane. The membranes were blocked for 1 hour at room temperature and subsequently probed with primary antibodies overnight at 4°C. After incubating the membrane with the appropriate secondary antibody conjugated to horseradish peroxidase, protein levels were detected with SuperSignal Dura substrate (Thermo Scientific). Primary antibodies were prepared in 5% Milk or BSA in TBST. Antibodies were purchased from the following sources: SRC-2 (BD Biosciences, 1:250); DKK4 (Abgent, 1:1000); CADM4 (Neuromabs, 1:500); THRSP (Santa-cruz, 1:500); FXR (Santa Cruz, 1:50). SHP overexpression was detected with a V5 antibody (Invitrogen, 1:5000).
Human Embryonic Kidney 293T (HEK 293T, 1x108) cells were co-transfected with pLKO shRNA constructs (TRC, GE Dharmacon), and PAX2, MD2 helper plasmids using Lipofectamine 2000 (Life technologies). Following transfection, the lentiviral supernatant was collected, filtered and supplemented with 8ug/ml hexadimethrene bromide (Sigma). Human HCC cell lines Huh7 and HepG2 (3x105) were infected overnight twice with the viral supernatant and 24h after the second infection transferred into fresh media containing Puromycin (2 μg/ml). Cells were selected in puromycin media for at least 7 days and then harvested for RNA or western blot analysis to assess extent of knockdown.
To overexpress candidate genes in human HCC cells, human ORFs corresponding to each gene were cloned into the PLX304 or PLJM1 lentiviral plasmids. PLJM-eGFP or PLX303-empty constructs were used as negative controls. HEK 293T cells (1x108 cells) were then co-transfected with lentiviral overexpression or control constructs and helper plasmids PAX2 and MD2 using Lipofectamine 2000 (Life Technologies). Following transfection, the lentiviral supernatant was collected, filtered and supplemented with 8 μg/ml hexadimethrene bromide (Sigma). Human Huh7 cells (3x105) were infected overnight twice with the viral supernatant and 24h after the second infection transferred into fresh media containing blasticidin (4 μg/ml) or puromycin (2μg/ml). Control cells and cells overexpressing SRC-2 or SRC-2 target genes were selected in antibiotic-containing media for at least 7 days and then harvested for RNA and western blot analysis to assess overexpression.
Human HCC cells (3–5 x 106) expressing shRNA lentiviruses or lentiviruses overexpressing candidate genes in PBS were injected subcutaneously into both the left and right flanks of 6 week-old immunocompromised athymic nude mice (Charles River, strain 490). Tumor volume was measured using calipers every 3–4 days until the average tumor mass reached 2cm3. Tumor volume was calculated using the formula (length x width2)/2. A total of five mice were injected per experimental group, corresponding to ten experimental samples per group.
To measure in vitro proliferation of cells, the CellTiter 96 Aqueous Non-Radioactive Cell Proliferation assay kit (Promega) was used. 1000 cells per well were plated in 96-well plates in triplicate overnight. The MTS/PMS agent was added to the media according to the manufacturer’s protocol and incubated at 37°C for 1.5 hours. Absorbance was then measured at 490 nm every 24 hours for 6–7 days. All experiments were performed in triplicate and performed at least two times.
To predict NRs that interact with SRC-2, promoter regions (spanning 10kb on either side) of candidate SRC-2 target genes were screened for NR binding motifs using the NHR scan tool (http://www.cisreg.ca/cgi-bin/NHR-scan/nhr_scan.cgi). SRC-2 binding regions in candidate genes were overlapped with predicted NR binding motifs to predict potential SRC-2/NR interactions.
5 x104 Huh7 cells expressing an eGFP control or SRC-2 lentivirus were seeded per well in 12-well plates in triplicate. Cells were transfected 24 hours later using Fugene HD (Promega) with 20 ng FXR plasmid, 80 ng SHP-LUC or SHPΔ215-569-LUC reporter plasmids, 1 ng Renilla control reporter plasmid and 199ng pUC19 plasmid to give a total of 300 ng DNA per well. Empty pCMX vector was used as a no receptor control. The same transfection plan was followed for a replicate set of plates for downstream protein analysis by immunoblotting. Cells were lysed 48 h later and luciferase activity was measured in Glo-Max Microplate reader (Promega) using the Dual Luciferase assay reporter system (Promega). Luciferase data was obtained by normalizing Firefly activity to Renilla control activity and fold change induction was calculated relative to activity in eGFP control cells.
A Student t-test was used for comparisons between two groups with normal data distribution (for real time qPCR, MTS, and xenograft assays). A nonparametric method (Wilcoxon Rank Sum test) was used when data were not normally distributed (for the liver tumor burden analysis). In the Wilcoxon Rank Sum test, the Src-2+/+ group served as the reference, and was compared to either the Src-2-/- or Src-2+/- groups (multiple comparisons were not adjusted). SAS 9.4 TS Level 1M2 (Cary, NC) was used for data analysis. For survival analysis (S12 Fig), survival functions were constructed using Kaplan-Meier method and were compared using the log-rank test.
|
10.1371/journal.pntd.0006415 | Quantification of pathogenic Leptospira in the soils of a Brazilian urban slum | Leptospirosis is an important zoonotic disease that causes considerable morbidity and mortality globally, primarily in residents of urban slums. While contact with contaminated water plays a critical role in the transmission of leptospirosis, little is known about the distribution and abundance of pathogenic Leptospira spp. in soil and the potential contribution of this source to human infection.
We collected soil samples (n = 70) from three sites within an urban slum community endemic for leptospirosis in Salvador, Brazil. Using qPCR of Leptospira genes lipl32 and 16S rRNA, we quantified the pathogenic Leptospira load in each soil sample. lipl32 qPCR detected pathogenic Leptospira in 22 (31%) of 70 samples, though the median concentration among positive samples was low (median = 6 GEq/g; range: 4–4.31×102 GEq/g). We also observed heterogeneity in the distribution of pathogenic Leptospira at the fine spatial scale. However, when using 16S rRNA qPCR, we detected a higher proportion of Leptospira-positive samples (86%) and higher bacterial concentrations (median: 4.16×102 GEq/g; range: 4–2.58×104 GEq/g). Sequencing of the qPCR amplicons and qPCR analysis with all type Leptospira species revealed that the 16S rRNA qPCR detected not only pathogenic Leptospira but also intermediate species, although both methods excluded saprophytic Leptospira. No significant associations were identified between the presence of pathogenic Leptospira DNA and environmental characteristics (vegetation, rat activity, distance to an open sewer or a house, or soil clay content), though samples with higher soil moisture content showed higher prevalences.
This is the first study to successfully quantify the burden of pathogenic Leptospira in soil from an endemic region. Our results support the hypothesis that soil may be an under-recognized environmental reservoir contributing to transmission of pathogenic Leptospira in urban slums. Consequently, the role of soil should be considered when planning interventions aimed to reduce the burden of leptospirosis in these communities.
| Leptospirosis is a globally distributed zoonotic disease that disproportionately affects vulnerable populations in urban slums. The disease is transmitted by direct contact with water, soil, or mud that has been contaminated with infected urine shed from chronically infected animals. Despite the critical role the environment plays in the epidemiology of the disease, the contribution of soil to the transmission cycle remains largely undescribed. Herein, we investigated the distribution of pathogenic Leptospira in soil samples from an endemic urban slum in Brazil. We found pathogenic Leptospira in nearly one-third of the soil samples, predominantly in low concentrations (<5×102 GEq/g). However, we observed considerable variation in the distribution and concentration of the pathogen at the fine spatial scale within the slum. Our results indicate that soil is likely an important additional environmental reservoir of pathogenic Leptospira in urban slums and prevention strategies should consider soil to help prevent the transmission of the disease in similar settings.
| Leptospirosis is a life-threatening, zoonotic disease of global importance, with more than 1 million cases and approximately 60,000 deaths estimated annually, predominately in developing tropical countries [1]. The disease is caused by spirochetes of the genus Leptospira, which contains 35 species, 13 of which in the pathogenic group [2,3]. Pathogenic Leptospira chronically colonize the renal tubules of animal reservoirs and are excreted with the urine. Humans are incidental hosts, and often acquire the infection after seasonal or intense precipitation events [4–6], when pathogenic Leptospira in contaminated soil, mud, or water penetrate abraded skin or wounds [7,8]. Clinical manifestations of leptospirosis range from mild or asymptomatic to severe illness, such as Weil’s disease [7,9], and pulmonary hemorrhage syndrome [9,10], which cause high case fatality rates.
Urban leptospirosis has emerged as a pandemic which disproportionately affects residents of slum communities around the world [11]. Poor sanitation and an abundance of food sources provide ideal conditions for the maintenance of large rodent populations, specifically the Norway rat (Rattus norvegicus), which are the primary animal reservoirs of pathogenic Leptospira in urban environments [12–16]. Infrastructure deficiencies facilitate the transmission to humans [17–19]: Open sewers and inadequate drainage systems allow contaminated water to pervade surrounding soil and water, and unpaved walkways expose residents to contaminated dirt and mud. Thus, the human-environment interface plays a critical role in the epidemiology and transmission of leptospirosis in urban slums. While previous studies have found Leptospira spp. in puddles, sewers, streams, and soil in endemic regions [20–25], the distribution and dynamics of leptospires in the environment, particularly in soil, are largely unknown.
To date, few publications have reported pathogenic Leptospira in soil, presumably because leptospirosis is generally considered to be a water-borne disease, and thus environmental studies have focused preferentially on aquatic matrices [20,21]. Previous studies that successfully analyzed soil reported very low prevalence among samples (4.9% in China [26]), or the isolation of only a few pathogenic strains in Philippines, Malaysia and New Caledonia [2,23,24,27,28]. These studies, however, were based on culture isolation of leptospires, followed by molecular identification. Culture techniques lack sensitivity because of pathogenic Leptospira spp. being overgrown by the autochthonous soil microbiota or saprophytic and intermediate Leptospira [23,24,29]. Notably, a recent study by Thibeaux et al. [25] reported a 57.7% prevalence in soil samples from a suspected environmental infection site in New Caledonia by using a qPCR targeting lipL32.
In this study, we aimed to quantify the pathogenic Leptospira load in soil samples from an urban slum community in Salvador, Brazil. Previous studies have shown this community has a high burden of disease [6,18,30] and a widespread presence of pathogenic Leptospira in its surface waters [31], making it an excellent location for an environmental survey of the pathogen. Since Norway rats are the predominant pathogenic Leptospira reservoir in this community [14], we hypothesized that the presence of rats would be associated with the occurrence and abundance of the pathogen in soil. As urban slum communities in tropical regions share many socioeconomic, structural, and environmental characteristics [32,33], our study may help inform potential public health interventions in similar epidemiological settings around the world.
We conducted this study in the community of Pau da Lima, an urban slum located in Salvador, Brazil. This site has been extensively described in previous studies [18,19]. Briefly, the community encompasses four interconnected valleys within 0.46 km2 (Fig 1A), with a population of 14,122 inhabitants residing in 3,689 households, according to a 2003 census [18]. The community lacks adequate sanitary infrastructure, resulting in untreated wastewater and rain runoff flowing through open sewers in the lower parts of the valleys. Leptospirosis is endemic in Pau da Lima, with a mean annual infection incidence of 37.8 per 1,000 inhabitants, and 19.8 severe cases per 100,000 inhabitants [30]. We selected three collection sites to represent a range of microenvironments present within the community (Fig 1B). Site A (900 m2) (12°55'22.2"S, 38°26'04.2"W) was located along an open sewer at the bottom of the valley, and included households, areas of domestic animal raising, and thick vegetation (Fig 2A). The steep, high banks of the sewer served as a barrier to separate households from the sewage, which limited potential flooding. Site B (900 m2) (12°55'17.9"S, 38°26'07.2"W) was situated at a higher elevation next to the valley slope and had closed sewage drains, paved stairs, and patios (Fig 2B). There was limited vegetation, although fences and barriers, coupled with the sheer valley slope, restricted access. Site C (400 m2) (12°55'24.8"S, 38°26'06.3"W) was situated at the bottom of the valley and in proximity to an open sewer with low embankments, allowing frequent flooding of surrounding areas during heavy rainfall periods. The thick vegetation and water-logged soil made part of the site inaccessible. We partitioned collection sites into 5 m x 5 m squares (A and C) or 10 m x 10 m (B) (Fig 2A and 2B). A larger grid was used at Site B, as many areas were impassible due to the decreased number of sampling locations and the challenging terrain (Fig 2B).
We collected soil and sewage samples in the rainy season between July and August 2014 (S1 Fig). A tracking board to monitor rat activity in the collection sites was placed as described previously [34] within all grid squares that were accessible and contained exposed soil. Tracking boards were evaluated daily over the course of three days for evidence of rat activity as ascertained by the identification of footprints, scrapes, and tail slides [34]. After the three days, one or two soil samples of 100–200 g were collected at a depth of 5–10 cm from non-adjacent areas within 30 cm of the edge of each tracking board between 9am and 12.30pm. Grid squares that were inaccessible because of private property, dense vegetation or water-logged soil, or contained no exposed soil were not included in the sampling. In total, we collected soil samples from 23, 7 and 11 grid squares in sites A, B and C, respectively for a total of 35, 14 and 22 soil samples. If a portion of an open sewer was included in the grid square and was accessible, two 50 mL samples for each sampling point were collected. The presence of vegetation and distance to open sewers and households was recorded for each grid square. The soil moisture and clay content were measured for one sample from each grid square. Samples were placed in aseptic containers, transported to the laboratory, and processed within 6 h.
To maximize the recovery of Leptospira DNA from soil samples, we developed an extraction protocol and determined its efficiency by performing spiking experiments with known concentrations of L. interrogans (S1 Supplementary methods). The final procedure consisted in the following steps: Subsamples of 5 g of soil were mixed with 40 mL of sterile double-distilled water and vortexed at maximum speed for 2 min. Samples were centrifuged at 100 rcf for 5 min. The supernatant was recovered and centrifuged at 12,000 rcf for 20 min at room temperature. The pellets were recovered, resuspended in 1.5 mL of sterile double-distilled water and centrifuged at 12,000 rcf for 20 min. Finally, the samples were decanted and the pellets were frozen at -80°C. Sewage samples (40 mL) were processed as described previously [35]. DNA was extracted from pellets using PowerSoil DNA Isolation Kit (Mo Bio Laboratories). An extraction blank (sterile double-distilled water) was added to each extraction batch to monitor for cross-contamination.
We quantified Leptospira spp. loads using two TaqMan assays targeting the lipL32 gene or the 16S rRNA gene as described previously [36,37] with minor modifications. Briefly, all reaction mixtures (25 μL) contained 12.5 μL Platinum qPCR SuperMix (Life Technologies), 0.2 μg/μL of bovine serum albumin (Ambion), and 5 μL of DNA template. The lipL32 reactions included 500 nM each of primers LipL32-45F and LipL32-286R, and 100 nM of LipL32-189P probe (S1 Table). 16S rRNA reactions included 300 nM each of primers Lepto F and Lepto R, and 200 nM of 16S probe. Amplifications were performed using a 7500 Fast Real-Time PCR System (Life Technologies) with the following conditions: an initial step of 2 min at 50°C, followed by 2 min at 95°C, and 40 cycles of amplification (15 s at 95°C and 1 min at 60°C). Genomic DNA obtained from L. interrogans serovar Fiocruz L1-130 [38] was used to construct calibration curves with concentrations ranging from 2 × 102 to 2 × 109 GEq/mL, which we included in each qPCR run. Efficiencies were always higher than 92.5%. All samples were run in duplicate, and included non-template controls in each plate row to detect any contaminating DNA. All negative controls (extraction controls and non-template controls) were negative. All DNA extractions and quantitative PCR (qPCR) analyses were performed according to the minimum information for publication of quantitative real-time PCR experiments (MIQE) guidelines [39].
To determine the specificity of the lipL32 and 16S rRNA qPCRs for pathogenic, intermediate and saprophytic Leptospira, we tested DNA extracts from 21 Leptospira species (S2 Table) as explained above, adjusting the concentration in each well to 107 GEq based on each genome’s size [40].
To confirm the specificity of the environmental qPCR reactions, we randomly selected and sequenced (Sanger method) 11/22 (50%) of the soil samples with positive results for lipl32. We also sequenced 27/78 (35%) of samples with a positive result for 16S rRNA qPCR (25 soil and 2 sewage samples). In brief, qPCR products were purified from 2% agarose gels using the QIAquick Gel Extraction Kit (QIAgen) following the manufacturer’s instructions. The products were sequenced using primers LipL32-45F or Lepto-F for lipL32 and 16S rRNA, respectively, corrected using BioEdit 7.2.0 (Ibis Biosciences), and the sequences compared using BLAST to those available in the NCBI. In addition, we performed a phylogenetic analysis for the lipl32 and 16S rRNA amplicons using the Phylogeny.fr platform [41]. Maximum Likelihood trees were inferred by PhyML 3.0 [42] using the Hasegawa-Kishono-Yano (HKY85) substitution model and 1000 bootstrap replicates. We used FigTree v1.4.2 (http://tree.bio.ed.ac.uk/software/figtree) to visualize and edit the final trees. The 16S rRNA tree included Leptonema illini DSM 21528 as the outgroup.
Leptospira concentrations were log10 transformed and analyzed as follows. The theoretical lower limit of quantification (LOQ) of the qPCR assays was calculated using the assumption that 1 copy of the targeted gene was amplified in a reaction (4 GEq/g). We assigned positive samples with concentrations below this threshold a value equal to the LOQ. Fisher’s exact test and the χ2 test were used to compare prevalence among sites and the associations of dichotomized environmental characteristics with the proportion of qPCR-positive samples. t-tests with a Welch’s correction were used to compare soil moisture content and clay component values between positive and negative qPCR samples. The median concentrations between sampling sites were compared using Kruskal-Wallis test with Dunn’s correction for multiple comparisons. Mann-Whitney test was used to compare the concentrations between soil and sewage at site C and the overall concentrations obtained by lipl32 and 16S rRNA qPCRs. Non-parametric tests were chosen due to the non-normal distribution of the data. Although equivalent parametric tests (one-way ANOVA and t-test) were inappropriate, technically, they supported the results of the non-parametric tests in all cases. To analyze the degree of agreement between lipL32 and 16S PCR detection methods, we used Cohen’s kappa coefficient. All statistical analyses were performed using GraphPad Prism 6.05 (GraphPad Software Inc.), and multivariate models were fit using the GENMOD procedure with a GEE model in SAS v9.3 (SAS Institute Inc., Cary, NC) to evaluate the relationship between the measured environmental characteristics and outcome of sample testing.
We collected 70 soil samples within the study area (34, 14, and 22 from sites A, B, and C, respectively) (Fig 2). Of those 70 samples, 22 (31%) were positive for pathogenic Leptospira as measured by the lipL32 qPCR (Fig 3A). There were no significant differences in the proportion of positive samples between the three collection sites (26%, 21% and 45%, respectively; p = 0.78) (Table 1), indicating that pathogenic Leptospira were widespread in the study site. These prevalences are lower than those recently reported using the same qPCR method in river bank soils from active leptospirosis transmission sites in New Caledonia (57.7%) [25]. Nevertheless, our results contrast with the very low number of pathogenic Leptospira isolates that are usually obtained from soil in endemic areas using culture-based approaches [23,24,26–28], suggesting that molecular approaches may capture better the presence of the pathogen in soil.
The concentration of pathogenic Leptospira among positive samples was predominantly low (median: 6 GEq/g), but varied by two orders of magnitude (range: 4–4.31×102 GEq/g) even within the same collection grid. This indicates a considerable heterogeneity of environmental loads within the slum microenvironment. Furthermore, there were no significant differences in the concentration of the qPCR-positive soil samples among the three collection sites (p = 0.16) (Fig 3A). Among the eight sewage-water samples collected from site C, seven (88%) were positive by the lipL32 qPCR with a median concentration of 0.5 GEq/mL.
Together our results showed that pathogenic Leptospira were present in low concentrations in soils sampled from diverse microenvironments within the urban slum. Contact with mud in the peridomiciliary environment was previously identified as a risk factor for leptospirosis infection in this setting [18], which suggests that soil may serve as an important environmental reservoir of the pathogen. Intense rainfall events during the rainy season would cause mobilization and dispersion of pathogenic Leptospira from soil to run-off as described for other environmental pathogens such as E. coli, Salmonella spp., Cryptosporidium spp. and fecal indicator bacteria [43–48]. Simultaneously, run-off originated in the higher areas of the valley may contaminate soil in the lower areas with pathogenic Leptospira through flooding and sewer overflow. Thus, soil may act as a source and a recipient of the pathogen depending on the specific location and weather conditions. Furthermore, the low concentrations found in soil are in agreement with those found in sewage and standing water in a previous study conducted in the same setting [31], which supports the hypothesis that the environmental load of pathogenic Leptospira is generally low, even in endemic areas. The dynamics and characteristics of water-based mobilization and dispersion of Leptospira to and from the soil reservoir within the slum community, and the role that low environmental concentrations may have on the risk of acquiring leptospirosis, will require detailed studies beyond the scope of our methods.
In contrast to the lipL32 qPCR results, the 16S rRNA qPCR detected Leptospira from 60/70 (86%) soil samples, significantly more than detected by the lipL32 qPCR (p<0.0001). Higher prevalences were found at sites A and C (88% and 100%, respectively) than site B (57%) (p<0.0014) (Table 2). Among positive samples, the median concentration was 4.16×102 GEq/g (range: 4–2.58×104 GEq/g), significantly higher than the one detected by lipL32 qPCR (6 GEq/g, p < 0.0001). All eight sewage samples from Site C were also positive using the 16S qPCR, with a mean concentration of 2.09×102 GEq/mL (range: 98–2.81×102), nearly 9-fold higher than that detected by lipL32 qPCR (24 GEq/mL, p = 0.0003). Notably, all soil and sewage samples that were positive using the lipL32 assay were positive with the 16S assay, though there was a poor agreement between the qualitative results obtained by the two methods (Cohen’s Kappa coefficient = 0.14± 0.05) (Table 2)
Given the discrepancy of results obtained with the two-qPCR methods (Table 2), we analyzed the specificity for detecting pathogenic Leptospira for each method. We sequenced the lipL32 amplicon from 11/22 (50%) lipL32 qPCR-positive soil samples. In all cases, the sequences presented similarities greater than 92% with lipL32 gene sequences from pathogenic Leptospira deposited in GenBank (S3 Table) and clustered with species from the pathogenic group (Fig 4A). These results confirmed that the lipL32 qPCR was highly specific for pathogenic Leptospira in soil as it was previously shown in sewage [31]. We also sequenced 27/78 (35%) 16S qPCR-positive samples (25 soil and 2 sewage samples). All the sequenced 16S rRNA amplicons clustered with species from the intermediate Leptospira group (Fig 4B). As observed elsewhere [49], a single mismatch in the approximately 50 bp fragment sequenced discriminates between intermediate and pathogenic Leptospira groups. Indeed, the reverse primer used in the 16S qPCR is degenerated at position 14 allowing for the hybridization with T and C bases, and thus, may detect both pathogenic and intermediate species (S1 Table). Of note, 6 sequences that were positive for both lipL32 and 16S presented a double peak in the sequencing chromatogram at the mismatch position. This indicates that both sequences coexisted in the sample [49], although in all cases the highest peak was the one belonging to the intermediate group.
To conclusively determine the specificity of the both qPCRs, we tested 21 type strains from the pathogenic, intermediate and saprophytic groups. Lipl32 qPCR showed signal only for pathogenic species, and excluded intermediate and saprophytic ones. In contrast, all pathogenic and intermediate species gave positive results for 16S rRNA qPCR and no signal was observed for the saprophytes (S2 Table). Therefore, 16S rRNA qPCR detects not only pathogenic Leptospira, but also intermediate species. Since the pathogenicity of the intermediate group is not well established, we considered only the results obtained with lipL32 qPCR for subsequent analyses.
Bivariate and multivariate analyses identified no significant associations between the presence of pathogenic Leptospira DNA in soil and environmental characteristics such as vegetation, distance to an open sewer or a house, or soil clay content (Table 1). However, we found that 62% of samples with a moisture content higher than 20% were positive, while only 21% were positive when the moisture was lower than 20%, which is consistent with previous observations that higher soil moisture content is associated with increased Leptospira isolation [23,50]. Additionally, previous studies have reported that Leptospira persist for a longer time in moist and super-saturated soils than in drier ones, although the duration of survival is also dependent on the serovar and other characteristics of the soil such as pH [51–53].
Furthermore, against our initial hypothesis, we did not find any significant association between rat activity and the presence of pathogenic Leptospira in soil. A potential explanation is that the direct association with rats is confounded by other animal sources of pathogenic Leptospira both domestic (cows, pigs and dogs) and wildlife (opossums and bats) that coexist in this urban slum [15,34]. Alternatively, as discussed above, some of the pathogenic Leptospira detected may have originated in other parts of the valley and contaminated the soil through run-off or floodwater, making rat activity an unreliable proxy for the presence of the pathogen. Finally, we cannot rule out that the tracking board method was not a sufficient to assess rat activity at the fine scale at which the variation of the presence and concentration of pathogenic Leptospira in the soil seems to occur. Studies with larger sample sizes and an increased diversity of sites sampled are required to track the origin of pathogenic Leptospira in soil and determine relationships with environmental characteristics potentially obscured by our relatively small sample size.
The concentrations of Leptospira in soil detected using the 16S qPCR were higher than those detected by lipL32 in all samples. Moreover, the difference between both measurements was higher than 0.60 log10 units in all but one sample (mean difference and SD: 2.05±0.89 log10 units). Since 16S qPCR detects pathogenic and intermediate species while lipL32 qPCR detects only pathogenic ones and both methods exclude saprophytic Leptospira (S2 Table), the observed concentration differences suggest that most of the signal detected with 16S qPCR originates from intermediate species. Hence, Leptospira species from the intermediate group may be more ubiquitous and present in significantly higher concentrations in soil from this community relative to pathogenic ones.
Intermediate species such as L. fainei, L. licerasiae, L. wolffii, and L. broomii have been linked to human leptospirosis cases [54–58], although they are not considered as virulent as the species from the pathogenic group, and thus are less relevant from a public health perspective. It is important to note that no cases of leptospirosis caused by an intermediate species has been reported in Pau da Lima during 15 years of active surveillance [6,18,30,59]. Previous studies of pathogenic Leptospira in the environment using 16S qPCR [20,49,60] might have led to a overestimation of the burden of the pathogen. Our results illustrate the importance of using highly specific tests to detect pathogenic bacteria in estimations of disease burden and environmental reservoir load.
Inherent limitations of qPCR in environmental samples may influence the accuracy of our estimates and ability to evaluate risk associations. First, qPCR may be detecting DNA from dead or damaged cells that have lost the ability to cause infection and therefore, our results may overestimate the environmental risk. On the other hand, although we optimized sample processing, DNA extraction, and detection methodologies to reduce DNA loss and PCR inhibition, we may have underestimated pathogenic Leptospira loads in the soil if they were below the LOD. Second, our sampling scheme may not have completely captured the heterogeneity in the urban slum environment. While we evaluated three study sites representative of different microenvironments within the slum, there may be additional heterogeneity with respect to soil type, climatic conditions, land use, and rat activity levels, which should be further explored. Finally, the relatively small sample size limited our ability to draw robust conclusions concerning environmental factors contributing to positive and negative soil samples.
To date, most research regarding the environmental reservoirs of pathogenic Leptospira has focused on water matrices such as sewage, puddles, wells, or freshwater. Our results are the first to successfully quantify the burden of pathogenic Leptospira in soil from an endemic region, and indicate that soil is an additional environmental reservoir in the life cycle of pathogenic Leptospira. As with other environmentally transmitted diseases, the mobilization of leptospires from the sub-surface soil, either by heavy rainfall, flooding, or excavation, would contribute to environmental exposures with a sufficient dose to produce infection in humans. Furthermore, understanding the specific role and impact of soil as an environmental reservoir and the relation of low environmental concentrations to the risk of human disease is critical to our knowledge of the leptospirosis transmission cycle. Importantly, our data suggest that efforts to eliminate or reduce access to recognized transmission sources, such as open sewers, may not be sufficiently effective to decrease the risk of infection. Consequently, the role of soil in the transmission dynamics and epidemiology of leptospirosis should be considered when designing public health interventions in endemic areas.
|
10.1371/journal.pgen.1007473 | RES complex is associated with intron definition and required for zebrafish early embryogenesis | Pre-mRNA splicing is a critical step of gene expression in eukaryotes. Transcriptome-wide splicing patterns are complex and primarily regulated by a diverse set of recognition elements and associated RNA-binding proteins. The retention and splicing (RES) complex is formed by three different proteins (Bud13p, Pml1p and Snu17p) and is involved in splicing in yeast. However, the importance of the RES complex for vertebrate splicing, the intronic features associated with its activity, and its role in development are unknown. In this study, we have generated loss-of-function mutants for the three components of the RES complex in zebrafish and showed that they are required during early development. The mutants showed a marked neural phenotype with increased cell death in the brain and a decrease in differentiated neurons. Transcriptomic analysis of bud13, snip1 (pml1) and rbmx2 (snu17) mutants revealed a global defect in intron splicing, with strong mis-splicing of a subset of introns. We found these RES-dependent introns were short, rich in GC and flanked by GC depleted exons, all of which are features associated with intron definition. Using these features, we developed and validated a predictive model that classifies RES dependent introns. Altogether, our study uncovers the essential role of the RES complex during vertebrate development and provides new insights into its function during splicing.
| RES complex is essential for splicing in yeast but its function and role during vertebrate development are unknown. Here, we combined genetic loss-of-function mutants with transcriptomic analysis and found that a subset of introns is particularly affected in RES complex knock-out background. Those introns display the major hallmarks of splicing through intron definition mechanisms (short introns, rich in GC and flanked by GC depleted exons). Moreover, bud13, rbmx2 and snip1 mutant embryos showed a marked brain phenotype with a RES-dependent introns enrichment in genes with neurodevelopmental function. Altogether, our study unveils the fundamental role of RES complex during zebrafish embryogenesis and provides new insights into its molecular function in splicing.
| Splicing is critical step in eukaryotic gene expression and is an important source of transcriptomic complexity [1]. Splicing is carried out by the spliceosome, a large macromolecular complex that includes five small nuclear ribonucleoproteins (snRNPs; U1, U2, U4, U5 and U6) and hundreds of core and accessory proteins that ensure the accurate removal of introns from pre-mRNAs [2]. Canonically spliced introns are removed through two transesterification reactions during a complex process involving the recruitment and release of multiple core splicing factors [3]. However, many introns are recognized by different mechanisms depending on their specific features. Short introns with high GC content are believed to be spliced through an “intron definition” mechanism, in which initial U1-U2 pairing occurs across the intron. On the other hand, long introns surrounding short exons are recognized and spliced through “exon definition” mechanisms, in which the initial pairing bridges across the exon [4]. While these mechanisms are widely accepted, little is known about the specific factors associated with each process.
The pre-mRNA REtention and Splicing (RES) complex is a spliceosomal complex conserved from yeast to human. It is organized around the U2 snRNP-associated protein Snu17p/Ist3p (RBMX2 in human), which binds to both the pre-mRNA-leakage protein 1 (Pml1p; SNIP1 in human) and bud site-selection protein 13 (Bud13p; BUD13 in human) [5–7]. Snu17p interacts directly with the pre-mRNA and with Bud13p, and both proteins have been involved in splicing; on the other hand, Pml1p has been mainly linked to the retention of unspliced pre-mRNA in the nucleus [6–9]. Furthermore, the components of the RES complex cooperatively increase the stability and the binding affinity of the complex for the pre-mRNA [10–13], highlighting the importance of cooperative folding and binding in the functional organization of the spliceosome [10–12].
In yeast, the RES complex interacts with the 3’end of the intron in the actin pre-mRNA and is required for the first catalytic step of splicing [7,13]. Additionally, microarray-based studies have shown a global effect of the RES complex on yeast splicing [14,15]. Mutations in the RES-complex genetically interact with other spliceosomal components, and several introns seem particularly sensitive to RES complex loss-of-function, often in association with weaker splice sites [6,9,13,16–18]. Interestingly, disruption of the genes encoding for the three subunits of the RES complex show consistent phenotypes including slow growth, thermosensitivity [6] and alteration in budding pattern [19].
While the RES complex was identified in yeast, its function in vertebrates, the features recognized by this complex and its role during development are unknown. Here, we found that expression of the RES complex is enriched in the CNS during early development. We generated loss-of-function mutants for the three components of the RES complex in zebrafish using an optimized CRISPR-Cas9 gene editing system [20]. The three mutants showed severe brain defects with a significant decrease in the number of differentiated neurons and increased cell death in the brain and the spinal cord. We observed a mild retention across most introns, consistent with a global effect on splicing. However, a subset of introns was strongly affected. Importantly, these retained introns showed the hallmarks of intron definition [4,21], as they: (i) were shorter, (ii) had a higher GC content, and (iii) were neighbored by lower GC content exons. We developed a logistic regression model with these and other genomic characteristics that allowed us to discriminate between RES-dependent and independent introns with high accuracy. Altogether, these results provide new insights into the function of the RES complex and identify the features associated with RES-dependent splicing.
To determine the role of the RES complex during vertebrate development, we first analyzed its expression pattern during development. bud13, rbmx2 and snip1 (Fig 1A) are maternally expressed (Fig 1B, S1D and S1F Fig), and later in development their mRNAs are strongly expressed in the central nervous system (CNS) (26 hours post fertilization, hpf), (Fig 1B) suggesting that RES complex may be required for brain development. Next, we generated mutant zebrafish lines for rbmx2, snip1 and bud13 using an optimized CRISPR-Cas9 system (S1A Fig) [20]. We identified a seven-nucleotide deletion in bud13 (bud13∆7/∆7), a sixteen-nucleotide deletion in rbmx2 (rbmx2∆16/∆16) and eleven-nucleotide deletion in snip1 (snip1∆11/∆11). These mutations are predicted to cause premature stop codons and disrupt protein function (Fig 1C–1E). Zygotic mutants for the three components of the RES complex showed a Mendelian ratio of homozygous mutant embryos. We observed strong structural brain defects and widespread cell death in the CNS at ~30 hpf in bud13 and ~48 hpf in rbmx2 and snip1 (Fig 1F–1H; S1B, S1C, S1E and S2B Figs). The embryo progressively degenerates and mutants die by 4–5 dpf. Earlier depletion of bud13 gene expression using morpholino antisense oligonucleotide targeting the AUG start site showed a more severe phenotype affecting not only the brain and CNS but also other tissues (e.g. mesoderm) (S2A Fig), consistent with a repression of the maternal contribution in the morphants compared to the zygotic mutants. Both, the mutant and morphant phenotypes were specific and fully rescued by injection of the cognate mRNA (bud13, rbmx2 or snip1) (Fig 1F–1H) or human mRNA (hBUD13) up to 6 dpf (S2B Fig). Interestingly, rescuing with lower amounts of hBUD13 mRNA partially phenocopied the rbmx2 and snip1 mutant phenotypes at 48 hpf (S2C Fig). These results suggest that the onset of the zygotic phenotype in rbmx2 and snip1 mutants (48 hpf) is likely due to differences in the maternal mRNA contribution and/or protein stabilities for these genes. Taken together, these results demonstrate that (i) the mutant phenotypes are specific to the targeted loci, (ii) the RES complex is essential for embryonic development and suggest that (iii) the biochemical function of Bud13, and presumably of the RES complex, may be conserved from human to zebrafish.
To determine the role of the RES-complex in splicing and gene expression, we analyzed polyA+ RNA from each mutant at the onset of the phenotype. As a control, we analyzed the transcriptome of stage-matched wild type siblings (See material and methods). We observed 621 up-regulated genes in all three mutants, that were significantly enriched for functions related to i) cell death (e.g. p53, casp8 and puma) (S4D Fig), consistent with the appearance of apoptosis in the brain (Fig 2A and S1B Fig), and ii) spliceosomal components, suggesting a compensatory effect upon a general splicing deficiency (See Materials and Methods for details, S4A, S4C Fig and S1 Table). In contrast, 745 genes were consistently down-regulated and were enriched for genes involved in transcriptional regulation (such as sox19b, atoh7, pou3f1) and nervous system development (e.g. neurod1/4/6b, sox1a) (S4B Fig and S1 Table).
Importantly, bud13∆7/∆7 mutants showed normal neural induction, morphogenesis and regionalization. For example, key brain areas such as the zona limitans intrathalamica (ZLI; dorsal shh in the diencephalon), the mid/hind-brain boundary (pax2a expression), and rhombomeres 3 and 5 (krox20 expression) were properly specified (S3 Fig). This suggests that the zygotic function of bud13 is not required for initiation of neural linage patterning and specification. In contrast, we observed a reduction in the number of differentiated neurons, including both excitatory and inhibitory neuronal populations, with significant decrease in glutamatergic as well as GABAergic neurons in the forebrain in bud13∆7/∆7 embryos at 32 hpf (Fig 2B and 2C). These results suggest that zygotic RES activity is required for neuronal differentiation and/or survival, but not for neural induction and early brain patterning.
To determine the global impact of the RES complex mutants on splicing, we analyzed the level of intron retention (IR) across the transcriptome. Briefly, for any given intron, the percent intron retention (PIR) is calculated as the average number of reads mapping to the 5’ and 3’ Exon-Intron (E-I) junctions over the average number of reads mapping E-I junctions plus any Exon-Exon (E-E) junction that supports removal of that given intron (Fig 3C) [22]. We found that 74–79% of introns showed increased retention (ΔPIR > 0) in the individual mutants compared to wild-type siblings (72,926 introns, with sufficient coverage, see Methods for details). In contrast, a significantly smaller set of introns (7–9%) showed decreased retention in the mutants (Fig 3A and S7C Fig, P<2.2e-16, Fisher exact test). Other types of splicing events were less impacted upon RES depletion (Fig 3B), although a substantial number of exons became skipped (S7D Fig) in the loss-of-function mutants, consistent with a general disruption of splicing. Interestingly, the three mutants shared 5,339 (~35%) introns with a medium-high level of retention (ΔPIR>5) (Fig 3D), consistent with a common function in the RES complex. This is likely an underestimate because each mutant was analyzed at the onset of the mutant phenotype, which is different across these mutants likely due to different level of maternal recue or protein stability. We observed mild widespread intron retention, yet ~3.5% of introns showed strong increased retention across each mutant (ΔPIR >15, S7C Fig), suggesting that a subset of introns have a stronger dependence for RES function in vivo. This effect was validated by RT-PCR for a subset of candidates for each mutant (Fig 3E, Fig 3F and S5 Fig), including wdr26b and ptch2, that when mutated in humans cause neurodevelopmental disorders [23,24]. Genes with strongly retained introns (ΔPIR >15) in at least two of the three mutants were enriched in transcriptional regulation and DNA binding (S6A Fig), including known regulators of vertebrate development and neuronal differentiation (e.g. irx1a, smn1, enc1, smad4, tbx2a, and nkx6.1/6.2; S2 Table). Further analyses revealed complex interactions between pre-mRNA splicing and mRNA abundance in response to RES complex depletion. Genes with at least one strongly retained intron (ΔPIR >15) had lower expression in the mutants compared to genes without intron retention (ΔPIR < 2) (S6B Fig; P<10–7 for the three mutants, Wilcoxon Sum Rank test). Conversely, genes that were differentially expressed in the three mutants (down- or up-regulated) show significantly higher retention (higher ΔPIR) in the mutants compared to their WT siblings (S6C Fig; P<10–8 for all comparisons in the three mutants, Wilcoxon Sum Rank test). Finally, genes with increased expression significantly overlapped with those with at least one strongly retained intron (ΔPIR>15) (S6D Fig). Taken together, our analyses indicate that the RES complex has a global effect on splicing, and is strongly required for a subset of introns in genes involved in transcriptional regulation and neural development.
To identify features that are primarily associated with RES complex function, we first defined a set of introns that were confidently dependent on RES for proper splicing (1,409 “RESdep” introns with ∆PIR>15, ≥1.5-fold net increase in intron reads; see Methods) (Fig 3C). As a control set (Ctr), we defined 5,574 introns with ∆PIR<0.5 in all three mutants, and evaluated the enrichment of 44 features, many of which have been previously associated with intron retention [22] (S3 Table). Consistent with the genome-wide patterns (Fig 4), strongly retained introns were enriched for last introns (Fig 4A and 4B and S7A Fig) and introns that do not trigger NMD when mis-spliced (Fig 4A and 4C and S7B Fig), suggesting that their accumulation is in part likely due to reduced degradation of the unspliced transcript isoform. However, a significant fraction of highly retained introns was predicted to elicit NMD upon inclusion. We thus hypothesized that this subset of NMD-triggering introns contains specific features that would maximally associate with RES-dependent mechanisms. Based on this, we also separately analyzed introns that were predicted to trigger NMD (574) and those that were not (569) (see Methods).
RES-dependent introns (i) were significantly shorter than the control set (Fig 5A, median of 281 nt versus 749 nt in the control, P<0.001, Mann-Whitney-U test), (ii) had elevated GC content (Fig 5B, P<0.001 Mann-Whitney-U test), (iii) were flanked by exons with lower GC content (Fig 5C and 5D, S8C and S8D Fig), and (iv) had weaker branch point (BP) consensus sequences [25,26] (Fig 5G and S8 Fig). Remarkably, at the genome-wide level, short introns and introns with high GC content also showed a higher retention across all three mutants compared to wild type siblings (S9 Fig). NMD-triggering introns further showed weaker core splicing signals, including acceptor (3’) and donor (5’) splice sites and BP consensus sequences than any other intron set (Fig 5E–5G), providing an explanation for their increased sensitivity upon RES complex disruption despite being subject to NMD.
Next, we assessed how genomic and transcript features define the dependence on the RES-complex. We applied a logistic regression model using 30 features as predictor variables (S3 Table), and defined a response variable classifying each intron as RES-dependent or non-dependent. We developed a model using 90% of the RESdep and a size-matched subset of control introns as training set, and validated the model on the remaining 10% of the data (see Material and Methods). We found a high performance in the classification of introns, with an average Area Under the ROC curve (AUC) of 0.821 (Fig 6A). This model was able to classify the impact of mutating each individual RES component with an AUC of ≥0.76. Subsequently, we analyzed the individual contribution of each feature to the model and their potential to reduce the null deviance (see Methods). Consistent with results in Fig 5, this analysis identified four important features i) the ratio between exon and intron length, ii) the ratio between exon and intron GC content, iii) gene expression levels and iv) the position of the intron within the transcript (last intron effect) (Fig 6B and S10 Fig).
To further test the model on an independent set of introns, we applied the model to 108,470 introns (S6 Table) without sufficient read coverage across our six RNA-seq samples and were not used for previous analyses. We then selected the top 100 introns predicted to be bud13-dependent and -independent and plotted their ∆PIR values based on RNA-seq data for the bud13 mutant and the corresponding control (S11A Fig). The majority of predicted bud13-dependent introns were substantially retained, whereas introns that were predicted to be unaffected showed a median ∆PIR close to zero (0.91) (S11A Fig). The false prediction rate (FPR) is 0.34 for the bud13 dependent introns and 0.16 for the unaffected introns, consistent with the AUC values reported above. Repeating this independent validation with rbmx2 and snip1 dependent introns (S11A Fig) showed FPRs of 0.33 and 0.03 for rbmx2, and 0.44 and 0.04 for snip1. The lower FPRs for the unaffected introns reflect that predicting unaffected introns can be obtained with higher performance, likely because the data contain a considerably larger amount of unaffected than affected introns. We further validated these predictions by RT-PCR assays for five predicted RES-dependent and five RES-independent introns (S11B Fig). Thus, altogether, our logistic regression analysis can identify introns dependent on the RES complex based on specific features within the genomic locus and the transcript.
Finally, to test our logistic regression model in isolation from the genomic context, we assayed four introns, two RES dependent and two RES independent, using minigene constructs including the tested intron plus the two flanking exons (Fig 6C and 6D and S12 Fig). Minigenes were cloned into a transgenesis vector (S12B Fig), and injected into 1-cell stage embryos. To assess the splicing pattern, we carried out a RT-PCR followed by PCR (S4 Table). From the subset of four different introns, three events were validated by the RT-PCR and gel electrophoresis as predicted by our model (Fig 6A, 6B, 6C and 6D and S12 Fig). These introns were in serpinb1l3 and col1a2 (predicted as RES independent) and wdr26b (predicted as RES dependent with ∆PIR>15) transcripts. Nevertheless, we detected no intron retention in ptch2, suggesting that its retention upon RES depletion may depend on its genomic context.
Splicing regulatory information is encoded by multiple sequence features, from the core signals (splice donor and acceptor and branch point) to other, less understood, sequence elements [27–29]. Our results identify intronic features that are associated with RES-dependent splicing across the transcriptome. These features can be used to discriminate a large fraction of RES-dependent from independent introns. Although loss-of-function for RES complex components caused mild intron retention across the transcriptome, we observed a subset of introns that were strongly accumulated across mutants of the RES complex. Recent in vitro studies on the single intron of the actin gene in yeast showed that the RES complex binds at the 3’ of the intron, between the BP and the acceptor site [12]. We observed that introns that more strongly depend on the RES complex show weaker BP consensus sequences. Furthermore, this subset of introns were shorter and had higher GC content, an association that is particularly striking in zebrafish, since short introns normally have lower GC content [21]. RES-dependent introns are flanked by exons with a lower GC content than RES-independent introns (Fig 5, S9C and S9D Fig). Therefore, these observations are consistent with a model whereby RES-dependent introns are mainly spliced through intron definition [4,21]. This association is surprising, since biochemical evidence suggest that the RES complex joins the spliceosome after recognition of the splice sites [12], and that the RES complex is not needed for spliceosome assembly in vitro but for U1 and U4 snRNP dissociation before the first catalytic step [7]. One possible explanation for this apparent discrepancy is that RES complex components play a role in early splice site recognition in vivo and therefore that the biochemical functions reported in a limited set of RNAs reflect limitations of in vitro splicing reactions. Alternatively, the RES complex may not be involved in early splice site recognition, but could be a limiting factor for splice site pairing or other steps in spliceosome assembly progression in vivo, particularly for introns defined by intron definition, highlighting differences in molecular pathways for intron- and exon-defined splicing. These concepts are in line with increasing evidence that splice site selection can be modulated at late stages of spliceosome assembly or even catalysis [14,27,30–35]. Finally, some RES-dependent introns also have weaker donor and acceptor splice site consensus sequences, and thus are expected to be more sensitive to defects on the splicing machinery. This is consistent with previous studies in yeast, which found that weaker 5’ splice sites increased susceptibility to RES loss-of-function [6].
An unexpected observation from genome-wide analyses of core splicing factor loss-of-function experiments is that each factor seems to differentially affect a specific subset of introns and exons [32,36]. This suggests that splicing of each intron in the genome is limited by specific core factors, depending on its combination of sequence features, as we observed for RES-dependent introns. As such, disruption of core splicing factors is predicted to produce unique phenotypes dictated by its expression, and the expression and function of genes that contain the subsets of introns sensitive to that factor. Consistent with this hypothesis, RES complex is required during brain development and neuronal survival, and mis-regulated introns are found in genes with well-known functions in neurodevelopment (e.g. irx1a, smn1, enc1, smad4, tbx2a, and nkx6.1/6.2). Specifically, zygotic mutants in bud13, snip1 or rbmx2 show microcephaly and decreased populations of GABAergic and glutamatergic neurons, despite normal specification and regionalization of the CNS (Fig 1, Fig 2 and S3 Fig). This phenotype is different from those described for a few other spliceosomal-related mutants in zebrafish [37–40]. For instance, while sf3b1 is required for early neural crest development [40], loss of another core component of the spliceosome, prpf8, results in massive neuronal cell death and impaired myeloid differentiation [37]. These differences might be caused by the different half-life of the maternal proteins in the zygotic mutants. Alternatively, different components of the splicing machinery might be essential in a cell-type/tissue specific manner during early development. This may also explain why mutations in specific spliceosomal components cause human diseases with diverse phenotypes, such as Taybi-Linder syndrome, microcephalic osteodysplastic primordial dwarfism type I and retinitis pigmentosa [41–43]. Interestingly, a mutation in human SNIP1 (p.Glu366Gly) has been associated with epilepsy and skull dysplasia [44]. Our data shows that human BUD13 can rescue loss of bud13 function in zebrafish, and future studies will be needed to determine whether Bud13 has a conserved function during brain development in humans (S2B and S2C Fig).
In summary, we have shown that RES complex disruption in zebrafish hinders splicing, but is not essential for the removal of most introns, indicating that such introns can be efficiently defined and spliced through RES-independent mechanisms. However, we found that a subset of introns is particularly affected by RES complex removal and that those introns display the major hallmarks of splicing through intron definition mechanisms. From a functional perspective, RES-dependent introns are in genes enriched for transcription factors and neurodevelopmental regulatory functions, thus resulting in brain developmental defects in loss-of-function zygotic mutants (Fig 7). Future studies will be needed to understand how spliceosomal mutations disrupt splicing of different genes by affecting specific limiting steps in pre-mRNA splicing resulting in diverse disease phenotypes.
Fish lines were maintained in accordance with research guidelines of the International Association for Assessment and Accreditation of Laboratory Animal Care, under a protocol approved by the Yale University Institutional Animal Care and Use Committee (IACUC).
Wild-type zebrafish embryos were obtained through natural mating of TU-AB and TLF strains of mixed ages (5–17 months). Selection of mating pairs was random from a pool of 48 males and 48 females allocated for a given day of the month. bud13Δ7/Δ7, rbmx2Δ16/Δ16, snip1Δ11/Δ11 were obtained through natural mating of heterozygous bud13+/ Δ7, rbmx2+/ Δ16, snip1+/ Δ11 mutants, respectively (see below gene editing using CRISPR-Cas9). Tg(dlx6a-1.4kbdlx5a/dlx6a:GFP) lines [45,46] were obtained from the laboratory of Marc Ekker and Tg(vglut:DsRed) [47] from the laboratory of Joseph Fetcho.
Embryos were analyzed using a Zeiss Axioimager M1 and Discovery microscopes and photographed with a Zeiss Axiocam digital camera. Images were processed with Zeiss AxioVision 3.0.6.
Whole-mount embryos were imaged in vivo by confocal microscopy (Leica TCS SP8 systems, Yale Center for Cellular and Molecular Imaging.) Mutant embryos and their wild-type siblings were scored at 30–32 hours post fertilization stage. Embryos were anesthetized using Tricaine and mounted in 0.6% agarose at an orientation where the frontal view of the brain was imaged. Embryos were imaged at 40x (1.3 Oil) using Z-stacks ranging from 70.41–96.02 μm, with a z stepping size of 0.4 μm. Z-stacks started at the first appearance of the GABAergic cells (GFP-labeled) and ended where GABAergic cells (GFP-labeled) could no longer be visualized. Each xy plane spanned 227.94 μm with a pixel size of 0.075 μm. Maximum intensity projections were shown for all confocal images. Images were processed using Fiji [48], Imaris (Bitplane) and Huygens deconvolution software (Scientific Volume Imaging). Figures were assembled using Illustrator (CC, Adobe). To quantify the number of glutamatergic (labeled in DsRed) and GABAergic cells (labeled in GFP) in the bud13Δ7/Δ7 mutant and their wild-type siblings respectively, two blinded raters segmented the raw z-stack images using ImarisCell module (Bitplane) and computationally counted the segmented cells in each channel (GFP and DsRed). Identical thresholds and parameters were applied to all samples for segmentation processing. Since the quantification performed by both independent raters yield consistent fold change in the respective cell counts between the bud13Δ7/Δ7 mutant and their wild-type siblings. Only one set of the analyzed results was displayed. Statistical analyses were conducted using Prism 6 (Graphpad).
To visualize apoptotic cells, vital dye acridine orange (Sigma) was used in live and dechorionated embryos. Embryos were incubated 2 minutes in PBS pH 7.1 with 2 ug/ml of acridine orange in the dark. After 3 brief washes in PBS, the embryos were placed in plates with 1% agarose and viewed with fluorescence microscopy, using the FITC filter set 1 [49].
Zebrafish bud13, rbmx2 and snip1 ORFs were PCR amplified (S4 Table) using cDNA from 2 and 6 hpf zebrafish embryos and cloned in pSP64T (bud13) in pT3TS [50] (rbmx2 and snip1). bud13 PCR product was cut using NotI and EcoRI and ligated into pSP64T. rbmx2 and snip1 PCR products were cut with NcoI and SacII and ligated into pT3TS. To optimize Kozak sequence, the forward oligonucleotide used for rbmx2 ORF introduced extra aminoacid in the 2nd position (GGC). Human bud13 ORF was cloned in pSCDest [51] using gateway gene cloning system (Thermo Fisher Scientific). Final constructs were confirmed by sequencing. To generate mRNAs, the template DNA was linearized using XbaI (pT3TS), BamHI (pSP64T) or KpnI (pSCDest) and capped mRNA was synthetized using the mMessage mMachine T3 (pT3TS), or SP6 (pSP64T and pSCDest) kit (Ambion), respectively and in accordance with the manufacturer’s instructions. In vitro transcribed mRNAs were DNAse treated and purified using the RNeasy Mini Kit (Qiagen). All mRNA rescued the mutant phenotypes when 50–100 pg were injected in one cell stage embryo.
The plasmid (modified from [52]) for intron retention validation in vivo, pTol2(hsp-MCS-polyA, CMV-eGFP-SV40), is a Tol2 transposon-based, bipartite construct consisting of heat-shock promoter (hsp), a multiple cloning site (MCS), to insert the desired validation cassette, flanked by Xenopus globin UTR and polyadenylation tail (polyA) as well as a cis-linked CMV promoter and SV40 poly(A)-regulated eGFP reporter. Briefly the plasmid was built as following: pT2(kop:Cre-UTRnos3,CMV:eGFP) [52] was cut with Kpn1 to remove kop:Cre-UTRnos3. A synthetic DNA fragment containing HSP, 5’UTR Xenopus globin, MCS and 3’UTR Xenopus globin was obtained from Integrated DNA Technologies (IDT) and PCR amplified with specific primers (S4 Table). In-Fusion cloning protocol (Clontech) was performed using the cut vector and the PCR product to get the backbone vector pTol2(hsp-MCS-polyA, CMV-eGFP-SV40). Cassette for validation were obtained as synthetic DNA fragments (Integrated DNA Technologies, IDT) or amplified from genomic DNA (see S4 Table for details) and cloned directionally in frame with XhoI and SacII. Final constructs were confirmed by sequencing.
A morpholino targeting bud13 mRNA start codon was obtained from Gene Tools and re-suspended in nuclease-free water. 1 nl of morpholino solution (0.6mM) was injected into wild-type dechorionated embryos at the one-cell stage.
A mix of 4 plasmid (15 pg/embryo) with the desired cassette (minigene), 2 predicted as no retained (serpinb1l3, col1a2) and 2 predicted as retained (wdr26b, ptch2) upon RES loss-of-function was injected in 1 cell stage embryo together with Tol2 mRNA (Addgene plasmid #31831) (33 pg/embryo). Embryos were sorted in bud13∆7/∆7 and bud13+/? at the onset of the phenotype (~ 28–30 hpf). Minigene expression under HSP promoter was induced by heat shock during 4 hs and then GFP positive embryos were collected for RT-PCR. PCR was performed using the specific Fw primer an a universal Rv globin (S4 Table, Fig 6C and 6D and S12 Fig).
CRISPR-Cas9-mediated gene editing was performed as described previously [53]. Briefly, 3 different sgRNAs (20 pg each) targeting bud13 gene (S4 Table) were co-injected together with 100 pg of mRNA coding for zebrafish codon optimized Cas9-nanos in one-cell stage embryos (S1A Fig). Cas9-nanos concentrates gene editing in germ cells and increases the viability of injected embryos [20]. F0 founders were mosaic and they were backcrossed with wild-type fish and then F1 fish were genotyped using their corresponding oligos per target site (S4 Table). Heterozygous adult fish bud13+/ Δ7 (Fig 1) were selected to generate bud13Δ7/Δ7 mutants. Similar approach was followed to generate rbmx2 and snip1 mutants but injecting 2 sgRNAs (S4 Table and Fig 1).
krox20, shh and pax2a in situ probes were previously described [54–56]. RES antisense and sense digoxigenin (DIG) RNA probes were generated by in vitro transcription in 20 μl reactions consisting of 100 ng purified PCR product (8 μl), 2 μl DIG RNA labelling mix (Roche), 2 μl ×10 transcription buffer (Roche), and 2 μl T7/T3 (antisense probes) and SP6 (sense probes) RNA polymerase (Roche) in RNase-free water and purified using a Qiagen RNEasy kit. In situ protocol was followed as detailed previously [57]. To reduce variability, wild-type sibling and bud13Δ7/Δ7 embryos were combined in the same tube during in situ hybridization and recognized based on their phenotype. Before photo documentation, embryos were cleared using a 2:1 benzyl benzoate:benzyl alcohol solution. Images were obtained using a Zeiss stereo Discovery V12.
Total RNA from 32 hpf bud13 Δ7/Δ7, 48 hpf rbmx2 Δ16/Δ16, 48 hpf snip1 Δ11/Δ11 embryos and their corresponding siblings was extracted using Trizol (ten embryos per condition). Strand-specific TruSeq Illumina RNA sequencing libraries were constructed by the Yale Center for Genome Analysis. Samples were multiplexed and sequenced on Illumina HiSeq 2000/2500 machines to produce 76-nt paired-end reads.
RNA used for intron retention validation experiments was treated with TURBO DNase (Ambion) for 30 min at 37°C and extracted using phenol chloroform. Then, Polyadenylated RNAs were purified using Oligo d(T)25 Magnetic Beads (Invitrogen) following manufacter recommendations. cDNA was generated by reverse transcription with random hexamers using SuperscriptIII (Invitrogen). RT–PCR reactions were carried out at an annealing temperature of 59°C for 35–40 cycles. PCR products were run in a 1.5% agarose gel. Primers are listed in the S4 Table.
For the qPCR experiment, total RNA was extracted as described above. GFP and dsRED mRNAs were used as spike-in RNA controls and 1 μg of total RNA was used to generate cDNA. 5 μl from a 1/50 dilution of the cDNA reaction was used to determine the levels of p53 in a 20 μl reaction containing 1 μl of each oligo forward and reverse (10 μM) (S4 Table), using Power SYBR Green PCR Master Mix Kit (Applied Biosystems) and a ViiA 7 instrument (Applied Biosystems). PCR cycling profile consisted of incubation at 50°C for 2 min, followed by a denaturing step at 95°C for 10 min and 40 cycles at 95°C for 15 s and 60°C for 1 min. Primers are listed in S4 Table.
Zebrafish embryos or a small amount of tissue from the end of the tail were used to extract DNA [58]. Briefly, embryos or fin clipped were incubated in 80 μl of NaOH 100mM at 95°C for 15 min producing a crude DNA extract, which was neutralized by the addition of 40 μl of 1 M Tris-HCl, pH 7.4 (Sigma-Aldrich). 1 μl of this DNA extraction was used as a template for PCR reactions using the primers described in S4 Table.
Gene expression levels for each condition were calculated from RNA-seq data using the cRPKM metric (corrected-for-mappability Reads Per Kilobasepair of uniquely mappable positions per Million mapped reads [59]. For this, a reference transcript per gene was selected from the Ensembl version 80 annotation for Danio rerio using BioMart (25,935 genes in total, S5 Table) and uniquely mappable positions for each transcript were calculated as previously described [59]. Quantile normalization of cRPKM values was done with ‘normalizeBetweenArrays’ within the ‘limma’ package.
To identify differentially expressed genes, we first filtered out genes that did not have cRPKM > 2 in all sibling control or all mutant samples and genes whose quantification was not supported by at least 50 read counts in at least 1 sample. Next, differentially expressed genes were defined as those that showed a fold change in expression of at least 1.5 in all 3 mutants and a fold change of at least 2 in 2 out of the 3 control vs. mutant individual comparisons (bud13, rbmx2 and snip1). Gene Ontology analysis was performed with the online tool DAVID (https://david.ncifcrf.gov/ Version 6.8) using as background all genes that passed the initial filters (minimum expression and read count).
Annotated introns for each reference zebrafish transcript in Ensembl version 80 were extracted and those that overlapped with other genes were removed yielding a total of 182,017 valid introns. To calculate the percent of intron retention (PIR) for each intron in a given RNA-seq sample, we used our previously described pipeline [22] with the following modification: to calculate intron removal, all exon-exon junctions supporting the splicing of the intron were used and not only those formed between the two neighboring exons. This was done to avoid false positives in the case of introns associated to cassette exons or other alternative splicing events. For all analysis, only introns with sufficient read coverage across the six samples were considered (at least 15 reads supporting the inclusion of one splice site and 10 of the other, or a total of 15 reads supporting splicing of the intron).
To define the confident set of highly affected introns, potential false positives were filter out by comparing the density of the mapped reads in the introns bodies in the mutant vs the control. For this purpose, we extracted all intronic sequences and calculated the number of uniquely mappable positions per intron following a similar strategy to that used to calculate cRPKMs [59] (see above). Specifically, every 50-nucleotide (nt) segment in 1-nucleotide sliding intronic windows was mapped to a library of full-length intronic sequences plus the whole genome, using bowtie with–m 2 –v 2 parameters (every intronic segment must map at least twice, to its own individual intron sequence and to the corresponding position in the whole genome). Segments that mapped more than twice were considered as multi-mappable positions, whereas those that did not map (e.g. due to undetermined (N) nucleotides in the assembly) were considered as non-mappable. The number of uniquely mappable positions of an intron is defined as the total number of segments minus multi- and non-mappable positions. Next, each RNA-seq sample was mapped to the same library of intronic plus full genomic sequences using–m 2 –v 2 to obtain the unique intronic reads counts. However, to minimize potential artifacts derived from the heterogeneity of the intronic sequences (e.g. high number of reads mapping to a transposable element or an expressed nested gene), if a given intronic position showed a read count more than five times higher than the median read count of the whole intron, then the read count of this position was set to 5 × median (if the median read count was 0, then the maximum read count for any given position was set to 5). Finally, ciRPKM scores were calculated (corrected-for-mappability intronic Reads Per Kilobasepair of uniquely mappable positions per Million mapped reads) for each intron and condition by dividing this number of counts by the number of uniquely mappable positions in that intron.
With this information, the set of confidently retained introns upon RES complex disruption (“RESdep” introns) were defined as those introns with ∆PIR > 15 and at least a 1.5-fold net increase in read density in the intron body calculated as:
[ciRPKMmut/(100‑PIRmut)]/[ciRPKMsib/(100‑PIRsib)]
in at least 2 out of 3 mutants (1,413 introns in total). As a control, we also define a set of confidently non-retained introns as those with a ∆PIR < 0.5 in the 3 mutants (“Ctr” set; 5,577 introns).
To analyze the number of retained introns in the 3 mutants and the inter-mutant overlaps Euler APE-3.0.0 software [60] was utilized.
To investigate the impact of non-sense mediated decay (NMD) on global intron retention upon RES mutation, all introns were separated as last or non-last introns (of the reference transcript) and between those predicted to trigger and not to trigger NMD. An intron was predicted to trigger NMD if its retention generated an in-frame stop codon that is located further than 50 nts upstream of an exon-exon junction [61]. By definition, last introns cannot trigger NMD. ∆PIR values were plotted as boxplots for each category, and two-sided Wilcoxon Sum Rank tests were used to evaluate statistical differences between the distributions.
To identify features discriminating introns highly retained upon RES depletion from un-retained/un-affected introns, we compared the sets of confidently introns (“RESdep” in Fig 5.) with control introns (“Ctr” in Fig 5.). Moreover, as introns that are predicted not to trigger NMD are expected to be more often accumulated unspecifically, two subsets for “RESdep” introns were generated: (i) those introns in genes with more than five introns, are not the last three introns of the gene, and that are predicted to trigger NMD (“NMD”, 577 introns); and (ii) predicted not to trigger NMD or cause a frame shift upon inclusion, unless they are the last intron of the gene (“no-NMD”, 569 introns). For these different sets of introns, 44 features were extracted (S3 Table), including intron and exon length and GC content, strength of 3' and 5' splice sites, branch point (BP) related features, and transcript length, using custom scripts in combination with the following two external tools: MaxEntScan scripts for determining the strength of 3' and 5' splice sites [62]; and SVM-BPfinder software for determining BP related features (BP strength, distance from BP to 3' splice site, and pyrimidine track length) [26]. For the latter analysis, the 150 nts upstream of the 3' splice sites were extracted and these sequences were used as input for SVM-BPfinder. Furthermore, we recorded the highest log-score of the SF1 position weight matrix binding model across these 150-nt intronic sequences [25].
We applied logistic regression models to the discrimination between differentially retained introns (retained) and non-differentially retained introns (control) upon RES mutations. We focused on the set of confidently retained introns (“RESdep”, 1,409 introns) versus control introns with an absolute ∆PIR < 0.5 in the three mutants (“Ctr”, 5,565; we removed 9 introns for which we could not determine all features). The binary response variable of the logistic regression models indicates for each intron if it belongs to the retrained or control group. As predictors we used 30 quantitative and qualitative features (S3 Table), including intronic and exonic characteristics, position along the transcript and gene expression in wild type conditions, among others. The binomial logistic regression models were learned using Lasso variable selection [63–65] available in R through library glmnet (2.0.10), and the generalized linear model function glm from the R stats library.
To investigate the overall classification performance, we randomly partitioned the data set “RESdep” of retained introns into 90%/10% (i.e., 1268/141) training/test data, and randomly sampled the same amount of training/test data from the control “Ctr” data set. We used the training data to learn a logistic regression model with Lasso variable selection and tested it on the test data. Next, to evaluate how well this model classifies specific subsets of "RESdep" introns and retained introns specific for each mutation, we applied the model trained with “RESdep” vs “Ctr” data having held fixed its parameters to the classification of the 141 control test-introns vs. 141 retained introns subsampled from the following sets: (i) “RESdep_∆PIR10” introns from the “RESdep” set with ∆PIR > 10 in all three mutants (871 introns); (ii) “NMD”, introns from the “RESdep” set predicted to trigger NMD when retained (574 introns); (iii) “bud13”, introns with ∆PIR>15 upon bud13 mutation at 32 hpf (2,363 introns); (iv) “rbmx2”, introns with ∆PIR>15 upon rbmx2 mutation at 48 hpf (2,186 introns); and (v) “snip1”, introns with ∆PIR>15 upon snip1 mutation at 48 hpf (2,675 introns). We repeated this procedure, including model training and classification of test data, 10,000 times and report average ROC curves (Fig 6A) and average model coefficients for each feature extracted from the trained models. These averages indicate the direction of the effect (e.g. positively [blue] or negatively [red] associated with retention upon RES mutation; Fig 6B; S10B Fig and S3 Table).
To study the potential of each feature to contribute to the discrimination between the "RESdep" and "Ctr" intron sets, we randomly partitioned the dataset of retained introns into 90%/10% (i.e., 1268/141) training/test data, and randomly sampled the same amount of training/test data from Ctr. Using the training data, we learned logistic regression models without Lasso variable selection using only a single feature at a time neglecting all other features. The test data were used to determine the AUC. This experiment was repeated 10,000 times and we report average AUCs for each feature in Fig 6B and S10B Fig. In addition, the fraction of the null deviance that was reduced by each single-feature model was recorded, and the average reductions of the null deviance for each feature are reported in S3 Table.
|
10.1371/journal.pgen.1004699 | daf-31 Encodes the Catalytic Subunit of N Alpha-Acetyltransferase that Regulates Caenorhabditis elegans Development, Metabolism and Adult Lifespan | The Caenorhabditis elegans dauer larva is a facultative state of diapause. Mutations affecting dauer signal transduction and morphogenesis have been reported. Of these, most that result in constitutive formation of dauer larvae are temperature-sensitive (ts). The daf-31 mutant was isolated in genetic screens looking for novel and underrepresented classes of mutants that form dauer and dauer-like larvae non-conditionally. Dauer-like larvae are arrested in development and have some, but not all, of the normal dauer characteristics. We show here that daf-31 mutants form dauer-like larvae under starvation conditions but are sensitive to SDS treatment. Moreover, metabolism is shifted to fat accumulation in daf-31 mutants. We cloned the daf-31 gene and it encodes an ortholog of the arrest-defective-1 protein (ARD1) that is the catalytic subunit of the major N alpha-acetyltransferase (NatA). A daf-31 promoter::GFP reporter gene indicates daf-31 is expressed in multiple tissues including neurons, pharynx, intestine and hypodermal cells. Interestingly, overexpression of daf-31 enhances the longevity phenotype of daf-2 mutants, which is dependent on the forkhead transcription factor (FOXO) DAF-16. We demonstrate that overexpression of daf-31 stimulates the transcriptional activity of DAF-16 without influencing its subcellular localization. These data reveal an essential role of NatA in controlling C. elegans life history and also a novel interaction between ARD1 and FOXO transcription factors, which may contribute to understanding the function of ARD1 in mammals.
| The development of a living organism is influenced by the environmental conditions such as nutrient availability. Under starvation conditions, the C. elegans larvae will enter a special developmental stage called dauer larva. An insulin-like signaling pathway controls dauer formation as well as adult lifespan by inhibiting the activity of FOXO transcription factor DAF-16 that regulates expression of stress-resistant genes. Here we isolate a new gene called daf-31; this gene encodes a protein that regulates C. elegans larval development, metabolism and adult lifespan. This protein has been found in other species to be part of an enzyme that functions to modify other proteins. We show that overexpression of our newly discovered protein stimulates the transcriptional activity of DAF-16. Interestingly, abnormal regulation of human proteins similar to DAF-31 results in tumor formation. It is known that human FOXO proteins prevent tumorigenesis. Therefore, it is possible that abnormal DAF-31 activity may lead to tumor growth by reducing DAF-16 activity. Thus, the present study may not only contribute to understanding the role of a universal enzyme in controlling development, metabolism and lifespan in other organisms besides worms but may also shed light on the mechanisms of tumorigenesis in humans.
| Animal development is a complex process that involves hierarchical gene regulatory networks and is influenced by environmental conditions. When food is abundant, the post-embryonic development of C. elegans consists of four larval stages (L1–L4) and the adult. During the L1 stage, environmental factors determine whether C. elegans molts to an L2 larva or a pre-dauer L2d larva [1]. At least three environmental cues have been defined: food supply, temperature, and a constitutively secreted dauer-inducing pheromone that signals population density [2]. The L2 larva is developmentally committed to continued growth, whereas the L2d larva can molt to a dauer larva if food is scarce and the animals are overcrowded, or to an L3 larva should conditions improve.
Mutations affecting dauer larval development include dauer-defective (daf-d) mutations that prevent entry into the dauer stage, and dauer-constitutive (daf-c) mutations that mandate entry into the dauer stage [2]. Based on epistatic relationships between daf-c and daf-d mutations, more than twenty genes controlling dauer formation have been ordered in a genetic pathway [2] representing generation of the pheromone signal [3], response by chemosensory neurons [4], [5] and transduction of the signal to other cells. Three functionally overlapping neural pathways control the developmental response to environmental cues. They involve DAF-7/TGF-ß [6], [7], DAF-11/cyclic GMP [8], and DAF-2/insulin-like [9], [10] pathways, which relay the environmental signals to a nuclear hormone receptor, DAF-12 [11], to control dauer versus non-dauer morphogenesis.
Mutations in two genes, daf-9 and daf-15, lead to non-conditional formation of detergent-sensitive dauer-like larvae [12]. These mutants form dauer larvae constitutively and display some characteristics of dauer larvae formed under starvation, such as a high density of intestinal and hypodermal storage granules. daf-9 encodes a cytochrome P450 related to those involved in the biosynthesis of steroid hormones in mammals [13], [14]; it was found to specify a step in the biosynthetic pathway for a DAF-12 steroid ligand called dafachronic acid [15]–[17]. daf-15 encodes the C. elegans ortholog of Raptor [18] that is proposed to interact with C. elegans target-of-rapamycin kinase (LET-363/CeTOR) to control C. elegans larval development [18]. Both daf-9 and daf-15 also regulate fat metabolism and adult lifespan [13], [14], [18].
The dauer-like mutants represent a mutant class distinct from the previously defined daf-c and daf-d mutants. Unlike most daf-c mutants, the dauer-like mutants are not ts, and they do not complete dauer morphogenesis. The daf-d genes such as daf-12 have non-conditional alleles and fail to respond to pheromone [1], but unlike the dauer-like mutants they can execute non-dauer development. The dauer-like mutants define a third class of mutants, one in which the animals are incapable of executing either complete dauer or non-dauer development.
The daf-31 mutant was isolated in genetic screens to identify genes similar to daf-9 and daf-15 [17]. The overall aim of the present study was to clone the daf-31 gene and characterize the DAF-31 function. Our genetic epistasis analysis suggests daf-31 functions downstream of or in parallel to daf-3, daf-12 and daf-16 dauer-defective genes, and acts upstream of or in parallel to daf-15/raptor. We cloned the daf-31 gene by positional cloning and showed that it encodes an ortholog of arrest-defective-1 protein (ARD1), the catalytic subunit of the major N alpha-acetyltransferase (NatA). Moreover, our data reveal that daf-31 has an essential role in controlling C. elegans larval development, metabolism and adult longevity.
Entry into the dauer stage is determined by the pheromone/food ratio, with high pheromone and low food supply favoring dauer formation [2]. Dauer larva is considered as an alternative L3 larval stage. Compared to L3 larva (Figure 1A), the dauer larva has a constricted pharynx (Figure 1B) and a special cuticle with dauer alae (Figure 1C). In the presence of dauer-inducing pheromones, daf-31 mutants cannot form SDS-resistant dauer larvae [17]. In order to determine whether daf-31 mutants enter the dauer stage in response to starvation, we examined the progeny of strain unc-24daf-31/nT1 under starvation conditions and observed uncoordinated (Unc) dauer larvae. These dauer larvae showed normal dauer features, such as a dark body, fully constricted pharynx (Figure 1D), and a cuticle with dauer alae (Figure 1E). However, daf-31 dauer larvae were not SDS-resistant like normal dauer larvae. Furthermore, daf-31 dauer larvae could not resume development when food was provided, dying shortly thereafter. Therefore, daf-31 mutants could not complete dauer morphogenesis under starved conditions, and those incomplete dauer larvae could not finish reproductive development after food was provided.
Fat accumulation is one characteristic of C. elegans dauer larvae. We examined fat accumulation in daf-31 homozygous mutants using Sudan Black B staining. As shown in Figure 1F, daf-2 mutant dauer larvae accumulate fat as described previously [9]. The daf-31 mutant worms also accumulate more fat droplets than wild-type worms and fat droplets in the daf-31 mutants are larger than those in wild-type worms (Figure 1G and 1H). To confirm this phenotype, Nile red was used to stain fixed worms; this approach has been reported to reliably detect fat droplets in C. elegans [19]. Similar to Sudan Black staining, Nile red also detected fat accumulation in daf-31 mutant worms (Figure S1). Therefore, daf-31 mutants shift metabolism to fat accumulation.
To position daf-31 in the dauer formation pathway, we examined the epistatic relationship between daf-31 and daf-d genes including daf-3, daf-12 and daf-16. The daf-31 mutation is epistatic to all three daf-d mutations as judged by the ratio of progeny (1∶2∶1) (Table 1). For the epistasis analysis with daf-16, the ratio of progeny is 1∶2 because nT1 homozygous animals are lethal. We repeated the epistasis analysis of daf-31 and daf-12 by using the daf-12(rh61rh411) null allele [11] and obtained a similar result (Table 1). These epistatic relationships suggest that daf-31 functions downstream of or in parallel to daf-3, daf-12 and daf-16 in dauer formation.
To examine the epistatic relationship of daf-31 and daf-15, we injected dsRNA of daf-15 into unc-24daf-31/nT1 young adult worms. Wild-type (N2) animals were treated equally and used as controls. We examined the phenotype of progeny reproduced at various time periods after injection. The progeny reproduced between seven and eighteen hours after injection arrested development at a dauer-like stage three days after egg lay (Table 2). These dauer-like animals have a similar phenotype to daf-15 mutants. Thus, regarding dauer entry, it appears that the daf-15 mutation is epistatic to the daf-31 mutation. We scored the recovery of both N2 and unc-24daf-31 dauer-like animals two days after dauer-like arrest. 48% of N2 dauer-like worms remained at dauer-like stage and the rest of the animals recovered and grew to L4 larval or adult size (Table 2). By contrast, 100% of unc-24daf-31 animals stayed at dauer-like stage (Table 2). For unc-24daf-31 dauer-like larvae without daf-15 RNAi treatment, the majority of these animals died within five days. However, surviving animals all grew to L4 larval or adult size (n = 52). Taken together, these data indicate that daf-15 is epistatic to daf-31 as unc-24daf-31 mutants treated by daf-15 RNAi form dauer-like larvae similar to daf-15 mutants. Moreover, these two mutants have a synergistic effect on C. elegans development because no unc-24daf-31 dauer-like larvae treated by daf-15 RNAi recovered. Thus, these two genes may function in the same pathway and daf-31 is upstream of daf-15. However, the possibility that these two genes act in parallel cannot be excluded.
A positional cloning strategy was used to identify the daf-31 gene on chromosome IV between unc-24 and fem-3 (Figure S2A). daf-31 was found to lie between the physical SNP markers T09A12 and F17E9 (Figure S2A). A genomic fragment corresponding to the K07H8.3 open reading frame fully rescued the genetic daf-31 null mutant [daf-31(m655)] phenotype, i.e. the transgenic animals did not form dauer-like larvae, but grew to fertile adults. Sequence analysis of the daf-31 gene in the mutant revealed a 393 bp deletion which removed 151 bp of promoter upstream of the ATG start codon and 242 bp of daf-31 coding region downstream of the ATG start codon, which may completely block daf-31 transcription as both the essential promoter region and the N-terminal portion of the gene were deleted (Figure S2B). Primers were designed to flank the deletion region and PCR analysis of mutant worms' genomic DNA detected a 1,449 bp band (393 bp smaller than the wild-type band) in both homozygous and heterozygous daf-31(m655) mutant worms (Figure S2C).
The daf-31 gene encodes the ortholog of ARD1 with a predicted molecular weight of 21.2 kDa. ARD1 is the catalytic subunit of NatA that catalyzes the acetylation of proteins beginning with Met-Ser, Met-Gly and Met-Ala [20]. Amino acid sequence alignment showed that DAF-31 shares 75% identity with human ARD1, 77% identity with mouse ARD1, 72% identity with Drosophila melanogaster ARD1 and 46% identity with yeast ARD1 (Figure S2D).
Given that there is only a single mutant allele of daf-31, we used RNAi to inhibit daf-31 in the N2 background to confirm the daf-31 mutant phenotype. Inhibition of daf-31 by feeding animals with E. coli that express daf-31 dsRNA did not induce a dauer-like phenotype. From our previous work, dsRNA injection can create a stronger mutant phenotype similar to that of a genetic null mutant [18]. In vitro synthesized daf-31 dsRNA was injected into gonads of N2 young adult worms and the progeny displayed a dauer-like phenotype similar to the daf-31(m655) mutants. The starvation-induced dauer morphology of daf-31(m655) mutants, such as dauer alae formation and contrasted pharynx (described in Figure 1) could not be examined using this RNAi method. Therefore, we examined fat accumulation in daf-31 RNAi-treated animals. Similar to the daf-31(m655) mutant, daf-31 RNAi-treated animals accumulated fat as detected by both Sudan Black and Nile red staining of fixed animals (Figure S3). Based on these results, we conclude that the daf-31 mutant phenotypes described in this study most likely resulted from daf-31 mutation instead of secondary mutations in the background.
In order to characterize the daf-31 expression pattern, we constructed a daf-31 promoter::gfp reporter construct. In N2 animals, GFP expression was detected from L1 to the adult stages in multiple tissues including the hypodermis, pharynx, intestine, and neurons (Figure 2). To confirm the GFP expression pattern of daf-31 promoter fusion reporter, we constructed daf-31 translational fusion reporter genes in which the GFP open reading frame was fused to the full-length daf-31 genomic DNA in frame either at the N-terminus or at the C-terminus. Both translation fusion reporter genes fully rescued the dauer-like phenotypes of daf-31 mutants. However, we did not observe GFP expression in the rescued daf-31 mutant worms, a phenomenon previously reported with other GFP fusion gene mutant rescues [21]. Thus, our observations of daf-31 expression pattern were limited to the daf-31 promoter fusion, which may not represent the endogenous expression pattern of the entire daf-31 gene if enhancer elements are present in introns or in 3′ untranslated sequences.
Increased adult longevity is a phenotype associated with many dauer mutants. Since daf-31 homozygous mutants arrest development at L4 stage, we inhibited the daf-31 gene by feeding RNAi. The RNAi treatment successfully reduced daf-31 mRNA level as measured by qRT-PCR (Figure S4). However, daf-31 RNAi treatment had no obvious effect on the lifespan of wild-type worms as the daf-31 RNAi-treated worms had similar mean and maximum lifespans as control vector RNAi-treated worms (Figure 3A and Table S1). When RNAi-sensitive rrf-3(pk1426) mutants were treated with daf-31 RNAi, their lifespans were significantly decreased (p<0.0001, log-rank test) (Figure 3B and Table S1). Compared to controls, the mean lifespan of rrf-3 mutants treated with daf-31 RNAi was shortened by four days (Figure 3B and Table S1).
To test if daf-31 mutations influence the longevity phenotype of daf-2 mutants, we treated the rrf-3(pk1426);daf-2(e1370) mutant with daf-31 RNAi. The mean lifespan of daf-31 RNAi-treated rrf-3;daf-2 mutants was five days shorter compared to that of control animals (p = 0.0005, log-rank test) (Figure 3C and Table S1). Thus, inhibition of daf-31 partially suppressed the longevity phenotype of daf-2 mutants. Based on this result, we postulate that overexpression of daf-31 may further increase the lifespan of daf-2 mutants.
To test this, daf-2(e1370) and daf-16(mgDf47);daf-2(e1370) mutants overexpressing daf-31 were constructed. The overexpression of daf-31 was confirmed by qRT-PCR (Figure S4). As shown in Figure 3D, daf-31 overexpression increased the lifespan of daf-2 mutant worms (p<0.0001, log-rank test) (Figure 3D and Table S1). The mean lifespan was increased by eight days and the maximum lifespan was extended by seven days (Figure 3D and Table S1). This increased lifespan was due to daf-31 overexpression as daf-31 RNAi treatment completely abrogated it (Figure 3E and Table S1). Interestingly, daf-31 overexpression failed to extend the lifespan of daf-16;daf-2 double mutants (Figure 3F and Table S1), indicating that DAF-16 is required for daf-31 overexpression to enhance the daf-2 longevity phenotype. We also measured the lifespan of N2 animals overexpressing the daf-31 gene. As shown in Figure S5 and Table S1, daf-31 overexpression did not extend the lifespan of N2 worms. In fact, it slightly decreased the lifespan of N2 worms. Finally, to confirm daf-31 functions through daf-16 in C. elegans lifespan regulation, we tested if daf-31 RNAi can further decrease the lifespan of RNAi-sensitive daf-16 mutants (daf-16;rrf-3). We found daf-31 RNAi had no obvious effect on the lifespan of daf-16;rrf-3 mutants (Figure S6 and Table S1).
The forkhead transcription factor DAF-16/FOXO controls the transcription of an array of genes essential for lifespan extension and oxidative stress resistance including the antioxidant enzyme superoxide dismutase (sod)-3 gene and beta-carotene 15,15′-monooxygenase gene (bcmo-2) [22]. We used qRT-PCR to measure the expression level of sod-3 and bcmo-2 in daf-31 overexpression strains and control animals. As shown in Figure 4A, the expression level of sod-3 was significantly increased when daf-31 was overexpressed in the N2 background and in daf-2 mutants. Similar to sod-3, the expression of bcmo-2 is also significantly increased when daf-31 is overexpressed in the daf-2 mutant background (Figure 4B). Taken together, these data indicate overexpression of daf-31 stimulates the transcriptional activity of DAF-16.
Reduction of daf-2 insulin-like signaling activity increases C. elegans lifespan by promoting nuclear localization of DAF-16 [23], [24]. We crossed the DAF-16::GFP reporter gene into daf-31 overexpressing animals to examine if daf-31 overexpression influences the subcellular localization of DAF-16. We found the percentage of animals showing DAF-16 nuclear localization was not significantly different between daf-31 overexpressing animals and control worms (Figure 4C and D).
daf-2 mutants are resistant to environmental stresses such as high temperature [25], [26]. We examined if daf-31 overexpression could enhance the thermotolerance of daf-2 mutants. As reported previously [25], the survival rate of daf-2 mutants at 35° is doubled compared to N2 worms (Figure S7 and Table S2). However, the mean survival for N2 and N2 overexpressing daf-31 was similar (9.8 hours vs. 10 hours) (p = 0.2420, log-rank test) (Figure S7 and Table S2). Similarly, daf-31 overexpression did not increase the survival of daf-2 mutants at 35° (p = 0.4623, log-rank test) (Figure S7 and Table S2).
We tested the influence of daf-31 overexpression on the reproduction of N2 and daf-2 mutant animals. daf-31 overexpression increased the brood size of N2 animals and daf-2 mutants by about 23% and 30%, respectively (Figure S8). While daf-31 overexpression increased the daf-2 mutant lifespan in a daf-16 dependent manner, daf-31 overexpression increased the reproduction of daf-2 mutants significantly in a daf-16 independent way. Overexpression of daf-31 increased the brood size of daf-16;daf-2 mutants by 40% (P<0.01, t-test) (Figure S8).
We demonstrated that daf-31 mutants form dauer-like larvae that share some characteristics of wild-type dauer larvae such as fat accumulation. Many daf genes, especially those from the insulin-like signaling pathway, are involved in the regulation of lifespan [27]. Mutations in daf-2, which encodes an insulin/IGF-1 receptor [9], convey a temperature-sensitive Daf-c phenotype, and the adults live twice as long as wild-type animals [9], [28], [29]. Mutations in some genes downstream of daf-2, such as age-1 and pdk-1, also extend lifespan [30], [31]. Conversely, mutations in other downstream genes, including daf-18 and daf-16, shorten lifespan [32]–[34].
We examined whether daf-31 is also involved in aging and found that daf-31 partially mediates the effect of reduced daf-2/IGF signaling pathway on C. elegans lifespan. Moreover, overexpression of daf-31 enhances the longevity phenotype of daf-2 mutants depending on the activity of DAF-16. Supporting this lifespan data, the expression levels of sod-3 and bcmo-2, the transcriptional targets of the DAF-16 FOXO3 transcription factor, are up-regulated in the daf-31 overexpression strains. Thus, it is reasonable to argue that DAF-31 regulates C. elegans lifespan by influencing DAF-16 transcriptional activity and daf-31 overexpression stimulates DAF-16 activity. Indeed both DAF-31 and DAF-16 are expressed in neurons and intestine, two major tissues essential for regulation of C. elegans lifespan by the DAF-2/IGF signaling pathway [35]–[37]. However, overexpression of daf-31 has no influence on the subcellular localization of DAF-16. It is consistent with the lifespan data that overexpression of daf-31 does not increase the lifespan of N2 animals. Thus, overexpression of DAF-31 only extends C. elegans lifespan in the daf-2 mutant background in which DAF-16 has entered the nucleus due to inhibition of the IGF signaling.
Previous studies show that the stress-resistance phenotype can be uncoupled from the longevity phenotype [38]. Indeed, although daf-31 overexpression further increases the long-lived lifespan of daf-2 mutants, it has no effect on the thermotolerance of daf-2 mutants. Interestingly, daf-31 overexpression increases the reproduction of both wild-type animals and daf-2 mutants, which is not dependent on DAF-16, suggesting DAF-31 functions through DAF-16 for lifespan regulation but not for reproduction. Since DAF-16 is required for the stress resistance of daf-2 mutants, it is likely that daf-31 overexpression extends the daf-2 mutant lifespan through DAF-16-dependent mechanisms other than increasing stress-resistance. DAF-31 is found in multiple tissues including neurons. It is known that many C. elegans neurons are refractory to RNAi treatment in wild-type background [39]. It is possible that neuronal DAF-31 activity is more important for lifespan regulation because daf-31 RNAi treatment only shows influence on lifespan of RNAi-sensitive mutants. Supporting this assumption, it has been reported that daf-16/FOXO activity in neurons accounted for only 5–20% of the lifespan extension seen in daf-2 mutants [37]. Since DAF-31 may only influence DAF-16 activity in neurons, its overexpression only increases the daf-2 mutant lifespan modestly.
We cloned the daf-31 gene and sequence analysis indicates DAF-31 is a worm ortholog of ARD1 that was first identified in yeast [40]. ARD1 is the catalytic subunit of the major NatA that transfers an acetyl group from acetyl coenzyme A to the N-terminal of nascent polypeptides. Yeast ARD1 mutants fail to enter stationary phase and sporulate during nitrogen deprivation [40]. The yeast stationary phase is comparable to C. elegans dauer stage and is essential for survival when nutrients are limited. C. elegans enters dauer stage during starvation or under high concentration of pheromone. Our data show daf-31 mutants could not complete dauer morphogenesis in response to pheromone and starvation, which indicates daf-31 is required for dauer formation. Thus, both yeast ARD1 and worm DAF-31 play an important role in the developmental switch in response to the environmental nutrient limitation. Additionally, daf-31 mutants could not grow to fertile adults in an environment with abundant food suggesting its essential role in normal development. Similar to our observation, it has been reported that loss of Ard1 is lethal for D. melanogaster and affects cell survival or proliferation, indicating ARD1 is required for D. melanogaster development [41]. In addition to developmental arrest, the daf-31 mutants shift metabolism to fat accumulation. Interestingly, yeast ARD1 mutants not only fail to enter stationary phase but also do not accumulate as much carbohydrates as wild-type yeast strains [40]. Thus, the function of ARD1 in regulating development and metabolism appears conserved from yeast to C. elegans.
N-terminal acetylation is one of the most common posttranslational protein modifications. It is estimated to occur on 50% of yeast proteins [20], 71% of D. melanogaster cytosolic proteins [20] and 84% of human proteins [42]. NatA plays the most prominent role in N-terminal acetylation. It would be interesting to know whether the pleiotropic phenotypes of ard1 mutants result from global changes of protein N-acetylation or from acetylation status of specific protein substrates. It has been reported that human ARD1 directly acetylates β-catenin and enhances its transcriptional activity [43]. We show that overexpression of DAF-31 stimulates the transcriptional activity of DAF-16. It would be interesting to examine if DAF-31 overexpression acetylates DAF-16. Alternatively, a suppressor screening of daf-31 mutants may help to identify the essential substrates of the DAF-31 acetyltransferase and contribute to understanding the mechanisms by which ARD1 influences development, metabolism and aging. Moreover, emerging evidence has revealed that abnormal regulation of ARD1 is associated with tumorigenesis and ARD1 represents a novel cancer drug target [44], [45]. Identification of DAF-31 substrate proteins may uncover new therapeutic targets of cancer diseases.
All strains were grown on NG agar plates seeded with E. coli strain OP50 [46]. Mutations are listed by linkage groups as follows: LG I: daf-16(mgDf47); LG II: rrf-3(pk1426); LG III: daf-2(e1370); LG IV: unc-24(e138), daf-31(m655); LG X: daf-3(e1376), daf-12(m20), daf-12(rh61rh411). All mutants are derived from the wild-type Bristol N2 strain.
To make the daf-16(mgDf47)I; daf-2(e1370)III double mutant, daf-2(e1370) males were mated with daf-16(mgDf47) hermaphrodites. Ten F1 adults (daf-16/+; daf-2/+) were incubated at 25°C. F2 dauer larvae (either +/+; daf-2 or daf-16/+; daf-2) were transferred to a fresh plate at 15°C for recovery. Then adults were shifted to 25°C. Since daf-16(mgDf47) can suppress the daf-2(e1370) Daf-c phenotype, non-dauer adults from the next generation were daf-16(mgDf47); daf-2(e1370) double mutants.
Double mutants were constructed for epistatic tests between daf-31 and daf-d mutants at 20°. However, daf-16 RNAi was used for epistasis analysis between daf-31 and daf-16. daf-16 RNAi fully suppressed dauer formation of daf-2(e1370) mutants. daf-12(m20) and daf-12(rh61rh411) mutations were used to construct the strain+daf-31(m655)/unc-24(e138)+; daf-12(m20) and +daf-31(m655)/unc-24(e138)+; daf-12(rh61rh411) using standard genetic methods. The daf-12(rh61rh411) mutation was confirmed by sequencing. +daf-31(m655)/unc-24(e138)+; daf-3(e1376) was constructed to determine the epistatic relationship between daf-31 and daf-3. As the daf-31(m655) mutation was not marked by a genetic mutation in daf-31;daf-3 and daf-31;daf-12 mutant worms, the daf-31(m655) deletion mutations were confirmed by using single worm PCR. Representative gel pictures are shown in Figure S9. Injection of RNAi was used to inhibit the daf-15 gene in unc-24(e138)daf-31(m655)/nT1 to examine the epistatic relationship between daf-15 and daf-31.
To construct the daf-31 overexpressing strains, the full-length daf-31 genomic DNA, including its native 760-bp promoter and 3′-UTR was cloned into pGEM-T (Promega); this construct successfully rescued the daf-31 dauer-like mutant phenotype. Multiple copies of the construct were integrated into chromosomal DNA by γ-irradiation to make an N2 transgenic line overexpressing daf-31. Then the daf-31 overexpressing chromosome was introduced into both daf-2(e1370) and daf-16(mgDf47);daf-2(e1370) mutants by genetic crosses.
To make a daf-31 promoter-GFP transcriptional fusion, the 760-bp daf-31 promoter was inserted into the GFP vector pPD95.70 (a gift from Dr. Andrew Fire at Stanford University) between the PstI and BamHI sites. The construct was injected into N2 adults at a concentration of 100 µg/ml. pRF4, which encodes a mutant collagen and induces a dominant roller (Rol) phenotype, was co-injected at the same concentration as a transformation marker. Rol animals were selected from the F2 generation and used to establish stable transgenic lines.
To evaluate the subcellular location of DAF-16, TJ356 (zIs356 [daf-16p::daf-16a/b::GFP+rol-6]) males were crossed to daf-31 overexpressing hermaphrodites. The roller progeny were mounted on 2% agar pads to examine DAF-16::GFP subcellular localization.
To stain fat using Sudan Black B, N2, daf-2(e1370) and daf-31(m655)IV/nT1[unc-?(n754) let-?](IV;V) synchronized L1 larvae were placed on NG agar plates, incubated at 20°C until they entered L3 or L4 stages, collected and washed two to three times with M9 buffer. Paraformaldehyde stock solution (10%) was added to a final concentration of 1%. The samples were frozen in dry ice/ethanol and then thawed under a stream of warm water. After a total of three freeze-thaw cycles, the worms were stained with Sudan Black B as described by Kimura et al. [9].
Nile red staining of fixed worms was performed as described by Pino et al. [47]. Worm samples were collected and washed twice with M9 buffer. After the final wash, worms were fixed in 40% isopropanol at room temperature for three minutes. The fixed worms were stained in Nile red/isopropanol solution for 30 minutes at room temperature with gentle rocking. The stained worms were washed once with 1 ml M9 buffer and mounted on a 2% agarose pad for microscopy under the fluorescence channel. In order to compare the fat content in different strains, the pictures were taken at the same camera setting under 20× magnification.
Three-factor-mapping with SNP markers and cosmid rescue were performed as previously described [18]. To determine the mutation in daf-31(m655), the K07H8.3 gene (GenBank accession # NM_068991.4) was amplified using primers 5′- GTG AGT CGA AAC CCA TTT TG -3′ and 5′- GAA TGA ACC AGT TGG AAA AGG -3′ from both N2 and daf-31(m655) mutant homozygotes. PCR products were cloned into the pGEM-T vector (Promega) following the manufacturer's instructions. T7 primer 5′- GTA ATA CGA CTC ACT ATA GGG -3′ and SP6 primer 5′- TAC GAT TTA GGT GAC ACT ATA G -3′ were used in DNA sequencing reactions.
Part of the daf-31 coding region was amplified from C. elegans genomic DNA using primers 5′- CGG GAT CCA TTC GTT GTG CTC GCG TG -3′ and 5′- CCC AAG CTT GCA GTG GTA TAG GCC TC -3′. The PCR products were then purified and cloned into the feeding RNAi vector L4440 (Addgene) between the BamHI and HindIII sites. The RNAi construct was transformed into E. coli HT115 (DE3) and RNAi feeding was performed as previously described [48].
To inhibit daf-15 and daf-31 genes by injection of RNAi, a 1 kb daf-15 cDNA fragment and the full-length daf-31 cDNA were cloned into pGEM-T vector (Promega), respectively. The gene identity was confirmed by sequencing. The Riboprobe Combination System-SP6/T7 (Promega) was used to transcribe RNA in vitro according to the manufacturer's protocol. Double-stranded RNA was synthesized and injected as described by Fire et al. [49].
Synchronized N2 L1 larvae were treated with either control (empty) vector or daf-31 RNAi by feeding as previously described [48]. Day 1 adult animals were collected for total RNA extraction using the Trizol kit (Zymo). Synchronized L1 larvae of daf-31 overexpressing strains were allowed to grow on OP50 food plates. Day 1 adult animals were collected for total RNA extraction using the Trizol reagent (Zymo). The first strand cDNA was synthesized using the ImProm-II reverse transcription system (Promega). SYBR green dye (Quanta) was used for qRT-PCR to measure the expression level of daf-31, sod-3 and bcmo-2 in corresponding worm samples. Reactions were performed in triplicate on an ABI Prism 7000 real-time PCR machine (Applied Biosystems). Relative-fold changes were calculated using the 2−ΔΔCT method. The primers used for qRT-PCR were: daf-31, 5′- GAA GAT CAC AAG GGA AAT GTT G -3′ and 5′- CTC TTG CGG TCT GAT CCA TC -3′; act-1, 5′- CAA TCC AAG AGA GGT ATC CTT ACC CTC -3′ and 5′- GAG GAG GAC TGG GTG CTC TTC -3′; bcmo-2, 5′- GCC GAT TTA GAG AAC GGA GAT CAC -3′ and 5′- TGA GAA TTC CGT CAT CTT CCC GA -3′; sod-3, 5′- GGA ATC TAA AAG AAG CAA TTG CTC -3′ and 5′- CGC GCT TAA TAG TGT CCA TCA G -3′.
About 120–150 L4 larvae raised at 20°C were transferred to ten NG agar plates (twelve to fifteen animals per plate spread with either OP50 or RNAi food) and incubated at 25°C. The first day of adulthood is day 1 in the survival curves. During the reproductive period, adult animals were transferred daily to fresh plates. Thereafter, animals were transferred every ten days (OP50 food) or every six days (RNAi food). Animals were scored as alive, dead, or lost every other day. Animals that do not move in response to touching were scored as dead. Animals that died from causes other than aging, such as sticking to the plate walls, internal hatching or bursting in the vulval region, were scored as lost. GraphPad Prism was used for statistical analysis and generation of survival curves. For the thermotolerance experiment, day 1 adult animals were incubated at 35°C and survival was scored as described above. To measure reproduction of worms, L4 larvae growing at 20°C were transferred daily to fresh plates and the progeny were counted.
|
10.1371/journal.pntd.0005829 | Heterologous expression of the antimyotoxic protein DM64 in Pichia pastoris | Snakebite envenomation is a neglected condition that constitutes a public health problem in tropical and subtropical countries, including Brazil. Interestingly, some animals are resistant to snake envenomation due to the presence of inhibitory glycoproteins in their serum that target toxic venom components. DM64 is an acidic glycoprotein isolated from Didelphis aurita (opossum) serum that has been characterized as an inhibitor of the myotoxicity induced by bothropic toxins bearing phospholipase A2 (PLA2) structures. This antitoxic protein can serve as an excellent starting template for the design of novel therapeutics against snakebite envenomation, particularly venom-induced local tissue damage. Therefore, the aim of this work was to produce a recombinant DM64 (rDM64) in the methylotrophic yeast Pichia pastoris and to compare its biological properties with those of native DM64. Yeast fermentation in the presence of Pefabloc, a serine protease inhibitor, stimulated cell growth (~1.5-fold), increased the rDM64 production yield approximately 10-fold and significantly reduced the susceptibility of rDM64 to proteolytic degradation. P. pastoris fermentation products were identified by mass spectrometry and Western blotting. The heterologous protein was efficiently purified from the culture medium by affinity chromatography (with immobilized PLA2 myotoxin) and/or an ion exchange column. Although both native and recombinant DM64 exhibit different glycosylation patterns, they show very similar electrophoretic mobilities after PNGase F treatment. rDM64 formed a noncovalent complex with myotoxin II (Lys49-PLA2) from Bothrops asper and displayed biological activity that was similar to that of native DM64, inhibiting the cytotoxicity of myotoxin II by 92% at a 1:1 molar ratio.
| Snakebite envenomation causes medical emergencies that, depending on the species responsible for the bite, involve different organs and tissues. Envenomation by snakebite is a worldwide problem, and Brazil presents a high incidence of Bothrops bites. Bothrops venoms cause pathological alterations with prominent local effects, such as edema, blistering, hemorrhage, dermonecrosis and myonecrosis, usually followed by poor tissue regeneration and permanent sequelae. Bleeding, coagulopathy, cardiovascular shock and renal failure are typical systemic effects of these venoms. The clinical treatment for snakebite envenoming is intravenous administration of the specific antivenom. However, serotherapy does not efficiently protect against local tissue damage. Additional challenges faced by classical antivenom therapy include the wide antigenic variation of venoms across species and even within the same snake species and the frequent occurrence of adverse reactions that are associated with the administration of immunobiologicals. The development of new effective toxin inhibitors based on the structure of natural antiophidic proteins is an attractive therapeutic alternative. DM64 is a myotoxin inhibitor that was isolated from opossum serum, and its expression as a recombinant protein is paramount to the characterization of its structure-function relationship, an essential step toward the development of alternative strategies to better manage bothropic snakebite envenomations.
| Accidents involving venomous snakes are medical emergencies that are often neglected in many tropical and subtropical countries [1]. Several epidemiological studies have tried to estimate the true burden of snakebite envenoming in the world. Overall, they have reported up to 5 million snake bites/envenoming per year, including tens of thousands of deaths and a much larger number of victims that are left with permanent sequelae [2–6]. The number of cases of snakebite envenomation is highest in rural regions and in cities that border on forests. Brazil also has a high level of snakebite accidents, most of which involve four predominant genera; Bothrops is the genus that is held accountable for the highest number of accidents [7]. According to the Brazilian Ministry of Health’s Notifiable Disease Information System (SINAN), 53,068 (provisional) Bothrops snakebite accidents occurred between 2013 and 2015 in the country [8].
Bothrops venoms contain complex mixtures of toxins that can cause the degradation of vascular basement membrane components and myonecrosis, resulting in local bleeding and tissue damage. In severe cases of envenoming, systemic bleeding, shock, hypotension and/or kidney injury may be observed, leading to high morbidity and mortality [9–11]. For longer than one century, snake envenomation has been treated using antivenoms that are based on horse antibodies. However, this remedy does not prevent local damage caused by some venomous snakes, and the antibodies can induce early or late adverse reactions [12].
The application of biochemical methods to the study of venoms has associated their pathological activities with proteins and peptides. Therefore, the study of venom proteomes (i.e., venomes) is regarded as one of the best approaches for characterizing snake venom compositions [13]. Several studies have identified and determined the relative abundances of certain classes of toxins in different snake venoms. It has also been reported that protein and peptide groups exhibit several kinds of activity on different targets, such as the cardiovascular and nervous system, blood components, and muscular and endothelial tissue [14–18]. Venomics studies of Bothrops have identified two of the most abundant protein groups: metalloproteases and phospholipases A2 (PLA2s). These protein groups are responsible for the most severe local clinical manifestations that are produced by these venoms, such as hemorrhage and muscular and endothelial damage [14,19–22].
Phospholipases A2 from snake venom have evolved into potent toxins that exhibit diverse activities such as neurotoxic, myotoxic, anticoagulant, hypotensive, cardiotoxic and edema-inducing [23]. PLA2s belong to one of the five principal groups that catalyze the Ca2+-dependent hydrolysis of the acyl ester at the sn-2 position of glycerophospholipids. Basic myotoxic phospholipases A2 are responsible for tissue degradation and necrosis at the bite site. These myotoxins bind to the plasma membrane of skeletal muscle cells, generating muscle necrosis. Some PLA2s contain a critical substitution at the calcium-binding site (Asp49 to Lys49) that renders them non-catalytic, yet they conserve their myotoxic activity [24–29].
Many studies have sought alternative sources of natural venom inhibitors to complement the action of antivenoms, particularly to help neutralize local tissue damage. Some reports have identified some natural components that inhibit PLA2 myotoxins from several snake venom species [30–33]. Additionally, several studies have investigated myotoxic inhibitors that have been isolated from reptile or mammalian sera [34–36]. DM64 is a myotoxin-specific inhibitor isolated from D. aurita (opossum) serum and has been characterized as an acidic glycoprotein of 64 kDa in size. DM64 has been structurally classified as a member of the immunoglobulin supergene family, showing five immunoglobulin-type domains that are similar to α1B-glycoprotein [35,37]. DM64 forms a noncovalent soluble complex and efficiently inhibits the myotoxic activity of both inactive Lys49-PLA2 and active Asp49-PLA2 but does not inhibit the catalytic activity of the latter [35].
Currently, many alternative treatments are being developed to supplement antivenom serum treatments and prevent local tissue damage. Several newly discovered natural antiophidic molecules from plant extracts have been reported [30–33]. However, their pharmacological reliability has not yet been demonstrated because these inhibitors have different functional groups that could interact with different molecular targets[30,32,33]. Myotoxin inhibitors isolated from snake and mammalian sera seem to be more specific, making their recombinant expression a promising strategy for new therapeutic developments [35–38].
DM64 is an efficient myotoxin inhibitor that was isolated from D. aurita serum, and the heterologous expression of this glycoprotein could be a breakthrough in the development of Bothrops envenomation treatment. Thus, the aim of this study was to express rDM64 using Pichia pastoris. The biological activity of the recombinant inhibitor on the myotoxin Lys49-PLA2 from B. asper was analyzed.
Yeast extract, peptone, biotin, dextrose, agar, peroxidase-conjugated anti-rabbit secondary antibody, Pefabloc SC (4-(2-aminoethyl)-benzenesulfonyl fluoride), sorbitol, EDTA (ethylenediaminetetraacetic acid) and DAB (3,3’-diaminobenzidine) substrate kit were purchased from Sigma (Missouri, USA). Glycerol and methanol were supplied by VETEC (Rio de Janeiro, Brazil). Pichia pastoris X-33, pPICZαA, and Zeocin were purchased from Invitrogen (California, USA). Restriction enzymes and PNGase F were furnished by New England Biolabs (Massachusetts, USA). Electroporation cuvettes, TMB (3,3’,5,5’-tetramethylbenzidine) EIA substrate kit and low-range SDS-PAGE standards were obtained from Bio-Rad Laboratories (California, USA). Trypsin was purchased from Promega (California, USA). Pierce Glycoprotein Staining kit, DMEM (Dubelcco’s Modified Eagle Medium), and fetal bovine serum were purchased from Thermo Fisher Scientific (Massachusetts, USA).
The DM64 gene sequence [35] was synthesized by Epoch Life Science (Missouri, USA). The gene was cloned without a signal peptide into the expression vector pPICZαA, which contained an alcohol oxidase 1 (AOX1) promoter. The gene sequence was cloned in-frame with the S. cerevisiae α-factor secretion sequence that is present in pPICZαA. A stop codon was inserted before the c-myc epitope and the polyhistidine (6x His) tag. The resulting plasmid (named pPICZαA-DM64) was linearized with SacI (a site present in the AOX1 promoter) prior to its transformation into P. pastoris X-33 cells via electroporation.
Transformed cell suspensions were diluted to guarantee the growth of the well-separated colonies on the surface of a solid medium. The colonies were grown in YPDS (1% yeast extract, 2% peptone, 2% dextrose, 1 M sorbitol, and 2% agar) medium containing 0.1 mg/mL Zeocin at 30°C for 3 to 5 days. A second selection was then performed in liquid YPD medium containing 0.1 mg/mL Zeocin at 30°C, 250 rpm for 24 hours. Fourteen clones that grew under these last conditions were submitted to a second screening by using increasing concentrations of antibiotics (0.2, 0.4, 0.6, 0.8, and 1 mg/mL Zeocin). The cells were grown at 30°C and 900 rpm, for 48 hours.
Pichia pastoris with the multicopy expression vector pPICZαA-DM64 was initially grown in BMGY medium, which contains glycerol as a carbon source (1% yeast extract, 2% peptone, 1.34% YNB, 4 x 10−5% biotin, 100 mM potassium phosphate, pH 6, and 1% glycerol) at 30°C and 250 rpm. The culture grew until 50-fold diluted aliquots reached an OD600nm of 0.36, which may have taken up to 24 hours. The culture medium was then changed to BMMY, which contains methanol for the purposes of induction and to be used as a carbon source (1% yeast extract, 2% peptone, 1.34% YNB, 4 x 10−5% biotin, 100 mM potassium phosphate, pH 6, and 1% methanol). The culture flask was then incubated at 30°C and 250 rpm, for 144 and 264 hours. One percent methanol was added every 24 hours to maintain protein expression. Yeast culture with 0.2 mM Pefabloc was added to the BMMY medium every 24 hours. The culture was centrifuged at 5,000xg for 20 minutes at 4°C, and the supernatant was collected and stored at -20°C. Aliquots of the culture were analyzed by using 12% SDS-PAGE gels [39] that were stained with silver nitrate [40]. The following low-range SDS-PAGE standards were used: phosphorylase b (97.4 kDa), serum albumin (66.2 kDa), ovalbumin (45 kDa), carbonic anhydrase (31 kDa), trypsin inhibitor (21.5 kDa), and lysozyme (14.4 kDa). Molecular mass estimates were calculated using Image Master 2D Elite software (GE Healthcare, version 4.01).
The expression medium (50 μL) was alkalized by adding 1 M sodium carbonate buffer (10 μL), and each aliquot that was taken at a different time after induction (0–264 hours) was incubated at 37°C for 2 hours in a 96-well polystyrene plate. The wells were washed three times with wash buffer (PBS buffer and 0.1% (v/v) Tween 20). Then, the wells were blocked with PBS buffer containing 5% (w/v) non-fat dry milk and incubated at 37°C for 2 hours. The wells were washed three times with wash buffer containing 5% (w/v) non-fat dry milk and then incubated for 1 hour at 37°C with fresh wash buffer containing 5% (w/v) non-fat dry milk and 100 μL of polyclonal anti-DM64 antibodies (0.23 mg/mL), prepared as previously described [41], except for the use of DM64 as antigen, instead of anti-hemorrhagic proteins. After this incubation, the wells were washed three times with wash buffer and then incubated at 37°C for 1 hour with 100 μL of peroxidase-conjugated secondary anti-rabbit antibodies (1:30,000) in wash buffer that contained 5% (w/v) non-fat dry milk. Finally, the wells were washed three times with wash buffer and then incubated for 20 minutes at room temperature with 100 μL of the Single Component TMB (3,3’,5,5’-tetramethylbenzidine) EIA Substrate kit. The reaction was stopped by the addition of 1 N H2SO4, and the absorbance at 450 nm was recorded. Native DM64 (0.1, 0.2, 0.4, 0.8 and 1.6 μg/50 μL) was used to build the standard curve.
Bands in silver-stained SDS-PAGE gels were digested as previously described [42], with modifications. They were first incubated twice with 100% (v/v) acetonitrile for 15 minutes and then dried under vacuum for 15 minutes. The bands were reduced by incubating the samples with 65 mM 1,4-dithiothreitol at 56°C for 30 minutes. The reduction buffer was removed, and the bands were washed twice in 100 mM ammonium bicarbonate for 10 minutes, followed by a 5 minute wash in 100% acetonitrile. The bands were dried under vacuum for 15 minutes; 20 ng/μL trypsin was added, and they were then incubated for 45 minutes on ice. Excess trypsin was removed, and 40 mM ammonium bicarbonate was added. Hydrolysis proceeded overnight at 37°C and then for an additional 45 minutes at 56°C. Finally, the digested products were desalted using tip columns that were packed with Poros R2 resin (Applied Biosystems) and equilibrated with 0.1% (v/v) formic acid in water. After washing away nonbound material (10 x 20 μL) by using equilibrium buffer, the peptides were eluted using 0.1% (v/v) formic acid in 50% (v/v) acetonitrile and dried under vacuum. Desalted tryptic peptides were redissolved in 1% (v/v) formic acid, and 4 μL of each sample was loaded onto a home-made capillary guard column (2 cm x 100 μm i.d.) that was packed with 5 μm, 200 Å Magic C18 AQ matrix (Michrom Bioresources, Auburn, CA, USA). Peptide fractionation was performed on an analytical column (10 cm x 75 μm i.d.) with a laser pulled tip (~5 μm) that was packed with the same matrix and coupled to an Ultimate 3000 RSLCnano chromatography system (Thermo Fisher Scientific, Waltham, MA, USA). The analysis was conducted using an LTQ Orbitrap XL mass spectrometer (Thermo Fisher Scientific, Waltham, MA, USA) with a capillary temperature of 200°C, tube lens voltage of 100 V and a nanoESI source with a spray voltage set to 1.9 kV and no sheath gas. Samples were added with water containing 0.1% (v/v) formic acid into the trap column at 2 μL/min, while the chromatographic separation occurred at 0.2 μL/min. The peptides were eluted with a 2–40% (v/v) acetonitrile gradient in 0.1% (v/v) formic acid over 32 minutes, which then ramped to 80% acetonitrile in 4 minutes and was followed by a final washing step at 80% acetonitrile for an additional 2 minutes. The mass spectrometer operated in data-dependent mode, using the following settings: for MS1, a 300–1700 m/z scan range, 1 x 106 automatic gain control (AGC), 500 ms maximum injection time (IT), centroid mode acquisition and resolution of 60,000 (FWHM at m/z 400). Up to 7 of the most intense precursor ions in each survey scan were selected for CID fragmentation in the LTQ using 35% normalized collision energy (NCE). MS2 analysis was performed with the following parameters: 1 x 104 AGC, 100 ms IT, and centroid mode acquisition. The isolation window was 2 m/z, and only precursor ions with a charge state ≥ 2 were selected for fragmentation, setting the dynamic exclusion to 45 s. The spectrometer was calibrated using a calibration mixture composed of caffeine, peptide MRFA, and Ultramark 1621, as recommended by the instrument manufacturer.
The Pichia pastoris sequence database was obtained from UniProt (Proteome ID UP000000314, containing 5,073 protein sequences). The sequences of DM64 (UniProt Q8MIS3, excluding the signal peptide), DM43 (UniProt P82957) and common contaminants (ftp://ftp.thegpm.org/fasta/cRAP) were also included in the protein database. Peaks Studio software (version 7.5) was used for de novo sequencing assisted database search [43] using the following parameters: monoisotopic masses, carbamidomethylation of cysteine (fixed modification), oxidation of methionine (variable modification), semi-tryptic digestion, 10 ppm peptide mass error tolerance, 0.6 Da fragment mass tolerance, up to 2 variable modifications per peptide and a maximum of 2 missed cleavages. The data were filtered using the PEAKS decoy fusion approach, and false discovery rates (FDR) were set to a maximum of 1% at the peptide level. Only proteins that were identified with at least 2 unique peptides were accepted (FDR values at the protein level ≤1%).
Expression products in the supernatant and purified rDM64 were separated by 12% SDS-PAGE gel [39], and the proteins were transferred to a 0.45 μm nitrocellulose membrane (GE Healthcare Life Science, USA) using ice-cold transfer buffer (25 mM Tris-HCl, pH 8, 192 mM glycine, and 20% methanol). The membrane was blocked in TBS-Tween (25 mM Tris-HCl, pH 8.0, 140 mM NaCl, 2 mM KCl, and 0.05% Tween 20) containing 5% (w/v) non-fat dry milk overnight at 4°C. The membrane was then washed three times in TBS-Tween and was incubated for 2 hours at room temperature with diluted [1:100 (v/v)] crude rabbit serum that was raised against native DM64 [41]. The membrane was subsequently washed three times in TBS-Tween and was incubated with 1:5000 (v/v) peroxidase-conjugated secondary anti-rabbit antibody. The rDM64 bands were visualized using a DAB substrate kit. The following prestained low-range SDS-PAGE standards were used: phosphorylase B (103 kDa), bovine serum albumin (77 kDa), ovalbumin (50 kDa), carbonic anhydrase (34 kDa), and soybean trypsin inhibitor (28.8 kDa).
The expression medium that contained the recombinant protein was concentrated using Amicon Ultra 15 mL centrifugal filters (Merck Millipore, USA) with a cutoff of 10 kDa. The expression medium was exchanged to buffer (20 mM Tris-HCl pH 7.5), and the sample was injected into a column that had already been equilibrated with the same buffer. This column, a HiTrap NHS (7 x 25 mm, GE Healthcare Life Sciences, USA) activated column that contained myotoxin II (from B. asper) that had been immobilized according the manufacturer’s instructions, was connected to an ÄKTA Purifier chromatography system (GE Healthcare Life Sciences, USA). The rDM64 fraction was eluted at a flow rate of 1 mL/min with 0.1 M glycine, pH 2.7, and collected over 1 M Tris to neutralize the solution. The fraction was concentrated using Amicon Ultra 4 mL centrifugal filters (Merck Millipore, USA) with a cutoff of 10 kDa. The recombinant protein fraction buffer was exchanged to the equilibration buffer of a Mono Q GL column (5 x 50 mm, GE Healthcare Life Sciences, USA) (20 mM Tris-HCl, pH 7.5). The recombinant protein was eluted during a 20 minute 0–1 M NaCl linear gradient at a flow rate of 0.5 mL/min. Protein concentrations in the collected fractions were determined using the bicinchoninic acid assay (BCA method) with BSA as standard [44]. Their electrophoretic profiles were analyzed by using SDS-PAGE under reducing conditions [39].
PNGase F was used to cleave between the innermost N-acetylglucosamine (GlcNAc) residues and the asparagine residues to which they are linked in this high-mannose-type recombinant glycoprotein. Initially, 3 μg of purified rDM64 was mixed with 0.5% SDS and 40 mM DTT, and the solution was incubated at 100°C for 10 minutes. Then, 50 mM sodium phosphate, pH 7.5, 1% NP-40, and 1,000 U of PNGase F were added, and the solution was incubated at 37°C for 1 hour. The reaction was stopped by adding 5 μL of denaturing buffer (5X, final concentrations: 0.3 M Tris-HCl, pH 6.8, 2% SDS, and 0.1 M DTT), and the separation of the reaction products was visualized using 12% SDS-PAGE [39] gel with silver staining [40]. Native DM64 was used as a positive control in this experiment.
The presence of sugars in rDM64 was determined with the periodic acid-Schiff method using the Pierce Glycoprotein Staining kit. The staining of 5 μg of rDM64 was done following the manufacturer’s instructions. Native DM64 was used as a positive control in this experiment. All samples were analyzed using 12% SDS-PAGE gels [39].
The interaction between myotoxin II (PLA2-Lys49) from B. asper and rDM64 was monitored by using a 12% native PAGE gel [39]. The myotoxin-rDM64 complex (2:1 molar ratio) was incubated in 20 mM Tris-HCl, pH 7.5, and 150 mM NaCl at 37°C for 30 minutes. Native DM64 was used as a positive control in this experiment and the gel was stained with silver nitrate [40].
Murine myoblast cell line C2C12 (ATCC CRL-1772), which can fuse and differentiate into myotubes, was used. The cells were grown in Dulbecco’s Modified Eagle Medium (DMEM) that was supplemented with 44 mM sodium bicarbonate, 19.5 mM glucose, 2 μM L-glutamine, 100 U/mL penicillin, 0.1 mg/mL streptomycin, and 10% fetal bovine serum (FBS) in a humidified atmosphere with 5% CO2 at 37°C. Cells were harvested from near-confluent cell monolayers grown in 25 cm2 bottles. After detaching the cells by using 1500 U/mL trypsin containing 5.3 mM EDTA for 2 minutes at 37°C, the resuspended cells were seeded in 96 wells at an approximate initial density of 1 x 104 cells/well in the same medium. Upon reaching near confluence in 3 days, the growth medium was replaced with a differentiation medium (DMEM supplemented with 1% FBS). When multinucleated myotube cells were observed (after 6 days of culture), they were utilized in a cytotoxicity assay [25]. Myotoxin II and recombinant or native DM64 were preincubated (1:1, 2:1 and 4:1, mol:mol) for 30 minutes at 37°C in 150 μL of DMEM supplemented with 1% FBS. After aspirating the old medium, the samples were added to the cell cultures that were growing in 96 wells to yield a total volume of 150 μL/well. After three hours of incubation at 37°C with 5% CO2, 20 μL aliquots of supernatant were collected to determine lactate dehydrogenase activity. Controls for 0% and 100% toxicity consisted of assay medium and 0.1% Triton X-100 in assay medium, respectively. The results are presented as the mean ± standard deviation (n = 3). Statistical analyses were by one-way ANOVA followed by Student–Newman–Keuls post hoc test (GraphPad Prism 5.0 software). P-values of 0.05 or less were considered significant.
Recombinant glycoprotein production was carried out in a 1 L culture flask using a two-step growth protocol that consisted of a glycerol batch phase and a methanol fed-batch phase. During the glycerol batch phase, P. pastoris was cultivated at 30°C for 24 hours, and the cellular concentration reached approximately 40 g/L. The methanol fed-batch phase was maintained at 30°C, and the induction lasted 264 hours, yielding 288 hours of total cell growth time. Pichia pastoris cells were grown in methanol in the absence (Fig 1A) or presence (Fig 1B) of 0.2 mM Pefabloc, a serine protease inhibitor. During the methanol fed-batch phase, cell growth without Pefabloc was maximum at 72 hours (96 hours of total cell growth time). From 72 to 264 hours of methanol induction, the cell concentration was constant, around 63 g/L (Fig 1A). On the other hand, in the presence of Pefabloc, the cell growth continued after 72 hours, reaching a biomass yield of 90 g/L at 264 hours (288 hours of total cell growth time)(Fig 1B).
The use of a serine protease inhibitor during fermentation was important to not only positively stimulate P. pastoris growth but also improve rDM64 expression yields. After 264 hours of methanol induction, the rDM64 concentration in the fermentation medium increased from 0.002 g/L without Pefabloc (Fig 1A) to approximately 0.02 g/L in the presence of Pefabloc (Fig 1B). Without protease inhibitor, the recombinant protein concentration in the fermentation medium increased until 144 hours of induction (168 hours of cell growth time) were reached and decreased thereafter until 264 hours (288 hours of cell growth time)(Fig 1A). The 65 kDa band in the SDS-PAGE gel that corresponds to rDM64 shows the same expression kinetics (Fig 2A), as does immunoblotting (Fig 2C). Several bands with lower molecular masses whose intensities progressively increase during 264 hours of methanol induction can also be observed by SDS-PAGE (Fig 2A). On the other hand, the amount of rDM64 in the culture medium containing Pefabloc favored a progressive increase (Fig 1B), which corresponds to the behavior of the main 65 kDa band observed by SDS-PAGE during the same time interval; there were also smaller amounts of protein bands with lower relative molecular masses (Fig 2B).
The bands indicated in red in Fig 2A were digested in gel, and the resulting peptides were analyzed by mass spectrometry (S1 Table). Database searches using the masses of these peptides and their fragment ions led to the unequivocal identification of the 65 kDa bands (bands # 3–5) as full-length DM64. The identity of DM64 was further confirmed by Western blotting using polyclonal antibodies, as shown in Fig 2C. Additional fainter bands with higher (bands # 1–2) and lower (bands # 6–10) relative molecular masses were also identified as DM64. In this last case, partial proteolytic degradation of rDM64 seems the most likely explanation, given the reduced abundance of these bands in the presence of Pefabloc. Higher molecular mass bands (band #1: 136 kDa and band #2: 88 kDa, Fig 2A) were also identified as DM64, and two explanations may be envisaged: either there is a strongly aggregated form of rDM64 and/or the recombinant protein has been expressed with different levels of glycosylation.
The culture medium containing the recombinant inhibitor was fractionated by liquid chromatography using an affinity column conjugated with myotoxin II, a Lys49-PLA2 isolated from B. asper venom [45]. DM64 specifically binds to the immobilized myotoxin, and for this reason, this first purification step efficiently removed most impurities (culture medium proteins), allowing for the recovery of a protein fraction that was enriched in rDM64 (Fig 3A). However, the SDS-PAGE profile of the sample under reducing conditions showed that the chromatographic peak corresponding to rDM64 was still heterogeneous, with both higher and lower molecular mass protein bands co-purified with full-length rDM64 (Fig 3C). Native DM64 is an acidic protein with a pI of 4.5 [35]; thus, a second chromatography step using an anion-exchange column was necessary to improve protein purification, as shown in Fig 3 (panels B and C). rDM64 was noted to be glycosylated following positive staining with periodic acid-Schiff reagent (Fig 3D) and was recognized by polyclonal antibodies raised against native DM64 (Fig 3E).
rDM64 was treated with the glycosidase PNGase F to remove N-linked oligosaccharides (Fig 4). Due to the limited amount of homogeneous protein, a partially purified fraction that was purified by affinity column was used instead of the protein preparation that was obtained by anion-exchange chromatography. Both native DM64 (71 kDa)(Fig 4, lane 1) and rDM64 (65 kDa)(Fig 4, lane 3) proteins were submitted to carbohydrate removal. Deglycosylated rDM64 showed a main molecular band at 56 kDa (Fig 4, lane 4), which is closer to the molecular mass of native DM64 without the glycan moiety (55 kDa)(Fig 4, lane 2). The theoretical average molecular mass of DM64 based on its primary sequence is 53.3 kDa (calculated using http://web.expasy.org/compute_pi). Therefore, given the accuracy (± 10%) of molecular mass estimates by SDS-PAGE [46], the molecular masses of deglycosylated DM64 (native/recombinant) proteins are in close agreement with the expected value.
The ability of the recombinant inhibitor to bind to myotoxin II (Lys49-PLA2) from B. asper venom was analyzed. Myotoxin II (mt II) and rDM64 were mixed at a 2:1 (mol:mol) ratio, and complex formation was monitored by electrophoresis under native conditions. Fig 5 (lane 5) shows a band with an electrophoretic mobility that corresponds to the complex. Native DM64 was used as a control for complex formation (Fig 5, lane 2). Due to the basic nature of myotoxin-II (pI 9.1)[47], it does not enter the gel under native conditions (Laemmli buffer system without SDS). The band corresponding to the rDM64-mt II complex (Fig 5, lane 5) shows a different mobility than that of DM64-mt II (Fig 5, lane 2) due to the difference in the net charges of the proteins; this factor is a major variable that influences electrophoresis mobility on native/non-denaturing gels. rDM64 has a different charge than that of native DM64 due to the absence of sialic acid, which is negatively charged.
The inhibition of the cytotoxicity induced by myotoxin II from B. asper on C2C12 myogenic cells by rDM64 was also evaluated. The toxin-inhibitor complex was incubated with myotubes for 3 hours and the rDM64 purified by affinity chromatography showed inhibitory properties, inducing a 92% reduction of the cytotoxic effect of myotoxin II when tested at a 1:1 molar ratio (myotoxin:rDM64)(Fig 6). When a 2:1 molar ratio was used, the cytotoxicity was inhibited by 65%, whereas a 15% inhibition was observed when a 4-fold molar excess of the toxin was tested. When the inhibitory effects of equivalent concentrations of native and recombinant DM64 were compared, a slight but significant difference was observed at a single molar ratio (i.e., 2 myotoxin:1 DM64). These results indicate that both DM64 and rDM64 show similar anticytotoxic effects.
Molecular biology techniques enable the production of recombinant proteins in large amounts; these proteins can then be used in therapeutics and scientific research. Currently, E. coli is the most widely used expression system, although many eukaryotic proteins are not efficiently expressed in this organism. Protein misfolding and the absence of post-translational modifications are important limitations in the use of E. coli. Pichia pastoris could be an eukaryotic organism alternative for the expression of recombinant proteins from mammals. This methylotrophic yeast presents advantages such as the presence of a strong promoter region regulated by methanol (AOX1), the possibility of allowing disulfide bond formation, the ability to secrete recombinant proteins, and the capacity to perform most post-translational modifications [48–50].
The vector pPICZαA was used for rDM64 expression (Figs 1 and 2). It is regarded as a good vector because it contains the Sh ble gene from Streptoalloteichus hindustanus, which confers resistance to Zeocin, allowing the best P. pastoris clone to be selected and screened for the inhibitor expression. The recombinant inhibitor was secreted due to the presence of the N-terminal α-factor signal peptide of S. cerevisiae. The vector also encodes a C-terminal c-myc epitope and a polyhistidine (6xHis) tag, enabling the easy detection and purification of recombinant proteins, respectively. However, this extra 21-amino-acid sequence may modify the structure of the recombinant protein, compromising its biological activity. Santos-Filho and co-workers [51] have previously reported using the pPICZαA vector to produce recombinant BaltMIP, a myotoxin inhibitor from B. alternatus serum, adding the c-myc epitope and the His-tag at the C-terminus of rBaltMIP. However, the inhibitory effect of the rBaltMIP was lower than that of the native BaltMIP. For this reason, we decided to maintain the natural structure of the rDM64, and such a C-terminal peptide was not added. Despite lacking these features, our results show that rDM64 was effectively purified using affinity (column conjugated with myotoxin II) and ion exchange chromatographies (Fig 3). For the unequivocal identification of the recombinant protein, we have used mass spectrometric data instead of just epitope recognition (S1 Table).
Fermenting Pichia pastoris generally starts in a shake-flask system before the culture is transferred to a larger volume fermenter. The shake-flask system provides suboptimal conditions due to the lack of data recording and regulatory control systems. Nonetheless, the production of rDM64 in shake flasks was advantageous because it is a low-cost and less complex method. However, proteolytic degradation is a recurrent problem when working in shake-flask cultures. In these experiments, methanol is the carbon source and induces the AOX1 promoter; therefore, recombinant protein induction also creates conditions that trigger an excess of protease production [48,52,53]. Methanol can also induce cell lysis by oxidative stress and heat-shock responses, eliciting a proteolytic response when the cells are growing exponentially, which results in high cell-density fermentation. In this regard, oxidative stress may also be responsible for recombinant protein degradation because of the increased amounts of reactive oxygen species that are produced during methanol induction. Our mass spectrometry analysis of protein expression in the medium identified proteases and intracellular proteins of yeast in addition to rDM64. This analysis (S1 Table) suggests that yeast lysis occurred during expression in shake flasks. The growth of P. pastoris in shake flasks created stress conditions, such as starvation, heat, pH changes, and/or toxic chemicals. Although expression in shake flasks may have also generated proteolytic products and/or glycoforms in minor quantities, it apparently did not affect rDM64 activity. The biological activity of the partially purified rDM64 fraction was similar to that of native DM64 (Fig 6).
During the expression of rDM64, we added Pefabloc, a potent and irreversible serine proteinase inhibitor, to the expression medium (Fig 1B). It has lower toxicity, improved solubility in water and higher stability in aqueous solutions than other inhibitors (e.g., PMSF and DFP). This serine protease inhibitor was used to protect against the proteolytic degradation induced by yeast serine proteases. This molecule may additionally inhibit the production of reactive oxygen species that are generated by NADPH oxidase [54–56]. Therefore, Pefabloc was also used to decrease ROS generation in yeast cultures in shake flasks, thus reducing rDM64 degradation. This report is the first to use a synthetic serine protease inhibitor during recombinant protein expression in a Pichia pastoris culture. However, as shown in Fig 2B, the use of Pefabloc during fermentation did not completely inhibit the proteolytic degradation of rDM64. Although its recommended working concentration ranges from 0.1 to 1 mM, a maximum concentration of 0.25 mM should be used when working with cell cultures. The concentration used in the present study (0.2 mM) may therefore represent a suboptimal inhibitor concentration considering a) the increased production of P. pastoris extracellular/cell-bound proteases that may be elicited by methanol induction and b) the release of intracellular proteolytic enzymes following cell lysis induced by oxidative stress. Nevertheless, Pefabloc favored both the growth of the cell culture and heterologous protein production, indicating that this strategy could be used for low-scale recombinant expression of other proteins.
Comparing the molecular masses of native and recombinant DM64 showed that rDM64 had a slightly smaller mass (Fig 4). Previously, our group reported that the N-glycan moiety of native DM64 is of the complex type, being composed of N-acetylglucosamine, mannose, galactose, and sialic acid [57]. However, the glycosylation of foreign proteins by P. pastoris includes only mannose residues. The N-deglycosylation assay showed that the molecular masses of the recombinant and the native protein were similar (Fig 4); hence the smaller molecular mass of the glycosylated recombinant protein is likely to be due to differences in the glycan moieties. It is important to observe that glycoproteins expressed in Pichia pastoris are less frequently hyperglycosylated than those produced in S. cerevisiae, although excessive glycosylation in P. pastoris has also been reported [48]. Glycosylation in the lumen of the endoplasmic reticulum after protein translation is similar in both yeast and mammalian cells, but in yeast Golgi, mannose residues are added and oligomannose units can be α-1,6 linked to the α-1,3 mannose in the Manα-1,3-Manβ-1,4-GlcNAc2 inner core. In this way, P. pastoris glycosylation results in diverse structural heterogeneity of the rDM64, which is attributed to the Golgi glycosylation enzymes [48,58,59]. In the heterologous expression in this study, the higher molecular mass bands (136 and 88 kDa) were also identified by mass spectrometry as rDM64 (Fig 2A). Periodic acid-Schiff (Fig 3D) results suggest that these bands may represent glycoforms that are produced by P. pastoris, although the deglycosylation assay does not seem to be conclusive (Fig 4).
Glycosylation is critical in several biological properties, including structural stability, biophysical characteristics, and resistance to proteolytic attack [60,61]. N-glycosylated proteins are abundant in eukaryotic cells, and N-linked glycans all contain a common trimannosyl-chitobiose core with one or more antennae attached to each of the two outer mannose residues [61,62]. Leon and co-workers [57] have previously shown that after sialic acid and galactose removal, DM64 was still able to interact with myotoxin II from B. asper. The same behavior was observed for DM43, an antihemorrhagic protein homologous to DM64 (71% sequence identity) that targets snake venom metalloproteases and does not inhibit myotoxins [63,64]. Interestingly, PNGase F-treated DM43 was half as effective as native DM43 in inhibiting the hydrolysis of azocasein by jararhagin [57]. On the other hand, our present results showed that the complex-type N-glycans of native DM64 can be fully replaced by a high-mannose N-glycan structure without impairing the inhibitory activity of rDM64 toward myotoxin II (Fig 6). Although other reports in the literature have shown that partial deglycosylation does not impair the structural stability of native proteins [65–67], the presence of carbohydrate moieties improves the solubility of proteins and may also be important during in vivo folding of nascent glycoproteins [61,68].
In summary, local myonecrosis caused by Bothrops species is an important public health problem, as it may cause permanent disability of the victims in addition to generating high medical costs due to increased hospitalization times [4,26,69]. This effect is mainly induced by myotoxic phospholipases A2 and, indirectly, by hemorrhagic metalloproteases. Therefore, the local administration of effective myotoxin inhibitors that are based on the structure of rDM64 may represent a valid alternative to reduce tissue damage at the bite site. The present study demonstrated the successful expression of rDM64 in P. pastoris cultures in small volumes. The maintenance of the native-like structure of the inhibitor was fundamental to preserving its anticytolytic effect, suggesting that rDM64 may also inhibit the in vivo myotoxic effect of myotoxin II. The production of a biologically active myotoxin inhibitor by yeast cells can contribute to the development of a therapeutic alternative for the treatment of envenomation by bothropic snakes. The production of rDM64 can also be exploited by studies aiming to map the structure-function relationship of toxin inhibitors and their molecular targets. This result is very important since, while many works have reported the primary structure of inhibitors isolated from snake and mammalian sera, none of these works have shown the three-dimensional structures of their inhibitors [38,63,64,70].
Our group has made multiple unsuccessful attempts to crystallize inhibitors from animal serum. Although the inhibitor DM43 has been crystallized, only low-quality diffraction patterns have been obtained. The heterogeneity of complex-type glycans tends to impair crystallization, and whether the high-mannose glycan structures inserted in recombinant proteins result in more homogeneous structures must yet be tested. The expression of DM64 in P. pastoris could also assist the development of alternative low-resolution structural analysis techniques, such as cross-linking-mass spectrometry (XL-MS), hydrogen-deuterium exchange-mass spectrometry (HDX-MS), and small-angle X-ray scattering (SAXS), using both full-length heterologous inhibitors and/or selected structural domains. Glycosylation is an important feature that maintains the structure of the inhibitor and its soluble state. Previous attempts to express toxin inhibitors using an Escherichia coli system were not successful, probably due to the absence of post-translational modifications. For future development, new research on rDM64 will be undertaken to increase the production scale for structural and in vivo assays.
|
10.1371/journal.pcbi.1002887 | Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling | Cellular signal transduction generally involves cascades of post-translational protein modifications that rapidly catalyze changes in protein-DNA interactions and gene expression. High-throughput measurements are improving our ability to study each of these stages individually, but do not capture the connections between them. Here we present an approach for building a network of physical links among these data that can be used to prioritize targets for pharmacological intervention. Our method recovers the critical missing links between proteomic and transcriptional data by relating changes in chromatin accessibility to changes in expression and then uses these links to connect proteomic and transcriptome data. We applied our approach to integrate epigenomic, phosphoproteomic and transcriptome changes induced by the variant III mutation of the epidermal growth factor receptor (EGFRvIII) in a cell line model of glioblastoma multiforme (GBM). To test the relevance of the network, we used small molecules to target highly connected nodes implicated by the network model that were not detected by the experimental data in isolation and we found that a large fraction of these agents alter cell viability. Among these are two compounds, ICG-001, targeting CREB binding protein (CREBBP), and PKF118–310, targeting β-catenin (CTNNB1), which have not been tested previously for effectiveness against GBM. At the level of transcriptional regulation, we used chromatin immunoprecipitation sequencing (ChIP-Seq) to experimentally determine the genome-wide binding locations of p300, a transcriptional co-regulator highly connected in the network. Analysis of p300 target genes suggested its role in tumorigenesis. We propose that this general method, in which experimental measurements are used as constraints for building regulatory networks from the interactome while taking into account noise and missing data, should be applicable to a wide range of high-throughput datasets.
| The ways in which cells respond to changes in their environment are controlled by networks of physical links among the proteins and genes. The initial signal of a change in conditions rapidly passes through these networks from the cytoplasm to the nucleus, where it can lead to long-term alterations in cellular behavior by controlling the expression of genes. These cascades of signaling events underlie many normal biological processes. As a result, being able to map out how these networks change in disease can provide critical insights for new approaches to treatment. We present a computational method for reconstructing these networks by finding links between the rapid short-term changes in proteins and the longer-term changes in gene regulation. This method brings together systematic measurements of protein signaling, genome organization and transcription in the context of protein-protein and protein-DNA interactions. When used to analyze datasets from an oncogene expressing cell line model of human glioblastoma, our approach identifies key nodes that affect cell survival and functional transcriptional regulators.
| Cellular signaling and transcription are tightly integrated processes that underlie many short- and long-term cellular responses to the environment. Dysregulation of these molecular events has been implicated in diverse diseases including neurodegeneration [1], [2], metabolic disorders [3], and every stage of tumor development and growth [4], [5]. Sophisticated algorithms have been developed to use transcription profiling data for discovery of regulatory networks in disease, either de novo or from an interactome network (see review of theory [6] and tools [7]). Despite the utility of these methods, they suffer from the limitation that they use gene transcripts as a proxy for proteomic changes. As a result, they are unable to capture post-transcriptional changes in proteins, which are an important part of signaling and regulation.
The advent of improved proteomic methods has the potential to provide a systematic map of critical signaling pathways that are altered in disease. Computational approaches to combine transcriptional and proteomic data have focused on assessing the correlation between the data sources [8]–[10]. Some network analyses of proteomic and transcriptional data treated both as evidence of changes in protein levels, which were then viewed in the context of known pathway models [11], [12]. By equating transcripts and the proteins they encode, such network models do not make full use of the data. Alternative approaches that treat proteomic and transcriptional data as distinct can examine how proteomic signaling drives changes in gene regulation. Methods that search for physical associations among proteins and between proteome and the genome are likely to be particularly important in the analysis of phosphoproteomic data from mass spectrometry. Phosphoproteomics selectively measures protein phosphorylation, a principal biochemical mechanism of cellular signaling controlling gene expression. Since changes in phosphorylation and transcription are poorly correlated ([13] and Figure S1), they are highly complementary, providing distinct windows into cellular processes.
Previously, we have shown that phosphoproteomic and transcriptional data from the yeast Saccharomyces cerevisiae pheromone response could be linked through physical networks in a framework of constraint optimization on interactome networks known as the prize-collecting Steiner tree (PCST) [14]. This approach revealed relevant proteins and pathways that could not be discovered when each type of data was analyzed in isolation. Furthermore, it identified a network more compact and functionally relevant than networks constructed from direct interactors of the phosphoproteomic hits and transcription factors or by a related network optimization method ResponseNet [15], [16]. However, the method depended on the availability of experimentally determined genome-wide binding locations for almost all the transcriptional regulators in yeast [17], [18]. Such data are unavailable for mammalian cells in which the most comprehensive analysis so far has produced data for fewer than 10% of the transcription factors and only in a limited number of cell types [19].
In order to study disease-related pathways in mammalian cells, we developed a combined computational and experimental strategy that predicts transcription factors with altered binding or activity relevant to a particular cell type under specific conditions. First, we used DNaseI-hypersensitivity site sequencing (DNase-Seq) [20] to identify regions bound by as yet unknown regulatory proteins in each condition of interest. Scanning these sequences with a library of motifs revealed preliminary candidates of relevant binding proteins [21]–[23]. To discover the subset of these motifs most likely to be regulatory, we employed a regression-based method to infer the activity of specific transcription factors [24]–[26]. For each potential regulatory protein, we tested whether the quality of transcription factor motif matches in differentially hypersensitive regions correlated with the expression level of nearby genes. We then searched for protein-protein interactions that link these transcriptional regulators to upstream phosphoproteomic events by solving a PCST problem on the interactome (Figure 1A).
Beginning with an interactome network in which the reliability of each interaction is weighted by experimental evidence, we find an optimal subnetwork of the most reliable interactions that include a subset of the phosphorylation events and the transcriptional regulators, with preferences given to phosphorylation events that undergo large changes and transcriptional regulators that show strong activities. An important aspect of the PCST algorithm is that it is able to naturally account for missing data and false positives. In particular, since the input experimental data do not capture every relevant protein, we allow the network to include proteins that were not explicitly measured in the proteomic or transcriptomic assays. In addition, to account for false positives in the data, we allow the algorithm to exclude experimentally determined proteins and genes that are not connected with high confidence interactions. As a result, the final networks are compact, enriched for functionally relevant proteins and the most reliable interactions that include these proteins, and can be used to guide subsequent experiments. To further account for possible noise in the input datasets, we merge an optimal PCST solution with a set of suboptimal solutions to obtain a robust final network (see Materials and Methods).
Here we apply our integrated approach to prioritize experiments for probing the signaling pathways downstream of a mutant epidermal growth factor receptor (EGFR) in glioblastoma multiforme (GBM). The variant III mutant of EGFR (EGFRvIII) is the most common deletion mutant of EGFR in human cancer [27] and its levels are highly correlated with poor prognosis in GBM [28]–[30]. The deletion of exons 2–7 removes most of the extracellular ligand binding domain, so it is unable to bind EGF or other EGFR-binding ligands [31]. Nevertheless, the mutant receptor is constitutively phosphorylated [32], and is capable of activating downstream signaling pathways at a low level. Unlike wild-type EGFR signaling, which shuts itself off through a process known as receptor-mediated down-regulation, EGFRvIII signaling does not, leading to its oncogenic properties [31]. To comprehensively identify the downstream signaling consequences of the EGFRvIII, we incorporated phosphoproteomic, transcription profiling and DNase-Seq data from U87MG glioblastoma cells expressing this oncogenic mutant receptor (Figure 1B). In addition to recapitulating many known components in EGFRvIII signaling and transcriptional regulation, our network predicts key signaling nodes not apparent from the experimental data and provides a method for prioritizing experimental tests. We validated several of these predictions through pharmacological tests and genome-wide protein-DNA binding measurements. We propose that combining epigenomic methods to uncover transcriptional regulators and constraint optimization on biological networks effectively organizes disparate transcriptional and proteomic data, and can be used to discover unknown components of biological responses leading, potentially, to new therapeutic strategies.
To understand the signaling pathways downstream of EGFRvIII, we used as a model two cell lines derived from human U87MG glioblastoma cells that were engineered to express high levels of EGFRvIII (U87H; 2 million receptors per cell) or a mutant form of the receptor with an inactive kinase (U87DK) [31], [32] (Figure 1B). U87MG is a widely used cell line model for human grade IV glioma and many aspects of its biology have been well characterized. Independent long-term cultures of these cells are genetically stable [33], making it possible to interpret results from different laboratories at different times. Many findings from EGFRvIII expressing-U87MG cells have been validated in vivo [31], [32], [34]–[37]. Since the signaling network for wild-type EGFR is well established but not for EGFRvIII, this system provided us the opportunity to explore the global biochemical events downstream of an important oncogenic mutant, the relationships between these events and their functional significance in a setting with established biological relevance. A previous quantitative phosphoproteomics study [37] detected 100 phosphorylation sites on 88 proteins in these two cell lines. In this study, we used microarrays to measure global expression differences between these cells, identifying 1,623 differentially expressed genes (Table S1).
In order to uncover links between the phosphorylation and transcriptional changes, we identified a set of transcriptional regulators that were most likely to be differentially active in the two cell types and responsible for the transcriptional changes. Our approach integrates sequence information with epigenomic and expression differences between the cell lines. We collected DNase-Seq data in each cell line and found 12,807 regions that showed quantitative changes in hypersensitivity. About 68% of these sites were hypersensitive in only one condition, while the rest were hypersensitive to different extents in the two cell lines. We then scanned these differentially hypersensitive regions for matches to known DNA binding motifs to compute affinity scores of the motifs in these regions. To select a subset of these motifs most likely to drive expression differences, we scored each motif using a univariate regression model relating its affinity scores in the differentially hypersensitive regions to differential mRNA expression of genes within 40 kb [24]–[26]. Each regression equation evaluated the function of one protein over potentially hundreds of proximal and distal regulatory elements associated with the set of differentially expressed genes. As a result, we were able to discover trans-acting transcription factors from the changes between two cell types.
With this approach, we identified 185 significant motifs that mapped to 297 proteins in the interactome (see Materials and Methods). These proteins represent candidate factors that show either partial or complete changes in their activity between the cell lines. Proteins identified either through phosphoproteomic measurements or by this regression analysis were provided as input to the PCST algorithm along with the interactome data (Figure 1B). The algorithm was run multiple times to find an optimal PCST and a set of ten related sub-optimal solutions. Merging these network solutions, we obtained the network presented in Figure 2. Adopting the PCST terminology, we refer to the phosphoproteins and transcription factor candidates as “termini” and the nodes supported only by the network analysis as “Steiner nodes”.
The interactome we used was derived from iRefIndex [38], a protein interaction database consolidating records from many primary interaction databases such as BIND [39], BioGRID [40], HPRD [41], IntAct [42] and MINT [43]. To account for the fact that the interactions vary in their reliability, we used the Miscore algorithm [44] to assign to each edge in the interactome graph a cost that is inversely related to a likelihood score for the interaction, taking into account the experimental methods that detect the interaction, the type of the interaction, and the number of publications supporting the interaction. The PCST algorithm aims to minimize the sum of the costs of edges required to connect the termini. Therefore, the optimal PCST solution represents a most likely network that links together the protein termini supported by experimental data.
The PCST network constructed from the U87 EGFRvIII dataset consists of 199 edges and 172 nodes, of which 65 are phoshoprotein termini and 63 are transcription factor termini (Figure 2). It gives a high-level view of the overall biological processes involved. Such processes include signaling pathways known to be activated downstream of EGFRvIII in U87MG cells, such as the phosphatidylinositol 3-kinase (PI3K) pathway and the Ras-Raf-MEK pathway (HRAS, RAF1 and various mitogen-activated protein kinases (MAPK) in Figure 2) [45], [46]. Others are more general processes induced by EGFRvIII that are known to contribute to tumor development, such as regulation of the cell cycle [45], DNA damage response [36], actin cytoskeleton organization and cell motility [47], [48]. We also note that the network includes the following two cell surface receptors: MET (met proto-oncogene), which contains a tyrosine phosphorylation site, and ERBB2 (v-erb-b2 erythroblastic leukemia viral oncogene homolog 2, neuro/glioblastoma derived oncogene homolog (avian)), for which no phosphorylated tyrosine residue was detected. Both were previously shown to be cross-activated by EGFRvIII [37], [49].
Intriguingly, the network includes two subnetworks containing neurological processes previously linked to GBM: synaptic transmission and axon guidance. The neurotransmitter glutamate, which mediates synaptic transmission through the N-methyl-D-aspartate glutamate receptors 2A and 2B (GRIN2A and GRIN2B) that appear in the PCST network, can promote glioma cell growth [50], and this pro-proliferative effect in U87MG cells is due to EGFR signaling [51]. In addition, genes associated with synaptic transmission and axon guidance are detected as frequently altered in large-scale sequencing and gene expression analysis of human GBM [52]. Although the genetic and transcriptional changes are associated with the neural subtype of GBM, where the EGFRvIII mutation was not found [53], our findings raises the possibility that EGFRvIII-induced post-translational modification may also alter these processes.
Our approach also recovered a number of transcriptional regulators already known to be involved in EGFRvIII signaling. For example, our network identifies the transcription factor signal transducer and activator of transcription 3 (STAT3), which has been previously implicated in EGFRvIII induced transformation [54]. The STAT proteins are unusual in that they were included due to three types of evidence: the expression regression procedure, a tyrosine phosphorylation site on STAT3 (Y705) [37] and direct interactions with EGFR (both wild-type EGFR and EGFRvIII) [54]–[57]. In contrast, the other transcription factors in the PCST network were present solely because of our modeling approach. Several of the prominent transcription factors in the PCST network, such as nuclear factor (NF)-κB (RELA and REL), activator protein 1 (AP-1; consisting of c-Fos (FOS), c-Jun (JUN), ATF proteins), and CCAAT/enhancer binding protein (C/EBP; family member C/EBP-β (CEBPB)), have been reported to be activated in EGFRvIII-expressing glioblastoma cells in an independent study [58]. Furthermore, the oncoprotein MYC, which was captured by our network, is known to be expressed in the U87MG cells and its transcriptional activity contributes to the undifferentiated state and consequently the high tumorigenicity of this cell line [59]. We emphasize that neither the phosphoproteomic data nor the transcription profiling data alone would have suggested roles for these proteins. No phosphorylation sites on these transcription factors were reported by mass spectrometry, and even the tyrosine phosphorylation site on STAT3 (Y705) showed less than 10% change in response to EGFRvIII expression. Nevertheless, in our network solution these transcription factors are prominently featured, demonstrating the value of the PCST algorithm for integrating these data by the interactome. By contrast, standard analysis of the promoter sequences of the differentially expressed genes for enriched transcription factor motifs resulted in only the cell cycle regulator E2F and a zinc finger protein (Table S2).
The PCST solution network was constructed by finding a small number of proteins that directly or indirectly interact with the set of termini in order to explain the measured changes between U87H and U87DK cells. Compared to networks constructed from proteins that directly interact with the set of phosphorylated proteins, transcription factor candidates or both (“nearest neighbor” networks), the number of nodes in the PCST solution network is about 4% to 7% of these nearest neighbor networks (Figure 3A). To systematically assess whether the highly compact view of the experimental data provided by the PCST solution is able to capture the underlying biological process, we compared the PCST solution to the alternate network approaches with respect to curated collections of genes known to be relevant to GBM. The Cancer Genome Atlas (TCGA) GBM 6000 Gene Ranker [60] scores and ranks over 7,600 genes for relevance to GBM based on gene expression, mutation, pathway analysis and literature curation. Figure 3B illustrates the scoring of all nodes reported by the various network solutions in addition to the PCST. As expected, the set of all termini (those directly detected in the phosphoproteomic data and inferred from differential expression) have high scores. However, we found that the proteins included in the PCST solution scored even higher, and also scored higher than the set of proteins in the input interactome that were excluded from the PCST solution (Figure 3B). A possible trivial explanation for this enrichment could have been that all proteins that interact with the termini are more relevant than the non-interactors, as a result of the biology or biases in the gene ranker curation. To explore this possibility, we compared the PCST to the nearest neighbor networks of various sets of termini and to ResponseNet, a previously published optimization-based network construction algorithm [16]. Both ResponseNet and PCST outperform the nearest neighbor networks. As the PCST solution is smaller than that of ResponseNet and achieves slightly better performance, it is likely to be better for guiding experiments. In summary, the PCST network approach selected a combination of experimentally determined proteins and “hidden” proteins that are relevant to GBM.
Since GBM is a heterogeneous disease where the EGFRvIII mutation is among the most common mutation of EGFR [53], we used independent datasets to validate that our network was specific to the EGFRvIII mutation among GBM cases. At the signaling level, we utilized a recently published global phosphoproteomic analysis of mouse GBM xenografts [61]. This study identified 225 tyrosine phosphorylation sites on 168 proteins, and is independent of the U87 cell line data that were the input to the PCST network. Comparison of xenografts expressing high level of EGFRvIII to xenografts expressing normal level of wild-type EGFR resulted in 11 differentially phosphorylated proteins. We asked how closely connected (in the protein interaction network) these EGFRvIII-specific phosphorylation events detected in xenografts were to proteins in the PCST network derived from the U87MG cell lines (Figure 3C). 158 of the 168 proteins in this dataset are present in the set of 11,637 proteins in the interactome that we could score for connectivity to the PCST. We found that 10 out of the 11 (91%) EGFRvIII-specific phosphorylation events fell in the top 6.4% of the proteins closest to the PCST compared to 36.7% of the phosphorylation events that were not EGFRvIII-specific (p<0.003 and Figure 3D). This suggests that the PCST solution, although constructed using experimental data from a tissue culture model, is closely related to protein signaling changes induced by the same oncogenic mutation in vivo.
To determine if the transcription factors in the PCST solution contribute to the transcriptional response to EGFRvIII mutation in human patients, we used publicly available TCGA GBM data [53], [62] to compare gene expression alterations in patients with and without the vIII mutation. More specifically, we extracted a set of samples that share mutation status with the U87 cells (p53 wild-type and p16 null) and classified those samples by EGFRvIII status based on EGFR exon probe set expression. Overall, the differences in expression between tumors with wild-type EGFR and those with the vIII mutation were small, which is consistent with the heterogeneous nature of GBM (Figure S2). We tested whether the targets of transcription factors in the PCST solution were associated with vIII-associated changes in gene expression as described below. First, we collected potential targets of each transcription factor in TRANSFAC by scanning for motif matches in genomic regions that show higher DNaseI hypersensitivity in each of the U87H and U87DK cells within 40 kb of transcription start sites. We then ranked all genes by the log-fold change between patients with and without the EGFRvIII mutation and used the minimum hypergeometric statistic (mHG; [63]) to ask whether the ranks of the putative targets of that transcription factors differ from the distribution expected from whole genome background by chance. Thus, for each transcription factor identified in the U87H and U87DK cells we obtained a p-value representing the significance of finding EGFRvIII responsive up- and down-regulated genes in the putative targets of this factor, depicted in Figure 3E.
The EGFRvIII up-regulated genes in the TCGA samples are more strongly enriched (lower mHG p-values) for targets of the transcription factors identified in the U87H cells (U87H TF) than the EGFRvIII down-regulated genes (Figure 3E top panel). Conversely, for the set of transcription factors that have motif matches in regions with increased DNaseI hypersensitivity in the U87DK cells (U87DK TF), their targets are more strongly enriched for EGFRvIII down-regulated genes than the up-regulated genes (Figure 3E bottom panel). These results support the utility of differential hypersensitivity for identifying transcription factors relevant to human tumors even though the DNase-Seq data were generated exclusively from cell lines and the EGFRvIII mutation has a modest effect on the global transcriptional programs in patients.
To assess the role of the PCST in selecting biologically relevant transcription factors, we classified factors based on whether they were included in the PCST solution. The U87H transcription factors included in the PCST network solution have even stronger enrichment in the EGFRvIII up-regulated genes than those factors excluded from the PCST solution (Figure 3F first panel), and the U87DK transcription factors included in the PCST show stronger enrichment in genes down-regulated in EGFRvIII patients than the excluded factors (Figure 3F fourth panel). Such differences between the included and excluded transcription factors are not observed for the associations between the U87H transcription factor targets and the EGFRvIII down-regulated genes, or between the U87DK transcription factor targets and the EGFRvIII up-regulated genes (Figure 3F second and third panel). This finding suggests that for transcription factor candidates selected based on motif evidence in open chromatin and correlation with target expression in cell line models, those that can be linked to upstream signaling changes are more likely to affect expression changes in vivo.
Having established that our approach recovered prior knowledge related to GBM, we used the network to prioritize experiments for testing new points of intervention in the network. We selected protein targets based on two criteria: (1) a quantitative score that measures how closely connected they were to the nodes in the PCST in the interactome (Figure 3C) and (2) whether they were targets of commercially available small molecules [64]–[66]. Our final list included fifteen nodes, all of which were among the 30% of the interactome most closely connected to the PCST. Eight of these were extremely high-scoring nodes (of which seven are in the original PCST and one is closely connected to the PCST), one was an intermediate scoring nodes, and six were lower-scoring nodes (Table 1). We treated the U87DK and U87H cells with small molecule antagonists for these highest-, intermediate- and lower-scoring nodes (“high-ranked”, “mid-ranked” and “lower-ranked” targets) at a wide range of concentrations and measured the resulting cell viabilities relative to those of vehicle treatment.
We first compared the effects of all the compounds at a concentration of 10 µM (except for harmine, which was only soluble at 5 µM in the organic solvent DMSO) (Figure 4A). With the exception of PDTC, which targets NF-κB (NFKB1), all compounds targeting the highest-ranked nodes reduced viability by at least 40% at 10 µM. In the cases of 4-OHT (targeting the estrogen receptor ESR1), 17-AAG (targeting heat shock protein 90kDa α HSP90AA1) and PKF118–310 (targeting β-catenin CTNNB1), more than 70% of the cells were killed at concentrations lower than 10 µM. The effects on the mid- and lower-ranked nodes were more modest, resulting in 10 to 50% reduction of viability. One trivial explanation for such difference would be if the compounds for the intermediate and lower-ranked targets happen to require higher doses in all biological contexts. However, this was not likely the case because these concentrations, which were only weakly effective in the U87-derived cells, were a few orders of magnitude higher than the GI50 values (the concentration that inhibit growth by 50%) in the NCI60 panel of cancer lines as reported by the In Vitro Cell Line Screening Project (IVCLSP) of the National Cancer Institute Developmental Therapeutics Program ([67] and http://dtp.nci.nih.gov/). By contrast, the doses we used to target the highest-ranked nodes were within the range of GI50 values (Figure S3). Therefore, we conclude that the highest-ranked targets have stronger effects on the viability of the U87 cells than the targets with lower ranks.
To further understand the behavior of these compounds on the U87 cells, we fit the measured dose responses to a four-parameter log-logistic (4-PL) function. For compounds that could be fit to the 4-PL model (lack-of-fit test p>0.05), we compared parameters of the fitted curves between U87H and U87DK cells (Figure 4B). Compounds for three out of the four highest-ranked targets, SAHA (targeting histone deacetylase HDAC1), 4-OHT (targeting estrogen receptor ESR1) and PFK118–310 (targeting β-catenin CTNNB1) were more toxic under EGFRvIII expressing condition (p<0.01), with LD50 values an average of two-fold higher in the DK cells. In contrast, cells treated with two of the three compounds for the lower-ranked targets, SB-505124 (targeting TGF-β receptor 1 TGFBR1, ranked 193 out of 11,637) and harmine (targeting dual-specificity tyrosine phosphorylation-regulated kinase DYRK1A, ranked 2,232 out of 11,637), exerted similar effects on the two cell lines. SB-431542 (for the lower-ranked targets TGFBR1, ranked 193 out of 11,637, and activin receptor type-1B ACVR1B, ranked 1,695 out of 11,637) appeared to exert a significantly different effect in the presence of EGFRvIII, but the difference in LD50 values was only 1% and more than 30 µM was required to reduce viability by 50% compared with much lower doses for the high-ranked targets. These results demonstrate that the compounds targeting the high-ranked nodes are more likely to have strong effects on cell viability and be associated with large differential sensitivity between two cell lines, implying the high-ranked targets give rise to important signaling differences induced by EGFRvIII.
Since transcriptional regulation is the first step that defines the long-term behavior of the cell including tumorigenesis, we sought to identify the transcriptional regulators responding to oncogenic signaling. In addition to several transcription factors known to be induced by EGFRvIII, our PCST network also included novel putative transcriptional regulators. Using the same scoring procedure as for ranking targets for small molecules (Figure 3C), we selected the highly ranked transcriptional co-regulator p300 (EP300) for experimental validation (Table 1). p300 is a particularly interesting candidate because it is the highest ranked transcriptional regulator (out of 937 annotated transcription factors, co-activators and co-repressors ranked by the network), and although it appears in the network it is not a terminal, i.e., it was not identified as a transcription factor candidate, nor does it contain phosphorylated tyrosine residues. We chose to perform a chromatin immunoprecipitation sequencing (ChIP-Seq) experiment for p300 since determining the genome-wide binding locations of a transcriptional regulator may suggest its biological functions and shed light on its regulatory mechanism. Both aspects could validate the predicted relevance of p300 in our experimental system and therefore demonstrate the capability of our method to uncover important regulators not present in the original signaling and transcription data.
We therefore performed a ChIP-Seq experiment in the U87H cells to identify the genes targeted by p300. 60.8 million uniquely mapped reads of 36 bp were obtained (Table S3). Peak calling by MACS reported 28,721 peaks at low stringency (p-value threshold 1E-05) and 7,657 peaks at high stringency (p-value threshold 1E-07), mapped to 6,391 and 1,969 genes within a 10 kb window, respectively.
p300 does not directly bind DNA but is instead recruited to its targets by sequence-specific DNA-binding proteins. Therefore, the PCST algorithm selected it solely by virtue of its protein-protein interactions. Nevertheless, if the network is correct, we should be able to detect the consequences of its recruitment to the DNA. Indeed, we found evidence that p300 is actively involved in chromatin remodeling of these cells: p300 binding sites are significantly enriched in regions that increase in hypersensitivity in the U87H cells compared to U87DK cells (p<1E-23; Figure 5B), suggesting that p300 may regulate transcriptional changes between the two cell types.
Our networks also included the DNA-binding proteins that recruit p300 to its targets. To test these predictions, we examined the sequence motifs present in the highest confidence p300 targets, defined as those regions that were present in the ChIP-Seq data and showed a change in DNaseI hypersensitivity between the cell lines. We used cross-validation to rank motifs by their ability to predict which differentially hypersensitive sequences were p300 targets (see Materials and Methods). Of the 151 non-redundant vertebrate TRANSFAC motifs, 52 correspond to proteins known to interact with p300; 33 of these were inside the PCST and 19 were outside of it. Among the 20 motifs that were the most predictive of p300 recruitment, at least 8 correspond to proteins inside the PCST (enrichment p-value = 0.039 by Fisher exact test), while only one was outside of the PCST (enrichment p-value = 0.94 by Fisher exact test) (Table S4). These motif results, together with the observed association between p300 binding and hypersensitive sites, provide solid support for the role assigned to p300 in the network as a transcriptional regulator responding to EGFRvIII.
Gene Ontology enrichment analysis of the p300 targets identified by ChIP-Seq revealed the potentially important role of this protein in regulating the transcriptional consequences of the EGFRvIII mutation (Table 2), specifically in cellular adhesion and response to hormone. Neither of these categories is enriched in p300-bound sites in the two other cell types in ENCODE [19] for which ChIP-Seq data were available (Table S5), suggesting that the functional role of p300 is likely to be specific to our system.
We have shown that the PCST solution provided an integrated view of the biological processes in the EGFRvIII network leading to directly testable predictions. Our analysis revealed proteins whose activities we subsequently targeted with small-molecule inhibitors to block the growth of tumor cell lines. In addition, the PCST solution network identified transcriptional regulators enriched at hypersensitive sites. Many of the proteins we targeted did not appear in the phosphoproteomic and transcriptional profiling data but were selected among thousands of other proteins in the interactome graph that interact directly or indirectly with the hits from the experiments. In particular, a network of direct interactors of the phosphorylated proteins contains 2,554 nodes. In the absence of a network optimization approach, it would be extremely difficult to prioritize experiments. Using the PCST approach, we were able to integrate the phosphoproteomic data with additional transcriptional data and provide a ranked list of proteins for experiments. Targets far from nodes in the PCST were less likely to exert differential cytotoxic effects in response to the oncogenic mutation, and in the cases in which targeting these nodes were cytotoxic, their effects tended to be weaker. Therefore, our approach provides a powerful way to prioritize targets based on experimental datasets that represent different aspects of the cell state such as protein signaling, chromatin conformation, and transcription output.
Our network-based approach was also able to identify transcriptional regulators that could not be found by other methods. Standard promoter analysis of the differential expressed genes in the current datasets yielded little information about potential DNA binding proteins besides the cell cycle regulator E2F and a zinc finger protein (Table S2), whereas integrating data from upstream signaling allowed us to identify promising candidates. The experimental validation of p300 exemplifies the power of our approach. p300 was not a hit in the tyrosine phosphoproteomic dataset, nor was it a sequence-specific DNA binding protein for which a motif can be correlated to differential mRNA expression from the regression analysis. However, it was included in the network due to its connectivity to the measured signaling events and to the sequence specific transcription factors. Identifying genome-wide binding locations of p300 by ChIP-Seq provided experimental support for its role in chromatin remodeling and tumorigenic processes.
At the genome-wide level, p300 targets were found to be enriched in genes involved in the response to hormone. Little is known about the role of p300 in regulating hormone response genes transcriptionally. However, it is known that p300 associates with multiple nuclear hormone receptor proteins and functions as a co-regulator [68], [69]. Our observation that nuclear receptor genes are p300 targets may represent a mechanism for the autocrine loop observed in EGFRvIII expressing glioma cells [70]. p300 targets were also associated with the process of cellular adhesion. In particular, several p300 target genes that were differentially expressed in the presence of EGFRvIII (Figure 5A) are well-characterized markers for the epithelial-mesenchymal transformation (EMT), a process that is known to alter cellular adhesion [71]. We observed that in the presence of EGFRvIII the cells have poor attachment to the tissue culture plate, consistent with alteration in cellular adhesion and possibly a partial mesenchymal phenotype. Therefore, our data point to the potentially important role of p300 in transcriptional regulation in our system and suggest a mechanism for the mesenchymal properties displayed by GBM cells [72], demonstrating how following up on high-ranked transcriptional regulators by ChIP-Seq can lead to new biological hypotheses.
The sequence specific factors responsible for p300 recruitment to EMT-related genes remain to be found. C/EBP-β (CEBPB) and STAT3 have previously been shown to synergistically induce mesenchymal transformation of glioma cells [73]. However, in our data, the levels of C/EBP-β transcript did not change in response to EGFRvIII expression and the phosphoproteomic data showed no significant change in the levels of activated STAT3. Our network results suggest a possible explanation for these findings. The network includes an interaction between C/EBP-β and SMAD4, and SMAD4 is known to repress the transactivation function of C/EBP-β [74]. We also note that the SMAD4 mRNA level is reduced by five-fold in the presence of EGFRvIII. Together, these data suggest that EGFRvIII expression leads to a decrease in SMAD4. This in turn activates C/EBP-β, which recruits p300 to EMT genes.
We tested the effects of seven compounds targeting high-ranked nodes predicted by the PCST solution, and six of these resulted in significant reduction in cell viability. These compounds represent both known and novel therapeutic agents for GBM (Table 3). Of these agents, dasatinib has the best-characterized effect on EGFRvIII glioblastoma. Dasatinib has anti-tumor effects on EGFRvIII-expressing glioblastoma models, including inhibition of invasion and induction of apoptosis [34] although its anti-proliferative effect has also been reported in U87 cells expressing wild-type EGFR [75]. Currently several clinical trials are ongoing for mono- and combination-therapy of dasatinib in GBM [76]. We used dasatinib to target the SRC and FYN kinases, and these have previously been reported to be activated by EGFRvIII [34]. The HSP90 inhibitor 17-AAG, which we found to be highly effective at sub-micromolar concentrations, has also been shown to be effective in a variety of human glioma cell lines and glioma models [77]. 17-AAG has entered clinical trials for several cancer types [78], but has not yet been tested in GBM. The HDAC inhibitor SAHA effectively inhibits tumor cell growth in multiple glioma cell lines and mouse models [79], [80], and a phase 2 clinical trial in patients with recurrent GBM showed modest single-agent activity [81]. In addition, co-delivery of SAHA and siRNA against EGFRvIII synergistically induced apoptosis in GBM cells [82]. Our data is the first demonstration that SAHA alone has additional potency in EGFRvIII-expressing cells. At a mechanistic level, this observation may be related to our discovery that the histone acetylase p300 is involved in chromatin remodeling induced by EGFRvIII.
Our experiments also illustrated that the selective estrogen receptor modulator tamoxifen reduced cell viability in both cell types but more potently in the EGFRvIII-expressing cells. Epidemiologic data on the effect of steroid hormone in the etiology of glioma are ambiguous (reviewed in [83]): prior to menopause, women are at lower risk of glioma than men, suggesting a protective role of estrogens; however, the relative reduction in glioma risk in women from the use of exogenous hormone is small and inconsistent. A high dose of tamoxifen was reported to reduce tumor volume and stabilize tumor progression in a subgroup of recurrent malignant glioma patients [84]. At the molecular level, there are considerable discrepancies regarding the expression of ESR1 in human glioblastoma (see recent summary in [83]). Further study is necessary to determine if particular selective estrogen receptor modulators given alone or in combination with other therapies might be more effective than the estrogen receptor modulator alone. Although ESR1 is a well-characterized transcription factor, the algorithm selected it only because of its interactions with other proteins, not because of the presence of the estrogen responsive element (ERE) sequence motif on DNA. Such ERE-independent actions have precedents; ESR1 is known to cross-talk with protein kinase cascades such as those of ERK (extracellular-signal-regulated kinases) MAPK and PI3K [85], [86] and it regulates transcription by interaction with other transcription factors and co-activators [87], [88]. Interestingly, non-genomic signaling by 17β-Estradiol can both stimulate and inhibit apoptosis [89].
In addition to these previously reported compounds, we identified ICG-001 and PKF118–310 – two agents that are effective against the U87 cells and that had not previously been reported in the context of GBM. ICG-001 inhibits CREBBP and was discovered from screening in a colon cancer cell line [90]. A more potent structural relative of ICG-001, PRI-724, entered phase 1 clinical trial for advanced colorectal and pancreatic cancer in February 2011. The 50% growth inhibitory concentration is in the micromolar range for colon carcinoma cells but ten-fold higher in normal colonic epithelial cells (4.43, 5.95, and 70.90 µM on SW480, HCT116 and CCD-841Co cells, respectively [90]), suggesting the greater than 50% reduction of cell viability we observed at 10 µM concentration is likely to be relevant physiologically. PKF118–310 is a potent inhibitor of the interaction between TCF4 (transcription factor 4) and β-catenin (CTNNB1) [91] and has shown growth inhibitory effects in cell line models of prostate cancer [92], osteosarcoma [93], hepatocellular carcinoma [94], and a mouse model of breast cancer [95], but effects on GBM have not been reported. However, there is evidence that the target of this compound is relevant to GBM: Wnt/β-catenin activation positively correlates with the progression of glioma [96], [97] and down-regulation of β-catenin inhibits glioma cell growth [97], [98]. In our system, expression of EGFRvIII leads to two-fold higher sensitivity to PKF118–310, bringing it to the sub-micromolar range of 50% growth inhibitory concentrations reported in osteosarcoma [93] and hepatocellular carcinoma cells [94]. It is therefore of great interest to determine whether a link exists between EGFRvIII status and β-catenin activation state in GBM patients.
We have demonstrated a method for uncovering a physical network of proteins and genes that respond to expression of an oncogenic mutation, which we have used to reveal new methods for specifically blocking growth of the tumor cells. Our approach for reconstructing mammalian signaling pathways uses epigenomic data to create a critical link between phosphoproteomic changes and downstream transcriptional events via physical associations from the interactome. Based on these networks, we were able to identify key transcriptional regulators and discover a number of compounds that killed EGFRvIII cells more effectively than control cells.
While our approach has a low false-positive rate, it is possible that it will miss many potential targets that lie far from the PCST. In fact, we found that an inhibitor for one of the lower-ranked targets (TUBB1, ranked 3,582 out of 11,637) did show a modest difference in cytotoxicity between DK and H cells. Such false-negatives arise because they are part of biological responses that are not connected to the available data, either because the interactome is incomplete or because these targets are functioning in biological processes that have little in common with the proteomic data on which we based our study.
The proteomic data selectively identified phosphorylated tyrosine residues, which are relatively rare compared to phosphorylated serine and threonine [99] but display faster and bigger fold changes [99]. However, the PCST formulation can be readily extended to incoporate phosphorylation data on these other residues, other post-translation modification on proteins, as well as other experimentally- and computationally-derived protein activities that can be used as constraints. In a similar vein, we can supplement or even completely replace the physical interactome on which to solve for the PCST by associations derived from purely data-driven approaches, which may create thousands of connections with different strengths of association. The optimization procedure provided by PCST builds parsimonius networks consisting of the strongest associations that satisfy the experimental constraints, providing a sound basis for designing experiments.
Our models, which treat transcriptional changes as downstream of proteomic changes, focused on identifying the differences between two cell types at steady state. In dynamic systems, it will be important to include feedback as the transcriptional changes also lead to changes at the proteomic level. Although differential equations provide a natural way to describe such feedback, these approaches are limited to relatively small systems. For example, a comprehensive differential-equations-based transcriptional and translational network for Escherichia coli [100] has been developed, but a genome-wide model for mammalian proteomics and transcription data is not yet feasible. We propose that our approach can be applied to “-omics” data to reduce the complexity of mammalian signaling and transcription to a point where differential-equations-based modeling becomes feasible. The combination of comprehensive “-omics” based analysis followed by quantitative modeling would provide a method for producing highly quantitative predictions of new therapeutic strategies even for a broad range of diseases.
The human glioblastoma cell lines U87MG expressing high levels of EGFRvIII (U87H, 2 million EGFRvIII per cell) and a kinase dead mutant of EGFRvIII (U87DK, 2 million kinase dead receptors per cell) were generous gifts from Dr. Paul Huang and Dr. Forest White at MIT. Cells were cultured in complete media (Dulbecco's Modified Eagle Medium (DMEM; Mediatech) supplemented with 10% fetal bovine serum, 100 units/mL penicillin, 100 mg/mL streptomycin (Invitrogen), 4 mM L-glutamine) and in a 95% air/5% CO2 humidified atmosphere at 37°C. Expression of EGFRvIII and DK receptors were selected by 400 mg/mL G418 (Calbiochem). To enhance cell attachment, tissue culture vessels with the Corning CellBIND surface (Corning) were used.
Total RNA was prepared from the U87MG derived cell lines by the RNeasy Plus Mini Kit (Qiagen) and quantified on the Affymetrix Human Genome U133 Plus 2.0 arrays. Labeling, hybridization, washing and staining were performed following the standard Affymetrix GeneChip protocol. The arrays were hybridized in an Affymetrix GeneChip Hybridization Oven 640 at 45°C at 60 rpm for 16 hours, washed and stained in Affymetrix Fluidics Station 450, and scanned with Affymetrix GeneChip Scanner 3000 7G. Two biological replicates were done for each cell line. The intensity values were normalized using the GC Robust Multi-array Average (gcrma) package [101] in the R BioConductor library and differential gene expression was calculated by the Linear Models for Microarray Data method [102] implemented as the limma package [103] in BioConductor.
The U87DK and U87H cells were seeded in parental media (complete media without G418). After 24 hours, the cells were washed gently with phosphate buffered saline (PBS) and cultured in serum free media for 24 hours. Nuclei extraction and DNaseI digestion followed published protocol [20], [104] for 50 million nuclei for each of the two biological replicates of each cell line. Sequencing libraries were prepared with the Illumina sample preparation kit and 100 to 300 bp fragments were specifically selected by gel electrophoresis. Each biological replicate was sequenced in one lane on a Genome Analyzer II sequencer (Illumina). The 35 bp-long sequencing reads were aligned to the hg18 genome by Illumina's Eland extended software with maximum two mismatches in the first 25 bp. The sequencing and alignment statistics are listed in Table S3.
The U87H cells were seeded in media without G418. After 24 hours, the cells were washed gently with PBS and cultured in serum free media for 24 hours. Crosslinking and cell lysis were done as previously described [105], and sonication was performed on a Bioruptor NextGen sonication system (Diagenode) with 10 cycles of 30 sec on, 30 sec off at high power setting. p300 ChIP was done on the SX-8G IPStar Automated System (Diagenode) with buffers from the Auto Transcription ChIP kit (Diagenode) following instruction manual version V1_07-10-10. The pre-set IP protocol “ChIP 22 hr IPure16 200vol” was used with 5 hours of antibody coating and 16 hours of ChIP reaction at 4°C. 3 µg of the p300 antibody sc-585x Lot#E2610 (Santa Cruz) was used on 25 µL of the sonicated chromatin diluted with 75 µL of ChIP Buffer T. The ChIP products were reverse-crosslinked at 65°C for 6 hours with occasional vortexing. ChIP DNA was purified by reagents in the Auto IPure kit (Diagenode) but done manually following the IPure kit (Diagenode) instruction manual version V2_12-05-10. Sequencing library was prepared from the purified DNA by the SPRI-te Nucleic Acid extractor (Beckman Coulter) with SPRIworks Fragment Library System I cartridges according to manufacturer's protocol. Enrichment was done with 2× Phusion Master Mix, PE PCR primer 1.0 (Illumina) and a barcoded paired-end PCR primer 2.0. The library was sequenced in one paired-end lane on Illumina Genome Analyzer II. The sequencing reads of 36 bp were aligned to the hg18 genome by the short reads aligner bowtie [106] version 0.12.5 suppressing all alignments for reads that align to more than one location (-m 1). The sequencing and alignment statistics are listed in Table S3.
4,000 cells in 100 µL of parental media were seeded per well in a 96-well CellBIND clear plate. Twenty-four hours later, the medium was aspirated, each well was washed with 150 µL of PBS, and 100 µL of fresh serum-free media (DMEM with no phenol red) containing the indicated concentrations of drugs was added. Six to eight within-day biological replicates were performed for at least three between-day biological replicates for each treatment of each cell line. To make stock solutions of small molecule drugs, dasatinib, rapamycin, paclitaxel (LC Labs), ICG-001, SAHA, SB-431542 (Selleck Chemicals), 17-AAG (AG Scientific), PKF118–310, SB-505124 (Sigma), D4476, and harmine (Cayman Chemical) were dissolved in dimethyl sulfoxide (DMSO), 4-OHT (Sigma) in pure ethanol and PDTC (Sigma) in PBS. All stock solutions were stored in the dark at −20°C and diluted to the desired concentration in cell culture media immediately prior to treatment. After 72 hours of drug treatment, cell viability was measured by the WST-1 reagent (Roche Applied Science). 10 µL of WST-1 was added to each well, the plates were incubated at 37°C for four hours and absorbance at 450 nm was measured by Varioskan Flash Multimode Reader (Thermo Scientific). Raw signals were normalized by a linear mixed-effects model (see below) to eliminate between-day batch effects. Relative viability values were computed as the ratios between the normalized signals of drug-treated cells and the corresponding vehicle control cells. Curve fitting to the four-parameter log-logistic model and statistical tests of fitted parameter values were performed using the R package drc [107]. Differential response between U87DK and U87H cells were assessed using ANOVA on the fitted parameters.
To eliminate batch effects between plates and experiments performed on different days, a linear mixed-effects model was fitted to the viability measurements from the vehicle control wells on each plate: signal = grand mean+cell line+day of the experiment+plate+residual error, where cell line was classified as a fixed effect term and day of experiment and plate were classified as random effect terms. Model fitting was performed by the R lme4 package [108]. The viability measurements from treatment wells were normalized by subtracting the random effect estimates from all raw signal values before calculating the relative viabilities and fitting the dose-response curves.
The parameter controls the size of the PCST network by balancing the edge costs and node penalties. We would like the solution network to be small and connect a large number of termini by a small number of Steiner nodes. Therefore, we defined the “efficiency ratio” of a PCST solution as the ratio of included terminal nodes to Steiner nodes and selected a value of such that the solution network is small and has a good efficiency ratio. Specifically, we ran the algorithm with a wide range of values and computed the efficiency ratios for the solutions (Figure S6). We found that the efficiency ratio was relatively stable for values between 40 and 100, although the rate of increase of this ratio was slightly larger between of 50 and 60 (so increasing network size is more efficient in connecting the termini) and the network solution is intermediate in size. The PCST presented in Figure 2 was from .
A complete list of tyrosine phosphorylated peptides was downloaded from the supplementary data of [61], which were collected from mouse xenograft samples established from patient surgical specimens. Among the eight samples, two express wild-type EGFR at normal level, three express amplified level of wild-type EGFR, and three express amplified level of EGFRvIII. For each phosphorylated peptide, Student's t-test was used to compare the three samples with amplified EGFRvIII to the two samples with normal level of wild-type EGFR, and those peptides with p-value less than 0.05 were considered differentially phosphorylated in response to EGFRvIII.
For each node in the interactome, either inside or outside of the PCST solution, we found the edges connecting this node to the nodes in the PCST solution and summed up the confidence scores on these edges ( described above). Since the score on each edge is a log-likelihood confidence score of this interaction, a large value of this sum means this node has more high confidence interactions with the nodes in the PCST solution. We then ranked all the nodes in the interactome by this score.
Level 3 exon array data of the TCGA GBM project [53], [62] were downloaded from the TCGA data portal (https://tcga-data.nci.nih.gov/tcga/). The U87 cell line is reported to contain wild-type p53 and deletion of p16 [117], which match the mutation status of 99 patients as reported in the cBio Portal [118]. The EGFRvIII status of these 99 patients were determined by testing for significantly lower expression of the exon array probe sets that map to exon 2–7 of the EGFR gene (p<0.05 by Wilcoxon rank-sum test). 19 such samples were found and were thus labeled as EGFRVIII. Gene level expression values from the exon array were then analyzed for differential expression between the EGFRvIII and non-EGFRvIII samples by limma [103]. Hierarchical clustering of the differentially expressed genes (p-value<0.01) show that the 17 of the 19 patients with EGFRvIII mutation were grouped together among the major clusters (Figure S2).
Separately, for each transcription factor that has a motif in TRANSFAC, we used the MATCH program [119] to score for matches to the motif in the DNaseI hypersensitive regions that are more hypersensitive in the U87H cells than in the U87DK cells. This resulted in a set of factors for the U87H cells (U87H TF). The set of U87DK TF were found similarly from regions with higher hypersensitivity in the U87DK cells. If there was a motif match within 40 kb of the transcription start site of a gene, this gene was considered a target of the TF. Then for each TF, we computed the minimum hypergeometric (mHG) p-value [63] for testing the enrichment of its target genes in the list of all genes in the TCGA expression dataset ranked by log fold change between patients with the EGFRvIII mutation and those without. To determine the role of transcription factors in up-regulation, we calculated the mHG p-value of the genes ranked from highest changes in expression in EGFRvIII patients to lowest changes in expression. To determine the role of genes in down-regulation, we reversed the order of genes. P-values in Figure 3E and Figure 3F were calculated using Student's t-test.
We searched for sequence motifs that could predict which of the regions that were more hypersensitive in U87H than in U87DK cells were bound by p300. Differentially hypersensitive regions overlapping with p300 bound regions in U87H from ChIP-Seq were considered positive and the rest were negative. We then computed matches to the sequence motifs in the TRANSFAC vertebrate non-redundant set in all the positive and negative regions. The motif match scores were used in a feature selection procedure by Wilcoxon test in five-fold cross validation to rank these motifs by their ability to classify the positive and negative regions. To compute the significance of the p300 interactors inside the PCST to those outside of the PCST, the top 20 motifs were selected in each iteration of the cross-validation. Those correspond to proteins that interact with p300 were recorded and enrichment p-values were computed by Fisher exact test.
The raw data for ChIP-Seq, DNase-Seq and microarray experiments have been submitted to GEO under accession number GSE36902.
|
10.1371/journal.pcbi.1000888 | MetMap Enables Genome-Scale Methyltyping for Determining Methylation States in Populations | The ability to assay genome-scale methylation patterns using high-throughput sequencing makes it possible to carry out association studies to determine the relationship between epigenetic variation and phenotype. While bisulfite sequencing can determine a methylome at high resolution, cost inhibits its use in comparative and population studies. MethylSeq, based on sequencing of fragment ends produced by a methylation-sensitive restriction enzyme, is a method for methyltyping (survey of methylation states) and is a site-specific and cost-effective alternative to whole-genome bisulfite sequencing. Despite its advantages, the use of MethylSeq has been restricted by biases in MethylSeq data that complicate the determination of methyltypes. Here we introduce a statistical method, MetMap, that produces corrected site-specific methylation states from MethylSeq experiments and annotates unmethylated islands across the genome. MetMap integrates genome sequence information with experimental data, in a statistically sound and cohesive Bayesian Network. It infers the extent of methylation at individual CGs and across regions, and serves as a framework for comparative methylation analysis within and among species. We validated MetMap's inferences with direct bisulfite sequencing, showing that the methylation status of sites and islands is accurately inferred. We used MetMap to analyze MethylSeq data from four human neutrophil samples, identifying novel, highly unmethylated islands that are invisible to sequence-based annotation strategies. The combination of MethylSeq and MetMap is a powerful and cost-effective tool for determining genome-scale methyltypes suitable for comparative and association studies.
| In the vertebrates, methylation of cytosine residues in DNA regulates gene activity in concert with proteins that associate with DNA. Large-scale genomewide comparative studies that seek to link specific methylation patterns to disease will require hundreds or thousands of samples, and thus economical methods that assay genomewide methylation. One such method is MethylSeq, which samples cytosine methylation at site-specific resolution by high-throughput sequencing of the ends of DNA fragments generated by methylation-sensitive restriction enzymes. MethylSeq's low cost and simplicity of implementation enable its use in large-scale comparative studies, but biases inherent to the method inhibit interpretation of the data it produces. Here we present MetMap, a statistical framework that first accounts for the biases in MethylSeq data and then generates an analysis of the data that is suitable for use in comparative studies. We show that MethylSeq and MetMap can be used together to determine methylation profiles across the genome, and to identify novel unmethylated regions that are likely to be involved in gene regulation. The ability to conduct comparative studies of sufficient scale at a reasonable cost promises to reveal new insights into the relationship between cytosine methylation and phenotype.
| New methods that assay epigenetic modifications over the whole genome promise to reveal insights into cell differentiation and development [1]–[15]. Moreover, incorporation of genome-scale epigenetic data into case-control studies is now becoming feasible, and has the potential to be a powerful tool in the study of disease [16]. Recent evidence has suggested that epigenetic variation is heritable, and may underlie phenotypic variation in humans ([17], our own observation with the human and chimpanzee methylomes). Such comparative studies rely both on the ability to obtain genome-scale epigenomic information cheaply and efficiently, and on the availability of methods for analysis of the data produced.
Cytosine methylation, which in vertebrates is mostly confined to CG dinucleotides, cooperates with other epigenetic modifications to suppress transcription initiation [3], [18] (in this paper we denote a cytosine that is followed by a guanine as CG, rather than CpG, and similarly CCGG is equivalent to CpCpGpG. We leave the notation for CpG islands unchanged). In vertebrates, most CGs are methylated. However, early experiments with the methylation-sensitive restriction enzyme HpaII showed that unmethylated CGs are clustered in “HpaII Tiny Fragment Islands” [19]. These unmethylated islands are frequently active promoter elements. Methods used to annotate them on a genomic scale have been based only on sequence composition, because until recently genome-scale assessment of HpaII fragments has not been practicable. The methylation status of these regions, known as CpG islands, has not been considered in their annotation and is generally unknown. Genome-scale survey of the methylation status of CGs would enable the annotation of CpG islands based on their methylation states, rather than their sequence. Patterns of unmethylated islands differ among tissues, and changes in the methylation states of certain regions are associated with disease, particularly cancer [2], [3],[20]–[22].
High-throughput sequencing technologies have catalyzed the development of new methods for measuring DNA methylation. These methods can be broadly classified as methyltyping versus methylome sequencing, in analogy with genotyping versus genome sequencing for DNA. Methyltyping technologies allow for the assessment of genome-scale methylation patterns, while emphasizing low cost at the expense of high resolution. Assays based on sequencing avoid problems associated with hybridization to arrays. Examples include MethylSeq, which is based on digestion with a methylation-sensitive enzyme and is the focus of this paper, and RRBS which is based on digestion with a methylation-insensitive enzyme followed by bisulfite sequencing [3], [23]. In contrast to methyltyping, whole-genome bisulfite sequencing offers the ability to measure absolute levels of DNA methylation at single-nucleotide resolution [7], [24], [25], but it is expensive because it requires sequencing of whole genomes. The issues of cost and coverage are complicated by a number of other issues. In the case of bisulfite sequencing, conversion may not always be complete. Also, the analysis requirements for the different assays vary in difficulty. For these reasons, there has been a proliferation of methods whose pros and cons are constantly changing as sequencing technologies change. A recent analysis (Table 2 in [26]) suggested that MethylSeq is the method with the most favorable profile of pros and cons, with respect to the measures chosen for comparison. Table 1 summarizes characteristics of MethylSeq and of the most commonly used alternative methods. MethylSeq retrieves information spanning more of the genome than RRBS, because of a more favorable profile of fragment sizes produced by HpaII relative to MspI (see the discussion of size selection bias below and the Methods section).
MethylSeq is a convenient methyltyping strategy because it is cost-effective, requires only small amounts of material, and avoids bisulfite conversion. Briefly, the assay works by digestion of DNA with a methylation-sensitive enzyme (HpaII) that cuts unmethylated CGs at CCGG sites. Subsequent sequencing and mapping to the genome reveals unmethylated CGs (Figure 1). Although the experiment is relatively simple, interpretation of the sequencing data is confounded by the dependence of read depth at a given site on the methylation status of neighboring sites. This has limited the use of MethylSeq; previous studies either pointed out the need for a method of site-specific normalization [1], or attempted to deal with the bias by removing problematic HpaII sites from the analysis [5](resulting in the loss from the analysis of more than 19% of HpaII sites in CpG islands, see Methods).
In order to make effective use of MethylSeq for genome-scale methyltyping we developed a freely available program, called MetMap, that infers methylation at individual CGs by modeling biases inherent in MethylSeq experiments. An additional important feature of MetMap is the annotation of strongly unmethylated islands (SUMIs) which, as opposed to the current definition of CpG islands, incorporate information from both a reference sequence and genome-scale methylation data. We have validated MetMap's site-specific analysis, as well as its unmethylated-island annotation, with bisulfite sequencing of specific sites.
We demonstrate the use of MethylSeq with MetMap by methyltyping four male human individuals, and annotating their unmethylated islands. We show that the picture revealed by such analysis is sufficient to survey methylation states across the genome. Such analysis gives significant insight into the methylome of each specimen, inside and outside of CpG islands, at site specific resolution. We show evidence that the mean extent of methylation of an island is more informative than the methylation state of the different sites in the island, because the correlation between the methylation states of any two samples improved when considering the mean. MetMap identifies numerous unmethylated regions, of varying lengths, which have not previously been annotated as CpG islands and are associated with other features indicative of transcriptional function. We conclude that MetMap leverages the cost-efficiency and practical ease of MethylSeq to produce informative genome-scale methylation annotations (methyltypes) that are suitable for both region- and site-specific comparative and case-control studies.
The remainder of this paper is organized as follows. We begin by explaining in detail significant biases present in MethylSeq experiments. We then describe the MetMap framework, which is designed to correct for such biases, starting with a description of MetMap's graphical model and continuing with a description of the software's different outputs. We then describe the validation of MetMap's procedure, using the methyltypes of four human individuals, and our discovery of new unmethylated regions in the human neutrophil genome, found through the use of MetMap on MethylSeq data. Finally, we discuss the advantages of using MetMap with MethylSeq to generate and analyze large numbers of samples, and outline our plans for the extension of MetMap's framework.
We carried out MethylSeq on specimens of a single homogeneous and uncultured cell type, the neutrophil, from four male humans. HpaII fragments were size selected in the range 50–300bp and sequenced on a first generation Illumina Genome Analyzer yielding 23,731,359 32bp reads. Although longer reads are currently available, reads for our assay only need to be sufficiently long so that they can be mapped correctly to the reference genome. The reads were aligned to the reference human genome (hg18 [28]) with Bowtie [29] resulting in 18,218,420 alignments (Table S1), and each of the four samples was analyzed with MetMap.
To infer methylation states from read depths, we first segmented the genome into 6,000 non-overlapping regions (of size 0.5Mbp) that could be analyzed separately. For each region, MetMap returned methylation probabilities for those CCGG sites for which information on site-specific methylation could be obtained from the MethylSeq experiment, and annotated SUMIs. The CCGG subset contained 59% of the CCGG sites (4.8% of all CG sites) in the human genome. Of the sites for which information could be obtained, 80% (1,035,243 sites) were outside CpG islands as annotated in the UCSC Genome Browser [30], and 20% (257,540 sites) were inside, resulting in a two-fold enrichment of the proportion of CCGG sites that are in such CpG islands.
To test whether MetMap was correcting bias in the raw counts (Figure 3), we directly determined the methylation status of 22 regions in the human genome using bisulfite sequencing [31] (Methods). Each CG in the bisulfite experiment received a score from the set (0,0.25,0.5,0.75,1) based on the observed proportion of alleles in which that site was unmethylated in a sample [32].
We correlated the bisulfite scores (taken as being the true methylation status) with the read counts and with the MetMap predictions. Each of the 46 validated sites had three different scores for the extent to which it was unmethylated: a bisulfite score, a read count score, and a MetMap score. The Pearson correlation coefficient between the raw read counts and the bisulfite values was 0.67 while the Pearson correlation coefficient between the MetMap methylation score of those sites and the bisulfite values was improved to 0.90.
As the bisulfite scores may be an imprecise measure of the true extent of methylation (Methods) we tested the sensitivity of our results to the bisulfite scores. We “adjusted” bisulfite scores, assigning to each value of the two sets of scores, the read-count set and the MetMap predictions set, a separate “adjusted” bisulfite value, that is within a predetermined range. The range available for adjustment was determined by the initial bisulfite score (Methods). After this adjustment, the correlation coefficient of the read counts with the bisulfite scores was 0.73 and the correlation coefficient of the MetMap scores with the bisulfite scores was 0.95. While the correlation values increased as expected, the difference between the performance of MetMap and that of read counts remains similar. This indicates that the improvement in using MetMap instead of raw read counts was not due to the procedure by which bisulfite scores were assigned.
Examples of MetMap's ability to accurately detect partially and fully methylated sites are shown in Figure 3, Figure S1 and Text S2. Both the extent and variability of methylation in a region are better predicted by MetMap than by the read counts.
To determine which parameter might be more informative for genome-scale methyltyping, we compared methylation states for individual sites and for SUMIs between pairs of samples. Although the methylation status of individual sites within SUMIs was variable, the average probability of methylation for the whole SUMI was consistent across individuals (Figure 4). This observation suggests that the mean methylation state of a SUMI is more constrained than the methylation states of the individual sites within it, and thus a change in mean SUMI methylation is more likely to have functional consequences than a change at a specific site. Based on this, we propose that the mean SUMI methylation status is the more informative parameter for comparative or association studies.
Similar read counts at orthologous restriction sites in two or more samples indicate that their methylation status is similar; however determination of their true extent of methylation requires a statistical method such as MetMap. Thus the degree of consistency observed among MetMap's site-specific inferences for different samples is supported by the high correlation of the corresponding raw read counts (e.g.: a correlation of 0.667 between sample 1 and sample 4).
We mapped the 20,986 SUMIs present in at least one of the four individuals, and examined their relationship to purely sequence based definitions of CpG islands (Figure 5.a). Of the 20,986 SUMIs present in at least one of the four individuals, 4,652 do not overlap UCSC CpG islands, and 7,055 do not overlap the “bona fide” islands [33] with an epigenetic score larger than 0.5 (as recommended by Bock et al. [33], termed here BF islands). This result is consistent with the higher specificity, but lower sensitivity, of BF compared to UCSC island prediction. Details regarding the extent of overlap between SUMIs and the BF and UCSC islands can be seen in Table 2.
We compared the length distribution of our SUMIs with the length distributions of both the UCSC and BF islands (Figure 5.b). SUMIs were similar to BF islands, but the length distribution of the UCSC CpG islands resembled a geometric distribution. The process by which UCSC CpG islands are annotated will produce false positives that follow a geometric length distribution, with the number of false positive CpG islands increasing as a function of decreasing length (Methods). Since the length distributions of SUMIs and BF islands do not follow the same trend as the UCSC CpG island distribution, it is probable that at the shorter lengths the majority of predicted UCSC CpG islands are false positives. SUMIs did not overlap completely with BF islands: of the 21,626 BF islands, 13,899 were identified as SUMIs. BF islands are determined with a support vector machine that uses epigenetic data from multiple sources to train its prediction model. In contrast, MetMap's SUMI predictions originate from an experimental signal for unmethylation in the cell type analyzed. The probable explanation for the MetMap/BF discrepancy is that the two methods have used epigenetic data from different tissues. More data from distinct cell types will shed light on this issue.
We therefore validated with direct bisulfite sequencing five regions that are annotated as part of both a UCSC CpG island and a BF island, and did not overlap with SUMIs; we also sequenced three regions in BF islands that did not overlap with SUMIs or with UCSC CpG islands. In all cases those regions were validated as methylated in the neutrophil samples (Figure S1.j–q). This is consistent with the notion that while these islands might be unmethylated in other cell types, they are methylated in the neutrophil. We analyzed four cases of SUMIs with scores higher than 0.5 that overlapped UCSC CpG islands but not BF islands (Figure S1.a–d). In each SUMI a region was picked and bisulfite sequenced. All four regions were determined as fully unmethylated (all CG sites received a score of 1).
3,797 SUMIs do not overlap with BF islands or CpG islands, revealing new regions that are unmethylated in neutrophil cells. Of these novel SUMIs, 2,317 (61%) are within regions experimentally determined by the ENCODE project as open chromatin (Methods), 1,882 (50%) are within regions determined as conserved by the 17-way UCSC conservation track, 2,274 (60%) are within 2Kbp of RefSeq genes, and 837 (22%) are within 2Kbp of the 5′ end these genes (Figure. 5.c and Table 3).
Consistently with their similarity to conventional CpG islands, SUMIs are enriched near the transcription start sites (TSSs) of RefSeq genes, with a preference for the downstream side (Figure 6.a). We observe the same property also when we consider novel SUMIs alone (Figure 6.b), or when we consider only SUMIs that do not overlap UCSC CpG islands (Figure 6.c) or BF-islands (Fig. 6.d). This indicates that the distribution of novel SUMIs around the TSSs does not originate from a characteristic present in only one of these sets. We find that the proportion of SUMIs that maps at a distance from TSSs is larger for novel SUMIs than for all SUMIs, but that novel SUMIs have a degree of association with open chromatin similar to that observed for all SUMIs (Table 3); this suggests that novel SUMIs may often represent distal regulatory sequences.
The possibilities and potential of DNA methylation analysis with new sequencing technologies have been described as a “revolution” [26]. The vast number of methods for methylation analysis, along with many papers describing exciting findings, suggests that this revolution is underway. For the foreseeable future, methods that rely on the construction of a sequencing library produced by methyl-sensitive enzymes, followed by sequencing to measure methylation, are the practical approach for the analysis of large numbers of samples [26]. The efficient use of MethylSeq data requires a computational method that can infer true methylation states by considering biases inherent in the technical method. We have developed MetMap, which makes it possible to use MethylSeq for genome-scale methyltyping. MetMap facilitates the rapid calling of restriction-site-specific methylation, and of unmethylated regions, to produce methylation maps that are suitable for comparative analysis. Validation of MetMap calls with bisulfite sequencing shows that it compensates for bias present in the MethylSeq data. MetMap can combine experimental data and genome sequence to identify many strongly unmethylated islands (SUMIs) that were previously unannotated, suggesting that it can identify novel functional regions.
The annotation of experiment-specific strongly unmethylated islands (SUMIs) reconciles the original definition of CpG islands, based on their sensitivity to methylation-sensitive restriction enzymes [19] with the sequence-based definitions now used. The definition of SUMIs is functionally more exhaustive than the standard definition of CpG islands, since it couples sequence clues to methylation (abundance of CpGs) with experimental measurements of methylation. In our comparison of four humans, we noted that the average methylation states of SUMIs were more conserved among individuals than the methylation states of sites within them, suggesting that average methylation is more likely to be functionally important and so is a more informative parameter. SUMIs lie proximal to genes (77% are within 2Kbp of genes; 60% are within 2Kbp of the 5′ end), and are likely to be directly involved in regulation of gene expression.
Overall, we predicted 3,797 SUMIs that do not overlap UCSC CpG islands or BF islands. Their sequence conservation and correlation with open chromatin suggests that they are functional, but they are less frequently associated with transcription start sites than the general set of SUMIs. We speculate that many novel SUMIs are enhancers. The discovery of these novel regions illustrates the utility of using experimental data to annotate CpG islands.
As more methylation data becomes available, the MetMap program we have developed can be refined and improved. For example, with the advent of methylation-based case-control studies, it should be possible to define methyl-haplotypes and to leverage MetMap to explore variation within and between individuals. MetMap's graphical model can also be used to learn the dependencies between the methylation states of neighboring CG sites, which will expand the scope of MethylSeq experiments to include sites that are not directly assayed. As more data-types are produced together with methylation experiments, we envision expanding MetMap to include information from related genomes, and possibly other related measurements. Ultimately, we look forward to the coupling of methylation data with other functional information, including expression measurements and chromatin structure assays, to fully reveal the roles and consequence of DNA methylation.
Human samples were collected with CHORI's IRB approval after obtaining informed consent.
The MetMap software takes as input: (1) the mapped reads of a MethylSeq experiment, (2) the boundaries on the lengths of the fragments sequenced (determined by the size-selection step), and (3) a reference genome. It outputs two files: (1) a list of the HpaII sites in the scope of the experiment with their MetMap scores, and (2) a list of SUMI regions with their scores.
MetMap is free, open source software, and can be downloaded from the following site:
http://www.cs.berkeley.edu/meromit/MetMap.html
In the Methylseq experiment, information regarding the methylation state of a CCGG site can be obtained for the subset of CCGGs that are present on some fragment that has CCGG sites at its ends and that passes the size selection step (see “CG sites in the scope of the MethylSeq experiment” section for details). We computed the number of CCGGs of the human genome that fulfill this criterion to be 1,349,378.
In the RRBS protocol the genome is digested with the methylation-insensitive restriction enzyme MSPI (which cuts at CCGG sites), and the fragments of size 40–220bp are size-selected and have their ends sequenced (after bisulfite treatment). For the human genome RRBS determines the methylation status of CGs [23].
We determine the span of a methyltyping method by considering regions in which that method profiles methylation. By doing so we gain an insight to the broadness of a method with respect to the regions for which it profiles methylation. In MethylSeq, methylation status is determined for a subset of the CCGG sites and in RRBS methylation status is determined for CG sites that are within fragments that have CCGG sites on both ends and which are of relative short length (up to 220bp). We therefore computationally categorized all CCGG sites of the human genome as 1/0 based on the ability to infer their methylation state with each method. All regions (bounded by CCGG sites) in which all CCGG sites received a “1” were considered as spanned by the method. When determining the span for CpG islands, the regions spanned were computed in the same manner, but considered only regions within CpG islands. In cases that the CCGG nearest to an edge of the island was determined as “1” the region between that CCGG and the edge of the island was also considered as spanned.
In the protocol used for this study the genome is digested with the methylation-sensitive restriction enzyme HpaII and only CCGG sites that follow certain criteria (as outlined in Ball MP et al.) are considered for their methylation status. One of the requirements is that the CCGG site be at least 40bp away from at least one of its two neighboring CCGG sites. In the human genome 19% of the CCGG sites have both of their neighboring CCGG sites at a distance smaller than 40bp, and are therefore excluded from the analysis.
CpG islands in the UCSC track are defined in [34] as regions with a GC content of 50% or more, a length greater than 200bp, and a greater than 0.6 ratio of observed CG dinucleotides to the expected number based on the GC content of the segment. The segments to consider are collected by scoring all dinucleotides (+17 for CG and −1 for others) and identifying maximally scoring segments. Under this model, the probability that a region from the null model (sequence which is not an unmethylted region) fulfills these requirements increases as the length of the region decreases. This statement holds for models in which the probability of observing an A/T in the null model is larger than that of observing a C/G. This is indeed the case in humans. The likelihood of false positives in the UCSC CpG island set has been noted [33], [35].
We obtained whole blood from four young adult male humans and obtained neutrophils by first isolating peripheral blood mononuclear cells by Ficoll separation, then purifying neutrophils with anti-CD16 antibodies conjugated to magnetic beads (Miltenyi); we verified that the purified samples contain neutrophils by Wright-Giemsa staining and visual inspection by a hematologist. Genomic DNA was isolated using the DNeasy Blood & Tissue isolation kit (Qiagen), quantified using a Nanodrop spectrophotometer, and quality-controlled for purity with an Agilent Bioanalyzer. Genomic DNA (2g) was digested with HpaII under conditions that make it very likely that digestion is complete (overnight with enzyme boosting), fragments 50–300bp long were isolated from an agarose gel, and single-read sequencing libraries were prepared following the manufacturer's protocol (Illumina). Libraries were sequenced on a first-generation Illumina Genome Analyzer and 32 base reads were generated. Only reads beginning with “CGG” (the sequence of the ends produced by restriction with HpaII) were retained and analyzed with MetMap.
MetMap receives as input the output of the MethylSeq experiment mapped to a reference genome, the reference genome, and the minimal and maximal lengths of the fragments sequenced, denoted by and . MetMap generates its graphical model (the , and variables along with their dependency relations) from the reference genome and the values of and . Having the graphical model's structure in place, MetMap incorporates the MethylSeq data by assigning values to all variables (all fragments that may be sequenced in the MethylSeq experiment): each variable is assigned a score between 0 and 9, by fixing a dataset-specific “capping” value, denoted (see next section), and to each , with paired-end read count , assigning . In case of a single-end dataset a transformation approximates a paired-end dataset, and the data is scaled as if it were paired-end (Text S1). MetMap is modular, allowing for potential incorporation of methods that normalize for biases generally present in short-read sequencing technologies [36].
Several types of probability distributions annotate the dependencies between the variables of MetMap's model. The transition probabilities between each pair of adjacent variables of type and/or (which represent adjacent CGs) take into account the reference genome but not the MethylSeq data, and are the prior distribution over the hidden states. In case that the two adjacent variables are of type , they take on a state (denoted by ) from . The transition probabilities are , , and , where is the distance between and . Each function determines the probability that is in its island state given that is in the given island state, that a CG is observed at and that no CGs are observed for the distance of . The parameters 0.00031434, 0.0257, 0.10178, 0.01298 and 0.013, respectively determine the probability of entering an island, of leaving an island, the probability of the sequence ‘CG’ occurring in an island, and out of one, and the initial probability that a site of the genome is in an island. These parameters were set using maximum likelihood estimates, calculated using chromosomes 21 and 22 of the human genome (Text S1). In cases where the successor variable of a pair of adjacent variables is of type , the methylation value of the state is considered. MetMap's current version assumes independence of the neighboring sites' methylation values, given the island values.
Parameters 0.2269, 0.05 and 0.7231 determine the probabilities of having an , or methylation value, given an unmethylated island status (). Parameters 0.8087, 0.05 and 0.1413 determine these probabilities, given an outside of unmethylated island status () (Text S1). The transition function to any state of is determined as the product of the transition probability considering only island values (as specified above) and the probability of observing the methylation value of the state at hand, given its island value. The third type of probability distribution in MetMap annotates the dependencies between the and variables. Each variable is dependent on the methylation values of the variables on the fragment it represents (Figure 3.b). Therefore, a dependency function is denoted for each variable as , where is a state of and is some configuration (assignment) of the values of all the restriction sites on fragment . We limit the number of variables in the interior of such structures to at most 3 (by random choice from the interior variables), and unite methylation configurations that are equivalent with respect to the probability function, resulting in lookup tables of size at most 5×10. The parameters in the lookup table were determined using a linear program that takes into consideration the internal constraints of the probability distributions (Text S1). Artificial restriction of the number of interior variables is not common because the maximum fragment length imposed by the size selection is relatively short.
MetMap infers the posterior probabilities of its hidden states by building the junction-tree graph and using belief propagation [27]. The structure of the graph makes this computation tractable and efficient: the running time for the inference procedure is less than an hour for large chromosomes on a small sized cluster.
MetMap generates two output files. One holds for each HpaII site in the scope of the experiment a MetMap score, indicating the inferred frequency of alleles in the MethylSeq sample that are unmethylated at that site. The second file holds the coordinates and scores of the annotated SUMIs.
To generate a value for , MetMap builds a histogram of the read count intensities for the subset of fragments of length 50–80bp, which do not hold internal restriction sites, and are located inside UCSC CpG islands. The fragments participating in the histogram contain a greatly reduced amount of bias (due to the lack of restriction sites in their interior) and are assumed to be mostly unmethylated (as they are in CpG islands). Under the assumption that the distribution of the histogram is close to Poisson, because the sequencing of fragments is equivalent to sampling them from the digest, we assume the variance is equal to the mean, and take to be the value two standard deviations away from the mean of the distribution. The procedure described is carried out to avoid setting in a way which is harshly influenced by PCR amplification bias, a phenomenon that causes some sites of the genome to receive extremely high counts, regardless of the extent to which they are methylated.
MetMap outputs methylation scores only for the HpaII sites (CCGGs) that are in the scope of the MethylSeq experiment. A HpaII site is in the scope of an experiment if and only if it lies on some fragment that has HpaII sites at its ends, and is of length such that , where and are the minimal and maximal fragment lengths for a specific MethylSeq experiment. MetMap's graphical model identifies these sites; they are all HpaII site variables ( variables) that have an edge to some fragment variable ( variable). Importantly, this condition does not require a site to be at an end of a fragment that satisfies the length requirements; a site may be in the interior of such a fragment.
The SUMI regions annotated by MetMap are the union of two sets of regions. The first set consists of those continuous regions in which each or variable of the MetMap model (CG sites) received a probability of being in an unmethylated island () that is larger than 0.1, and in which the MethylSeq data directly supports the presence of at least two fragments. This set will include regions with relatively weak direct experimental evidence but with strong sequence evidence for being unmethylated. The second set is generated by setting a 600bp interval around each HpaII site that had a value smaller than 0.1 and a value higher than the prior probability of being unmethylated outside of an unmethylated-island (0.1663). All overlapping windows are concatenated and the regions taken are those in which at least 30% of the HpaII sites had a p(U) larger than the prior-set threshold (0.1663), and in which the MethylSeq data directly supports the presence of at least two fragments. This set includes regions with weaker sequence support for unmethylation, but with extensive evidence that they are unmethylated. Each SUMI receives a score, specifying the mean of the MetMap scores at all of the sites within the SUMI.
The SUMI lists for the four human neutrophil samples can be found at:
http://www.cs.berkeley.edu/meromit/SUMIs_Human_Neutrophil/
DNA was treated with the MethylEasy bisulfite conversion kit (Human Genetic Signatures), PCR-amplified with locus-specific primers that recognized human target sequences, and sequenced using standard Sanger chemistry. Since all epialleles from a single specimen were sequenced in bulk in the same mixture, we estimated the ratio of unmethylated/methylated alleles at each CG in the sequence by examining the relative heights of the ‘C’ and ‘T’ traces in the sequencing output. Each CG site received a score from the set (0,0.25,0.5,0.75,1), based on the relative C/T peak height [32]. A score of 1 indicates the site is fully unmethylated, meaning that only the ‘T’ trace was observed at the C position of a given CG, while a score of 0 indicates the site is fully methylated, meaning that only the ‘C’ trace was observed at the C position of a given CG.
We tested the extent to which our results may be affected by the representation of the bisulfite scores on a discrete five-point scale, since the true proportion of alleles that are unmethylated is a close to continuous measure. Each data point was assigned an ‘adjusted’ bisulfite score, within a tolerance window specified by the true bisulfite value of that data point. The ‘feasible ranges’ allowed for the ‘adjusted’ bisulfite scores were as follows: (0,0.15) for a 0 bisulfite score, (0.15,0.35) for a 0.25 score, (0.35,0.65) for a 0.5 score, (0.65,0.85) of a 0.75 score and (0.85,1) for a 1 score. For example, for a site with bisulfite score 0.25, read count score 0 and MetMap prediction 0.4 we would get two pairings (0.15,0) for (adjusted-bisulfite, read count score), and (0.35,0.4) for (adjusted-bisulfite, MetMap score). The score ranges were based on an assumption that assignments of “0.5” scores were the least precise. The adjustment of the bisulfite score to the read counts was done by generating a normalized read count value, in the 0–1 range, using the same “capping” value as MetMap.
One file of open chromatin was compiled from:
ftp://hgdownload.cse.ucsc.edu/goldenPath/hg18/encodeDCC/wgEncodeChromatinMap/
using the files: wgEncodeUncFAIREseqPeaksH1hesc.narrowPeak
wgEncodeUncFAIREseqPeaksNhek.narrowPeak
wgEncodeUncFAIREseqPeaksGm12878V2.narrowPeak
wgEncodeUncFAIREseqPeaksHuvec.narrowPeak
wgEncodeUncFAIREseqPeaksPanislets.narrowPeak
|
10.1371/journal.pgen.1002806 | TDP-1/TDP-43 Regulates Stress Signaling and Age-Dependent Proteotoxicity in Caenorhabditis elegans | TDP-43 is a multifunctional nucleic acid binding protein linked to several neurodegenerative diseases including Amyotrophic Lateral Sclerosis (ALS) and Frontotemporal Dementia. To learn more about the normal biological and abnormal pathological role of this protein, we turned to Caenorhabditis elegans and its orthologue TDP-1. We report that TDP-1 functions in the Insulin/IGF pathway to regulate longevity and the oxidative stress response downstream from the forkhead transcription factor DAF-16/FOXO3a. However, although tdp-1 mutants are stress-sensitive, chronic upregulation of tdp-1 expression is toxic and decreases lifespan. ALS–associated mutations in TDP-43 or the related RNA binding protein FUS activate the unfolded protein response and generate oxidative stress leading to the daf-16–dependent upregulation of tdp-1 expression with negative effects on neuronal function and lifespan. Consistently, deletion of endogenous tdp-1 rescues mutant TDP-43 and FUS proteotoxicity in C. elegans. These results suggest that chronic induction of wild-type TDP-1/TDP-43 by cellular stress may propagate neurodegeneration and decrease lifespan.
| TAR DNA Binding Protein 43 (TDP-43) is implicated in several human age-dependent neurodegenerative disorders, but until now little was known about TDP-43's role in the aging process. Here we used the nematode Caenorhabditis elegans to study the role of the TDP-43 orthologue tdp-1 in aging and neurodegeneration. In this study we discovered that tdp-1 is a stress-responsive gene acting within the Insulin/IGF signaling pathway to regulate lifespan and the response to oxidative stress. We found that, although worms missing tdp-1 were stress-sensitive, elevated expression of tdp-1 was toxic. We asked if tdp-1 also responded to the stress caused by toxic proteins found in Amyotrophic Lateral Sclerosis (ALS). Using worm models for ALS, we discovered that mutant TDP-43 generated oxidative stress and induced tdp-1 expression with negative consequences on neuronal function and lifespan. Consistently, removing tdp-1 rescued toxicity in our worm ALS models. tdp-1's role in the cellular stress response likely reflects an ancient adaptation to deal with unfavorable environmental conditions that is inappropriately activated and maintained by genetic mutations leading to proteotoxic and oxidative stress. We predict that similar mechanisms may exist in humans, helping explain the involvement of TDP-43 in a growing number of neurodegenerative disorders.
| TDP-1 is the Caenorhabditis elegans orthologue of the multifunctional DNA/RNA binding protein TDP-43 (TAR DNA Binding Protein 43). Mutations and accumulations of TDP-43 have been found in patients with Amyotrophic Lateral Sclerosis (ALS), Frontotemporal Dementia, and in a growing number of neurodegenerative disorders [1]. As part of its numerous roles in RNA metabolism, TDP-43 is a component of the cytoplasmic ribonucleoprotein complexes known as stress granules that form in response to environmental stresses like heat shock, oxidative and osmotic stress among others [2], [3]. ALS is an age-dependent neurodegenerative disorder and given that TDP-43 is a stress responsive protein we hypothesized that TDP-1 may regulate longevity and the cellular stress response.
In worms a major axis of stress-response signaling and longevity is the Insulin/IGF-signaling (IIS) pathway. IIS follows an evolutionarily conserved and genetically regulated pathway that regulates numerous processes including development, metabolism, fecundity, cellular stress resistance and aging [4]. In C. elegans, IIS initiates a phosphorylation cascade through the DAF-2/Insulin-IGF receptor that phosphorylates the forkhead transcription factor DAF-16 and retains it in the cytoplasm [5]–[8]. Hypomorphic daf-2 mutants relieve DAF-16 phosphorylation causing DAF-16 to translocate to the nucleus where it activates a large number of genes resulting in increased lifespan and augmented stress resistance [9]. daf-2 mutants are also resistant to numerous stresses including oxidative, osmotic, thermal and proteotoxicity [10]. The IIS pathway likely employs multiple mechanisms to combat a variety of cellular insults but little is known about how these separate functions are regulated by DAF-16.
We attempted to address this issue by asking whether TDP-1 participated in the cellular stress response and longevity pathways in C. elegans. We asked if TDP-1 participated in the IIS pathway to specify developmental, stress response, and longevity outcomes. Given the importance of human TDP-43 to neurodegeneration we also asked if tdp-1 regulated age-dependent proteotoxicity. Our work points to TDP-1 as a key stress response protein at the crossroads of IIS, proteotoxicity and endoplasmic reticulum (ER) stress. Moreover, persistent induction of TDP-1 may actively promote neurodegeneration.
Downregulation of DAF-2 extends lifespan and promotes stress resistance via regulation of DAF-16 transcriptional activity [9]. To determine if tdp-1 participated in the IIS pathway we used a deletion mutant tdp-1(ok803) that removes the two RNA Recognition Motifs of TDP-1 (Figure S1A). Looking directly at the IIS pathway and longevity, daf-2(e1370) animals were long-lived but this phenotype was significantly reduced in a daf-2(e1370);tdp-1(ok803) double mutant strain (Figure 1A, Table S1). Conversely, daf-16(mu86) mutants were short-lived compared to wild type worms and the inclusion of tdp-1(ok803) did not further reduce daf-16 mutants' lifespan (Figure 1A, Table S1). We confirmed that tdp-1 regulated daf-2 mediated longevity with RNA interference and observed that tdp-1(RNAi) reduced the extended lifespan of daf-2 mutants (Figure 1B, Table S1). These data suggest that tdp-1 regulates the lifespan phenotypes of reduced IIS and the inability of tdp-1 to further reduce daf-16 lifespan suggests that this effect is dependent on daf-16.
Surprisingly, despite tdp-1(ok803) limiting the extreme lifespan of daf-2 mutants we observed that at 20°C tdp-1(ok803) mutants had a small but significant increase in lifespan versus N2 controls (Figure 1C, Table S1). However this effect was lost when the worms were grown at 25°C (Figure 1D, Table S1) suggesting that tdp-1 may respond to stressful environmental conditions. Consistently, opposite to reducing tdp-1 function, overexpression of TDP-1 fused to GFP from the endogenous tdp-1 promoter (tdp-1p::TDP-1::GFP) greatly reduced lifespan compared to wild type N2 worms and tdp-1 deletion mutants at 20°C (Figure 1E). The lifespan of a tdp-1(ok803);TDP-1::GFP strain was greater than TDP-1::GFP alone, but less than either N2 worms or the tdp-1(ok803) deletion mutant (Figure 1E, Table S1). Finally, TDP-1::GFP transgenics showed a further decrease in mean and maximum lifespan when grown at 25°C compared to N2 worms (Figure 1F, Table S1). These data show that tdp-1 has a dose-dependent effect on lifespan in worms and that lifespan is sensitive to tdp-1 expression levels, which is consistent with studies of TDP-43 in flies [11] and mice [12] where overexpression also reduces lifespan.
Intimately linked to IIS regulation of lifespan in worms is dauer formation and stress resistance [4]. daf-2(e1370) mutant larvae show near complete dauer formation when grown at 25°C, but this phenotype was not altered in daf-2(e1370);tdp-1(ok803) mutants (Figure S2A) suggesting that tdp-1 does not function in the dauer formation axis of IIS in worms.
To further define the role of tdp-1 in the in vivo stress response, we challenged worms against juglone induced oxidative stress. Juglone is a natural product derived from the black walnut tree that raises intracellular oxide levels [13]. Adult wild type N2 worms transferred to juglone plates showed complete mortality after 14 hours, while daf-2(e1370) worms were completely resistant (Figure 2A). If tdp-1 were required for protection against stress we would expect the mutant tdp-1 worms to be hypersensitive to juglone-induced toxicity. Interestingly, tdp-1(ok803) mutants were more sensitive to juglone than N2 worms, and tdp-1(ok803) completely abolished daf-2's resistance to juglone in the daf-2(e1370);tdp-1(ok803) double mutant strain (Figure 2A). To corroborate the role of tdp-1 in resistance to oxidative stress we used hydrogen peroxide, which is another oxidative stress enhancer. Tested on hydrogen peroxide plates we observed that tdp-1(ok803) animals were more sensitive than wild type N2 worms, and tdp-1(ok803) sensitized the normally resistant daf-2 animals to stress-induced mortality (Figure 2B). These data show that tdp-1 is required for daf-2 mediated resistance to oxidative stress. Given the effect of tdp-1(ok803) on daf-2 mutants' stress resistance compared to the lack of effect on dauer phenotypes we wondered if tdp-1 might have a more specific role in the stress response axis of IIS.
Not all forms of stress are equal at the cellular level so we investigated several additional forms of cellular stress including hypertonic stress with elevated sodium chloride (NaCl) or sorbitol, increased temperature, low oxygen and ultraviolet irradiation. Compared to N2 worms, tdp-1 mutants were sensitive to hypertonic stress from NaCl (Figure 2C). Interestingly, both daf-2(e1370) and daf-2(e1370);tdp-1(ok803) mutants were resistant to this hypertonic stress suggesting that tdp-1 may not function, or may function in parallel to the IIS pathway to regulate the response to osmotic stress. To confirm this hypothesis we used sorbitol, another hypertonic stress producer, and similarly observed that tdp-1(ok803) mutants were sensitive to osmotic stress but daf-2(e1370) and daf-2(e1370);tdp-1(ok803) mutants were both resistant after either 14 hours (Figure 2D) or 48 hours of exposure (Figure S2B). Finally, tdp-1 had no discernable role in the response to thermal stress, hypoxia, or radiation since there was no difference between N2 and tdp-1(ok803) worms, as well as daf-2(e1370) and daf-2(e1370);tdp-1(ok803) mutants in response to these stresses (Figure S2C–S2E). Thus, tdp-1 functions in the IIS pathway to specify the response to oxidative stress and lifespan, but tdp-1 is dispensable for the regulation of hypertonic stress by the IIS pathway. A second tdp-1 deletion allele, ok781, was available for investigation and although we observed that tdp-1(ok781) animals had a modest increase in lifespan, they were statistically indistinguishable from wild type N2 worms in stress assays (Figure S3).
A component of DAF-2's response to cellular stress is induction of DAF-16 stress response genes [9]. Thus we examined if DAF-16 targets implicated in oxidative stress signaling were affected by tdp-1. We used RT-PCR to examine expression levels of three DAF-16 genes including the catalases ctl-1 and ctl-2, as well as the superoxide dismutase sod-3 in various genetic backgrounds. First of all, no expression of tdp-1 was observed in strains containing the deletion allele tdp-1(ok803) (Figure 2E). Next we observed that ctl-1 and clt-2 expression levels were unaltered by the deletion of tdp-1, either alone or in combination with daf-2(e1370) (Figure 2E). However, we observed that sod-3 expression was greatly reduced in daf-2(e1370);tdp-1(ok803) animals (Figure 2E). These data suggest that under low IIS conditions tdp-1 is required for the expression of certain DAF-16 targets implicated in oxidative stress and may partially explain the sensitivity of tdp-1 mutants to juglone and hydrogen peroxide.
Given the remarkable sensitivity of tdp-1(ok803) mutants to oxidative and osmotic stresses we next wanted to see if there were in vivo changes in TDP-1 expression in response to these two stress conditions. In wild type unstressed animals TDP-1::GFP is lowly expressed (Figure 3A), primarily nuclear and is expressed in most tissues (Figure 3B–3D). We next tested if shifting TDP-1::GFP animals to either NaCl or juglone plates affected TDP-1::GFP expression. Young adult TDP-1::GFP worms were exposed to either NaCl or juglone for 90 minutes and live animals were examined for changes in TDP-1::GFP expression. Strikingly, exposure to hypertonic or oxidative stress greatly increased TDP-1::GFP expression compared to untreated control worms as detected by visual inspection and quantification of images (Figure 3E and Figure S4A). Endogenous tdp-1 was not required for expression of the TDP-1::GFP transgene since expression was induced by stress in tdp-1(ok803) deletion mutants (Figure 3F and Figure S4B). As a control for generic effects of stress on transgene expression we tested two other strains and observed no induction with a neuronal GFP reporter, while juglone and NaCl induced expression of the well-characterized sod-3p::GFP reporter (Figure S5).
Having shown that our TDP-1::GFP transgenics are potent stress reporters, using daf-2 and daf-16 mutants we tested if IIS regulated this effect. We observed that TDP-1::GFP was highly induced in the daf-2(e1370);TDP-1::GFP strain (Figure 3G) compared to TDP-1::GFP controls (Figure 3A) and was further elevated in daf-2(e1370) mutants exposed to stress (Figure 3G and Figure S4C). Given the pleiotropic phenotypic effects of daf-2 mutations we tested another allele and observed that opposite to e1370, TDP-1::GFP expression remained low under normal and stress conditions in the presence of the daf-2(e1368) mutation (Figure 3H and Figure S4C) suggesting a complex role for IIS in TDP-1 expression.
Looking deeper into the IIS pathway, previous research has shown that the increased stress resistance of daf-2 mutants is dependent on daf-16 mediated nuclear transcription [14]. TDP-1::GFP was not upregulated by mutation in daf-16 under normal conditions (Figure 3I). Next, to directly test if daf-16 was essential for the stress dependent induction of tdp-1 we exposed daf-16(mu86);TDP-1::GFP transgenics to stress. Exposure to NaCl continued to induce TDP-1::GFP expression in daf-16 mutants (Figure 3I) while treatment with juglone failed to induce TDP-1::GFP expression in the daf-16 mutants (Figure 3I and Figure S4D). These data suggest that the expression of TDP-1 in response to oxidative stress is dependent on daf-16, while induction of TDP-1 by osmotic stress is independent, which is consistent with our data showing that tdp-1 is not required for daf-2's resistance to hypertonic stress (Figure 2C and 2D). Additionally, the upregulation of TDP-1::GFP in daf-2(e1370) mutants was abolished by daf-16(RNAi) directly linking the IIS pathway to tdp-1 expression (Figure S6A and S6B, and Figure S4E). Finally, stress resistance and lifespan phenotypes from low IIS requires nuclear localization of DAF-16 and transcriptional activation of target genes [6], [9], [14] but we observed that tdp-1(ok803) had no effect on DAF-16::GFP stress-induced nuclear localization (Figure S6C–S6F).
Since TDP-43 is known to shuttle between the nucleus and cytoplasm we wondered if the subcellular distribution of TDP-1::GFP was altered under stress conditions. Examining TDP-1::GFP animals under high magnification we observed that when exposed low IIS from the daf-2(e1370) mutation TDP-1::GFP was no longer restricted to the nucleus and was observed in the cytoplasm (Figure 3J and 3K). These data suggest stress and/or low IIS significantly influences the expression and cellular distribution of TDP-1 proteins.
To make certain that the induction of TDP-1 by stress was not a phenomenon restricted to our transgenic reporter strain we examined endogenous TDP-1 protein levels under normal and stress conditions. Using a monoclonal antibody against C. elegans TDP-1 (Figure S7) and western blotting we observed a significant increase in TDP-1 protein levels in N2 worms exposed to hyperosmotic or oxidative stress compared to untreated controls (Figure 4A). To confirm the opposing effects of daf-2 mutations on TDP-1 expression we examined daf-2(e1370) and daf-2(e1368) mutants under normal and stressed conditions. As seen with our TDP-1::GFP reporter strain, we observed that TDP-1 protein levels were elevated in daf-2(e1370) animals, and we observed a further increase in TDP-1 levels in daf-2(e1370) animals exposed to oxidative stress (Figure 4B). Consistent with what we observed in the daf-2(e1368);TDP-1::GFP animals, we observed that TDP-1 protein levels remained low in daf-2(e1368) animals under normal and stress conditions (Figure 4C). Finally, endogenous TDP-1 protein was lowly expressed in daf-16 mutants, greatly increased upon exposure to osmotic stress, but again remained low in daf-16 animals treated with juglone (Figure 4D) in agreement with our findings with the daf-16(mu86);TDP-1::GFP reporter strain. In summary, TDP-1 is a stress responsive protein whose expression is greatly influenced by the IIS pathway especially in the context of oxidative stress.
Our findings begin to outline a complex role for TDP-1 in lifespan and the cellular stress response in relation to the IIS pathway. As one of our main interests is understanding the role of aging and stress signaling in the context of age-dependent neurodegeneration we next investigated how proteotoxicity contributed to these processes.
A feature of many late neurodegenerative disorders is proteotoxic stress caused by misfolded proteins and mutations in human TDP-43 are believed to cause proteotoxicity leading to neuronal dysfunction and cell death [15]. We examined this directly with transgenic worm strains expressing human wild type and mutant TDP-43 in worm motor neurons [16]. HSP-4 is a C. elegans Hsp70 protein orthologous to mammalian Grp78/BiP, and the transgenic C. elegans hsp-4p::GFP reporter is activated in response to misfolded proteins within the endoplasmic reticulum (ER), including chemically by compounds like tunicamycin (Figure 5A and 5B) [17]. Using this reporter we observed that transgenic strains expressing wild type TDP-43 did not induce reporter expression whereas transgenics expressing mutant TDP-43 strongly induced hsp-4p::GFP expression (Figure 5C and 5D). These data indicate that mutant TDP-43 toxicity may activate the ER unfolded protein response (UPRER). We did not observe induction of GFP in reporter strains for the mitochondrial chaperones hsp-6/Hsp70 and hsp-60/Hsp10/60 or the cytoplasmic chaperone hsp-16.2/Hsp16 (Figure S8).
Since we identified TDP-1 as a stress responsive protein we wondered if it also responded to ER and proteotoxic stress. We noticed increased expression of TDP-1::GFP in worms grown on plates with the ER stress inducing compound tunicamycin compared to untreated controls (Figure 6A and 6B). Looking at proteotoxic stress with our TDP-43 transgenics we observed that wild type TDP-43 had no effect on TDP-1::GFP expression while mutant TDP-43 strongly induced TDP-1 expression (Figure 6C and 6D, and Figure S4F). To confirm that induction of TDP-1::GFP was due to protein misfolding and proteotoxicity and not to an artifact of mutant TDP-43 transgenes we examined another proteotoxicity model based on the expression of wild type and mutant human FUS in motor neurons [16]. FUS is a nucleic acid binding protein related to TDP-43 that has also been implicated in ALS and dementia and the expression of an ALS-linked FUS allele in C. elegans motor neurons produces degenerative phenotypes similar to mutant TDP-43 [16]. Similar to the TDP-43 model, we observed no effect on TDP-1 expression from wild type FUS but mutant FUS greatly induced TDP-1 expression (Figure 6E and 6F, and Figure S4F). These data suggest that misfolded mutant TDP-43 and FUS initiate the UPRER, which in turn activates expression of TDP-1. Finally, activation of TDP-1 by ER stress converged on the IIS since induction of TDP-1::GFP expression by tunicamycin was blocked by a null mutation in daf-16 (Figure 6G).
A consequence of processing misfolded proteins within the ER is the production of reactive oxygen species as part of what is termed the integrated stress response [18]. Our data show that tdp-1 responds to oxidative stress in a daf-16 dependent matter so we hypothesized that the ER stress produced from mutant TDP-43 and FUS may generate oxidative stress thus linking proteotoxicity to the IIS pathway. To test this hypothesis we stained our transgenic TDP-43 and FUS strains with dihydrofluorescein diacetate (DHF), a compound that fluoresces when exposed to intracellular peroxides associated with oxidative stress [18]. We observed no DHF signal from wild type TDP-43 and FUS transgenics but strong fluorescence from mutant TDP-43 and transgenics (Figure 7A–7D). These data suggest that in addition to activating the UPRER mutant TDP-43 and FUS generate oxidative stress. To further establish the link between proteotoxicity and the IIS we examined the subcellular localization of DAF-16. DAF-16::GFP is typically cytoplasmic under non-stressed and/or low insulin signaling conditions (Figure 7E). Consistently we observed nuclear localization of DAF-16 in a TDP-43[A315T];DAF-16::GFP strain indicating that mutant TDP-43 causes cellular stress that is transmitted by the IIS pathway (Figure 7F).
Several studies have shown that reduced daf-2 function suppresses proteotoxicity [19]–[22]. However we found that daf-2 mutations have opposite effects on TDP-1 expression and if elevated TDP-1 expression were cytotoxic then we would expect to see opposing effects of daf-2(e1368) and daf-2(e1370) on TDP-43 toxicity in our models. To examine this directly we created daf-2(e1368);TDP-43[A315T] and daf-2(e1370);TDP-43[A315T] strains and scored paralysis phenotypes. Interestingly, we observed that daf-2(e1368) suppressed paralysis while daf-2(e1370) enhanced paralysis compared to TDP-43[A315T] alone (Figure 8A). This intriguing finding suggests that daf-2 can have variable effects on proteotoxicity. Furthermore, daf-2(e1368) significantly reduced the motor neuron degeneration caused by mutant TDP-43 [16] while daf-2(e1370) enhanced degeneration (Figure 8B). Since the IIS pathway is believed to regulate the expression of numerous protein quality control genes we examined if the two daf-2 alleles had an effect on misfolded mutant TDP-43 with a biochemical assay to detect protein aggregation. Here, homogenized protein extracts from transgenic worms are separated into two fractions, supernatant (detergent-soluble) and pellet (detergent-insoluble) and by western blotting with human TDP-43 antibodies we have previously shown that mutant TDP-43 is prone to aggregation and is highly insoluble [16]. Looking at protein extracts from TDP-43[A315T], daf-2(e1368);TDP-43[A315T], and daf-2(e1370);TDP-43[A315T] animals we observed that daf-2(1368) greatly reduced the amount of insoluble TDP-43 compared to control transgenics while daf-2(e1370) had no effect (Figure 8C). In summary these data suggest that different daf-2 alleles have widely variable effects on proteotoxicity but are consistent for each allele: e1368 reduces mutant TDP-43 insolubility, suppresses mutant TDP-43 induced paralysis, neurodegeneration, and keeps TDP-1 expression low while e1370 has no effect on protein insolubility, enhances paralysis, neurodegeneration and induces TDP-1 expression.
Although loss of tdp-1 sensitizes worms to oxidative and osmotic stress, elevated and chronic expression of TDP-1 leads to decreased lifespan suggesting that the induction of TDP-1 by proteotoxicity, oxidative stress or the IIS pathway may exacerbate neuronal toxicity and decrease neuronal survival. We directly tested this hypothesis by crossing TDP-1::GFP worms with our mutant TDP-43[A315T] and FUS[S57Δ] transgenics and scored for survival and the onset of paralysis. We observed that TDP-43[A315T] and FUS[S57Δ] strains containing the TDP-1::GFP transgene had significantly decreased lifespans compared to control transgenics (Figure 9A and 9B, Table S1). The expression of either mutant TDP-43[A315T] or FUS[S57Δ] causes motility defects leading to progressive paralysis and strains also expressing TDP-1::GFP had accelerated rates of paralysis compared to single transgenic controls (Figure 9C and 9D). These data suggest that induction of wild type TDP-1 expression by proteotoxicity has negative consequences on survival and neuronal function.
To rule out the possibility that the negative effects observed were simply due to transgene effects from TDP-1 overexpression we predicted that removing endogenous TDP-1 would reduce proteotoxicity. To test this we constructed TDP-43[A315T];tdp-1(ok803) and FUS[S57Δ];tdp-1(ok803) double mutant strains and observed that these animals had a significantly lower rate of paralysis compared to single transgenic TDP-43[A315T] or FUS[S57Δ] worms (Figure 9E). The expression of mutant TDP-43[A315T] or FUS[S57Δ] is accompanied by the age-dependent degeneration of motor neurons that was reduced in tdp-1(ok803) mutants (Figure 9F). Finally, to examine if protein misfolding was reduced in tdp-1(ok803) strains co-expressing mutant TDP-43 or mutant FUS, we examined the solubility of mutant TDP-43 and FUS proteins with our biochemical assay [16]. We observed no change in protein solubility after deletion of the endogenous tdp-1 suggesting that the protective effects are not due to down-regulation or clearance of mutant proteins (Figure 9G and 9H).
This work identified tdp-1 as a key stress responsive gene at the interface of longevity, stress resistance and neurodegeneration (Figure 10). The role of TDP-1 in lifespan is complex and suggests that worms like other species are sensitive to TDP-1/TDP-43 levels [23]–[26]. TDP-1 overexpression reduces lifespan while deletion of tdp-1 in unstressed worms promotes a modest increase in lifespan but leaves worms sensitive to specific environmental stresses.
TDP-1 also has a complex role in the cellular stress response. We showed here that TDP-1 specifies the response to cellular stress with roles in oxidative, osmotic and protein-misfolding stress, but independent of high temperature, low-oxygen and damage from radiation. TDP-43 is a component of the ribonucleoprotein complexes known as stress granules that form under stressful conditions where they perform molecular mRNA triage where mRNAs are sorted for storage, degradation, or translation during stress and recovery [3]. The stress-inducible aspect of TDP-1/TDP-43 function likely reflects an ancient mechanism for enduring acute environmental stress until conditions improve. While the role of TDP-43 in response to acute stress is being actively studied [27]–[29], its role in response to chronic stresses like protein misfolding during aging is unknown.
Although the tdp-1 alleles ok803 and ok781 are predicted to be molecular null alleles they did not show identical phenotypes. Both alleles extended lifespan but ok781 did not affect several stress phenotypes. The difference between the two alleles is that ok781 still maintains the putative Nuclear Localization Signal (Figure S1) and any potential protein product may behave differently than ok803. Furthermore our predictions show that the ok781 allele may also produce an amino acid sequence with no known homology thus limiting its biological relevance. We believe the lifespan and stress phenotypes observed in the tdp-1(ok803) mutants are truly linked to this gene based on several lines of experimental evidence.
Concerning the lifespan phenotypes, tdp-1(RNAi) reduces the long-lived phenotype of daf-2(e1370) worms in agreement with the tdp-1(ok803);daf-2(e1370) strain mutant. We also tested if introducing wild type tdp-1 DNA sequence could rescue the long-lived phenotype of tdp-1(ok803) mutants. Using a strain expressing a full length TDP-1 open reading frame driven by the endogenous tdp-1 promoter we observed that it could partially rescue the extended lifespan phenotype of tdp-1(ok803) mutants. This is a direct proof that the lifespan phenotypes we observe in tdp-1(ok803) mutants is due to loss of the sequence and that we can correct this phenotype by introduction of wild type tdp-1 sequence.
We corroborated the role of tdp-1 in the cellular stress response with several experimental approaches. First, tdp-1(ok803) mutants are sensitive to oxidative and osmotic stress and we observed that our TDP-1::GFP reporter is induced by these same stresses. Second, western blotting with a TDP-1 antibody showed that endogenous TDP-1 protein levels are also induced by oxidative and osmotic stress. Third, our genetic experiments with tdp-1(ok803) showed that these mutants are sensitive to oxidative stress via the IIS pathway, but sensitivity to osmotic stress is independent of the IIS pathway. Our experiments with the TDP-1::GFP reporter or immunoblotting with the TDP-1 antibody fully support the observations made with the tdp-1(ok803) mutant. Finally, tdp-1(ok803) suppresses proteotoxicity while TDP-1::GFP enhances toxicity. Again, this is akin to a classic genetic rescue experiment and provides direct evidence that the phenotype observed by deletion of tdp-1 can be modified by the introduction of wild type tdp-1 sequence. In total we believe our hypothesis that tdp-1 has roles in lifespan and stress is well supported by multiple approaches. It is not clear why we see very little effect for the ok781 allele in stress assays, but at the same time there is not yet general consensus for any of the phenotypes observed for the tdp-1 mutants.
An early location for mutant TDP-43 and FUS toxicity may be within the ER. The ER has critical cellular functions including protein folding, and misfolded proteins within the ER cause stress leading to activation of the UPRER to restore homeostasis [30]. Cellular insults can lead to increased protein misfolding as can the expression of genetically encoded proteins like mutant TDP-43 or FUS [16]. Early phases of the UPRER are protective but ER stress also stimulates the clearance of misfolded proteins from the ER through ER-associated degradation (ERAD) by transporting misfolded proteins from the ER lumen to the cytoplasm for degradation by the ubiquitin-proteasome. ERAD is energetically costly, redox intense and leads to substantial production of oxidative stress and if the ER stress is not resolved it can lead to cell death [30]. We hypothesize that protein misfolding is an early step in neurodegeneration that may result in at least three overlapping mechanisms of toxicity: primary toxicity from misfolded proteins, secondary toxicity from increased oxidative stress and tertiary toxicity propagated by stress induced TDP-1 expression (Figure 10). This feed forward mechanism originating in the ER may drive cytotoxicity and neurodegeneration. If this mechanism is conserved, these data may help explain why the intracellular accumulation of wild type TDP-43 is observed in a growing number of neurodegenerative disorders. Furthermore, given TDP-43's propensity to aggregate and its inherent cytotoxicity, wild type TDP-43 may actively contribute to neurodegeneration. Indeed recent hypotheses suggest that mutant proteins may act as seeds for the accumulation of wild type TDP-43 into pathogenic conformations as described for prion toxicity [31].
Our model complements the “two-hits” hypothesis for sporadic diseases that highlights the role of environmental stresses in combination with cytoplasmic accumulations of TDP-43 as part of the trigger for pathogenesis [32]. Proteostasis is essential to survival and healthy aging but gradually becomes less efficient as organisms age and may contribute to the accumulation of TDP-43 [33]–[35]. Add to this the stress-induced expression of TDP-43 and the cell is faced with increased cytoplasmic aggregation leading to pathogenesis.
Many TDP-43 models have been described and they all show toxicity from the over expression of wild type TDP-43, which is sometimes less toxic than mutant but not always [11], [12], [21], [24], [36]–[41]. These findings again demonstrate that control of TDP-43 levels is important for cell survival and that wild type TDP-43 may contribute to neuronal toxicity. We genetically tested this premise in C. elegans by creating transgenic strains that either overexpressed stress-activated TDP-1 or were missing the worm's endogenous tdp-1. Paralysis and lifespan phenotypes of TDP-43 and FUS transgenics were worsened by increased expression of TDP-1. Consistently, deletion of endogenous tdp-1 from strains expressing mutant TDP-43 or FUS remarkably reduced paralysis and motor neuron degeneration phenotypes. Taken together we directly showed that wild type tdp-1 plays an important role in neurodegeneration caused by mutant TDP-43 and FUS proteins.
However two studies in C. elegans have shown no effect on TDP-43 toxicity in animals mutant for endogenous tdp-1 [24], [37] while another study has shown that deleting tdp-1 reduces TDP-43 and SOD1 toxicity [42]. The reason for the differences are not clear but may be due to differences between models, where in our model animals express mutant TDP-43 in only 26 GABAergic motor neurons [16] while the other models described rely on the expression of TDP-43 transgenes throughout the worms entire nervous system [24], [37], [42]. In our model we observe adult-onset, progressive motility defects and neurodegeneration [16], while the other models describe uncoordinated motility problems from earlier stages. It may be that the pan neuronal expression of TDP-43 in some models causes phenotypes that are too severe for modulation by reducing endogenous tdp-1. Additionally, the fact that we see a reduction of mutant FUS toxicity by deleting tdp-1 bolsters the notion that proteotoxic induction of TDP-1 propagates toxicity and this may not be a phenomenon restricted to mutant TDP-43.
A surprising finding is that the UPRER appears to be activated in a cell non-autonomous manner. The hsp-4p::GFP reporter is expressed primarily in the worms intestinal cells while mutant TDP-43 and FUS are expressed in motor neurons. Thus ER stress generated within neurons is capable of signaling to other cells and tissue types perhaps as part of a coordinated organism wide response. Cell non-autonomous signaling has been described in C. elegans for mitochondrial stress regulating longevity [43] and the heat shock response [44]. Whether mutant TDP-43 and FUS similarly induce a system wide ER stress response in mammals awaits investigation.
Concerning the regulation of lifespan, our work describes a complex relationship between tdp-1 and daf-2. tdp-1(ok803) mutants have extended lifespan but are stress sensitive, while daf-2(e1370);tdp-1(ok803) mutants are stress sensitive and have decreased lifespan compared to daf-2(e1370) alone. In all systems examined increased TDP-43 expression is toxic [45] and there is widespread speculation in the field that wild type TDP-43 may contribute to cytotoxicity and neurodegeneration over long term settings [31]. Thus deleting tdp-1 from worms leaves them sensitive to stress but frees them from potential long-term cytotoxic effects. We are not alone in the observation that removing tdp-1 increases lifespan [42] but we are the first to look at tdp-1's role in the cellular stress response. Looking at daf-2, the difference here may be that tdp-1 is essential for the stress resistance aspects of daf-2 mutants and concomitant long-lived phenotype. Thus removing tdp-1 renders daf-2 animals sensitive to stress induced damage limiting their extended lifespan.
A surprising finding was the opposing effects of daf-2 mutations on proteotoxicity and TDP-1 expression. Several publications have reported that reduced daf-2 function suppresses proteotoxicity [19]–[21]. Our experiments are not fully comparable to these studies since some studies rely solely on daf-2(RNAi) or on a single daf-2 allele (e1370) to investigate proteotoxicity while we looked at two different daf-2 alleles, e1368 and e1370. Further complicating the matter is the fact that some models are based on muscle-cell expression vectors and/or have severe developmental effects like the recently described TDP-43 model [21].
daf-2 mutations are grouped into a complex allelic series comprising two classes. Class 1 alleles are less severe and mutations fall within the extracellular regions of the DAF-2 receptor protein [46]. Class 2 alleles tend to be more severe, display pleiotropic effects and the mutations are found within the ligand binding area of the receptor or the tyrosine kinase domain [46]. Using the class 1 allele daf-2(e1368) and the class 2 allele daf-2(e1370) we observed that the two alleles had opposing effects on mutant TDP-43 toxicity and TDP-1 expression. In our mutant TDP-43 worms we observed that daf-2(e1368) suppressed mutant TDP-43 induced paralysis, while the class 2 allele daf-2(e1370) enhanced paralysis. Furthermore, daf-2(e1368) reduced the insolubility of mutant TDP-43 protein, while daf-2(e1370) had no effect on mutant TDP-43 solubility. Using a TDP-1::GFP reporter or western blotting with a TDP-1 antibody we observed that daf-2(e1370) increased the expression of TDP-1 while daf-2(e1368) had no effect. Given that the IIS pathway regulates both the toxicity of mutant TDP-43 proteins and the expression of endogenous TDP-1/TDP-43, altered IIS may directly contribute to ALS pathogenesis.
The pleiotropic and sometimes opposite effects of daf-2 are known and they are consistent within each class [46]. Regarding stress, a notable example is that class 1 daf-2(e1368) mutants are sensitive to hypoxia while class 2 daf-2(e1370) mutants are highly resistant to hypoxia [47]. Here we observe opposing effects for daf-2 on proteotoxicity and the cytotoxic induction of TDP-1 expression both of which are consistent for each allele. Our data are in agreement with studies showing that reduced IIS diminishes proteotoxicity but the question remains as to why the more severe class 2 daf-2(e1370) allele enhances toxicity in our TDP-43 transgenics. Part of the reason may lie with intrinsic differences between models. Our transgenics are based on the expression of TDP-43 in 26 GABAergic neurons and the worms do not show motor impairment until well into adulthood [16] while the other models show early effects and the animals are severely impaired by the time they reach adulthood [19], [21]. Thus it may be that stronger daf-2 mutations are required to suppress severe phenotypes but are too strong for the milder, late-onset toxicity in our TDP-43 transgenics. There may be an optimal rate of IIS unique to each situation that has been referred to the Insulin Signalling Paradox [10]. Our data begin to shed light on this phenomenon where changes in signaling can have dramatically different effects and suggests that the role of IIS in neurodegeneration is more complicated that currently appreciated and requires further investigation.
It is still unclear if mutant TDP-43 results in a gain, a loss of function or both but work from zebrafish and fly TDP-43 models suggest that it may be both [11], [41]. Our data introduce a novel gain of function mechanism where the increased expression of wild type TDP-1 is induced by proteotoxic stress. Several strategies come to mind to alleviate the neuronal toxicity caused by wild type and mutant TDP-43 including reducing levels of wild type TDP-43, mutant TDP-43, and/or reducing UPRER stress by promoting protein folding. In the future it will be important to determine if strategies to reduce TDP-43 neuronal toxicity may be applicable to additional neurodegenerative disease as a shared mechanism of cell death in the development of new therapeutics.
Nematode strains used were described previously [16] or received from the Caenorhabditis elegans Genetics Center CGC (St Paul, MN). All strains were maintained following standard methods on OP50 bacteria plates. Strains used in this study include: tdp-1(ok803) and tdp-1(ok781) both outcrossed 5 times to N2 prior to use, daf-2(e1368), daf- 2(e1370), daf-16(mu86), gpIs1[hsp-16.2::GFP], oxIs12[unc-47p::GFP;lin-15(+)], xqIs93[tdp-1p::TDP-1::GFP], xqIs132[unc-47p::TDP-43-WT;unc-119(+)], xqIs133[unc-47p::TDP-43[A315T];unc-119(+)], xqIs173[unc-47p::FUS-WT;unc-119(+)], xqIs98[unc-47p::FUS[S57Δ];unc-119(+)], zIs356[daf-16p::daf-16-gfp; rol-6], zcIs4[hsp-4p::GFP], zcIs9[hsp-60p::GFP],gpIs1[hsp-16.2p::GFP] and zcIs13[hsp-6p::GFP].
Sixty synchronized L4 were grown on OP50 bacteria plates (20 animals/plate) and three independent assays were performed. Lifespan analyses were performed at 20°C and 25°C and worms were scored every 1–2 days from adult day 1 until death. Worms were scored dead if they didn't respond to tactile or heat stimulus.
RNAi-treated strains were fed with E. coli (HT115) containing an Empty Vector (EV), daf-16 (R13H8.1), or tdp-1 (F44G4.4) RNAi clones from the ORFeome RNAi library (Open Biosystems). RNAi experiments were performed at 20°C. Worms were grown on NGM enriched with 1 mM Isopropyl-b-D-thiogalactopyranoside (IPTG). For lifespan, worms were transferred to RNAi 5-fluorodexyuridine (FUDR, 12.5 mg/L) plates at adult day 1 until death. Worms were declared dead if they didn't respond to tactile or heat stimulus. Experiments were conducted with 20 animals/plate by triplicates.
Young adult daf-2(e1370) were allowed to lay eggs overnight at 20°C. The eggs were then transferred to 25°C and scored for dauer formation 5 days later. Three different trials on different days were performed.
Stress tests were performed at 20°C (oxidative, osmotic and UV stress), 25°C (hypoxia) and 37°C (thermal stress). Worms were grown on NGM and transferred to NGM plates + 240 µM juglone (oxidative stress), or NGM plates + 10 mM Hydrogen peroxide (oxidative stress), or NGM plates + 400 mM NaCl (osmotic stress), or NGM plates + 611 mM Sorbitol (osmotic stress), all at adult day 1. For the oxidative, osmotic and thermal stress assays, worms were evaluated for survival every 30 minutes for the first 2 hours and every 2 hours after up to 14 hours; for sorbitol we also performed a test over 48 hours, starting the counts after 14 hours on the compound. For UV irradiation, adult day 1 worms were transferred to NGM plates without any food source and exposed to UV (1200 J/m2). Worms were then transferred to NGM plates with OP50 bacteria and died animals were counted every 2 hours till 14 hours after irradiation. For hypoxia experiments young adult animals were transferred to a new plate and subjected to low oxygen conditions with the AnaeroPack system (Mitsubishi Gas Chemical America) for 24 hours and assayed for survival. For all experiments nematodes were scored as dead if they were unable to move in response to heat or tactile stimuli. For all tests worms, 20 animals/plate by triplicates were scored.
Gateway system (Invitrogen) compatible tdp-1 promoter and open reading frame plasmid clones were obtained from Open Biosystems and recombined with plasmid pDES-MB14 (kindly donated by M. Vidal, Harvard), and verified by sequencing to create a tdp-1p::TDP-1::GFP plasmid, which was injected at 5 ng/µl into unc-119(ed3) animals along with myo-3p::mCherry, myo-2p::mCherry comarkers at 5 ng/µl and wild type transformants expressing GFP were kept. The transgene was integrated using UV radiation and wild type, GFP positive animals were kept for further study. Multiple stable transgenics were isolated and outcrossed to N2 4 times before use. Strain XQ93 xqIs93[tdp-1p::TDP-1::GFP] was used in this study.
For visualization of TDP-1::GFP animals, M9 buffer with 5 mM levamisole was used for immobilization. Animals were mounted on slides with 2% agarose pads. TDP-1::GFP expression was visualized with a Leica CTR 6000 and a Leica DFC 480 camera. L4 animals were grown on NGM plates and transferred to NGM plates + 240 µM juglone (oxidative stress) or NGM plates + 400 mM NaCl (osmotic stress) for 90 minutes, and examined for fluorescence with the Leica system described above. Some animals were also stained with DAPI (1∶1000, diluted in 1× PBS). Image processing was done with Adobe Photoshop. For images of TDP-1::GFP alone, images were converted to black and white and the images reversed to allow for better contrast and visualization. Quantification of TDP-1::GFP levels was done with ImageJ (NIH) and the mean and SD was calculated from 5 images for each strain and experimental condition. For visualization of DAF-16::GFP, hsp-4p::GFP, hsp-6p::GFP, hsp-16.2p::GFP and hsp-60p::GFP animals, M9 buffer with 5 mM levamisole was used for immobilization. Animals were mounted on slides with 2% agarose pads and examined for fluorescence with the Leica system described above.
For visualization of oxidative damage in the transgenic strains the worms were incubated on a slide for 30 minutes with 5 mM dihydrofluorescein diacetate dye (Sigma-Aldrich) and then washed with 1× PBS three times. After the slide was fixed fluorescence was observed with the Leica system described above.
RNA was extracted with an RNAeasy kit (Qiagen) and reverse transcribed with QuantiTect (Qiagen). Primers used include: ctl-1 forward, AGGTCACCCATGACATCACCAAGT; ctl-1 reverse, GAT TGCGCTTCAGGGCATGAATGA; ctl-2 forward, TTCGCTGAGTTGAACAATCCG; ctl-2 reverse, GTTGCTGATTGTCATAAGCCATTGC; tdp-1 forward, AAAGTGGGATCGAGTGACGAC; tdp-1 reverse, GACAGCGTAACGAATGCAAAGC; sod-3 forward, CGAGCTCGAACCTGTAATCAGCCATG; sod-3 reverse, GGGGTACCGCTGATATTCTTCCACTTG; act-3 forward, GTTGCCGCTCTTGTTGTAGAC; act-3 reverse, GGAGAGGACAGCTTGGATGG.
Worms were collected in M9 buffer, washed 3 times with M9 and pellets were placed at −80°C overnight. Pellets were lysed in RIPA buffer (150 mM NaCl, 50 mM Tris pH 7.4, 1% Triton X-100, 0.1% SDS, 1% sodium deoxycholate) + 0.1% protease inhibitors (10 mg/ml leupeptin, 10 mg/ml pepstatin A, 10 mg/ml chymostatin LPC;1/1000). Pellets were passed through a 271/2 G syringe 10 times, sonicated and centrifuged at 16000 g. Supernatants were collected.
For TDP-43 and FUS transgenics, soluble/insoluble fractions, worms were lysed in Extraction Buffer (1 M Tris-HCl pH 8, 0.5 M EDTA, 1 M NaCl, 10% NP40 + protease inhibitors (LPC;1/1000). Pellets were passed through a 271/2 G syringe 10 times, sonicated and centrifuged at 100000 g for 5 minutes. The soluble supernatant was saved and the remaining pellet was resuspended in extraction buffer, sonicated and centrifuged at 100000 g for 5 minutes. The remaining pellet was resuspended into 100 µl of RIPA buffer, sonicated until the pellet was resuspended in solution and saved.
All supernatants were quantified with the BCA Protein Assay Kit (Thermo Scientific) following the manufacturer instructions.
Worm RIPA samples (175 µg/well for transgenic worms; 15 µg/well for non transgenics) were resuspended directly in 1× Laemmli sample buffer, migrated in 10% polyacrylamide gels, transferred to nitrocellulose membranes (BioRad) and immunoblotted. Antibodies used: rabbit anti-TDP-43 (1∶200, Proteintech), rabbit anti-FUS/TLS (1∶200, Abcam), rabbit anti-TDP-1 (1∶500, Petrucelli laboratory) and mouse anti-Actin (1∶10000, MP Biomedical). Blots were visualized with peroxidase-conjugated secondary antibodies and ECL Western Blotting Substrate (Thermo Scientific). Densitometry was performed with Photoshop (Adobe).
For lifespan and stress-resistance tests, survival curves were generated and compared using the Log-rank (Mantel-Cox) test, and 20–30 animals were tested per genotype and repeated at least three times. For progeny counts, dauer-formation assays and hypoxia tests the mean and SEM were calculated for each trial and two-tailed t-tests were used for statistical analysis.
|
10.1371/journal.pgen.1002449 | An siRNA Screen in Pancreatic Beta Cells Reveals a Role for Gpr27 in Insulin Production | The prevalence of type 2 diabetes in the United States is projected to double or triple by 2050. We reasoned that the genes that modulate insulin production might be new targets for diabetes therapeutics. Therefore, we developed an siRNA screening system to identify genes important for the activity of the insulin promoter in beta cells. We created a subclone of the MIN6 mouse pancreatic beta cell line that expresses destabilized GFP under the control of a 362 base pair fragment of the human insulin promoter and the mCherry red fluorescent protein under the control of the constitutively active rous sarcoma virus promoter. The ratio of the GFP to mCherry fluorescence of a cell indicates its insulin promoter activity. As G protein coupled receptors (GPCRs) have emerged as novel targets for diabetes therapies, we used this cell line to screen an siRNA library targeting all known mouse GPCRs. We identified several known GPCR regulators of insulin secretion as regulators of the insulin promoter. One of the top positive regulators was Gpr27, an orphan GPCR with no known role in beta cell function. We show that knockdown of Gpr27 reduces endogenous mouse insulin promoter activity and glucose stimulated insulin secretion. Furthermore, we show that Pdx1 is important for Gpr27's effect on the insulin promoter and insulin secretion. Finally, the over-expression of Gpr27 in 293T cells increases inositol phosphate levels, while knockdown of Gpr27 in MIN6 cells reduces inositol phosphate levels, suggesting this orphan GPCR might couple to Gq/11. In summary, we demonstrate a MIN6-based siRNA screening system that allows rapid identification of novel positive and negative regulators of the insulin promoter. Using this system, we identify Gpr27 as a positive regulator of insulin production.
| Pancreatic beta cells are the only physiologic source of insulin. When these cells are destroyed in type 1 diabetics, there is uncontrolled hyperglycemia from complete insulin deficiency. In type 2 diabetes, these same cells fail to increase insulin secretion to compensate for peripheral insulin resistance leading to relative insulin deficiency. We constructed a novel screening system to find new regulators of insulin production in this critical cell type. Here, we describe a screen of the G protein coupled receptors (GPCRs) and show a role for orphan GPCR, Gpr27, in insulin promoter activity and insulin secretion. We propose that Gpr27 is a novel target for diabetes therapeutics.
| Nearly 13% of American adults have diabetes and these numbers continue to rise, mostly from an increase in type 2 diabetes [1], [2]. Although insulin resistance is a cardinal feature of type 2 diabetes, most people with insulin resistance do not develop diabetes because their pancreatic beta cells are able to compensate by increasing insulin production. However, if insulin production cannot match the increased demand imposed by insulin resistance, hyperglycemia and frank diabetes ensues. Over time, beta cell function further declines in most people with type 2 diabetes, resulting in the eventual failure of oral medications and the necessity of insulin therapy [3].
Improving insulin production and beta cell function is therefore a universal goal of diabetes therapeutics. We reasoned that an unbiased search for regulators of insulin production might reveal new diabetes drug targets. Therefore, we constructed a novel screening system to screen for genes important for insulin promoter activity. By screening siRNAs targeting all GPCRs, we identify several GPCRs that regulate insulin promoter activity and specifically characterize Gpr27 as a novel regulator of insulin production.
To allow rapid evaluation of insulin promoter activity, the MIN6 mouse beta cell line was infected with a lentivirus that stably expresses destabilized GFP under the control of the proximal 362 base pairs of the human insulin promoter (Figure 1A) [4]. This insulin promoter fragment maintains a substantial proportion of promoter activity and tissue specificity while being compact enough to allow lentiviral delivery [5].
To favor single copy integration, the construct was delivered at a low multiplicity of infection (MOI) and a clonal line was selected. To generate an internal control reporter, the GFP positive subline was subsequently infected at a low MOI with a second lentivirus containing mCherry under the control of the constitutive rous sarcoma virus promoter (RSV) (Figure 1A). A stable clone expressing both constructs was isolated. In these cells, the ratio of GFP to mCherry fluorescence indicates human insulin promoter activity.
When transfected into this reporter line, siRNAs targeting activators of insulin gene transcription would be expected to reduce insulin promoter activity and reduce the GFP/mCherry ratio, while siRNAs targeting negative regulators of the insulin promoter should increase the GFP/mCherry ratio (Figure 1B). Indeed, transfection of an siRNA targeting the insulin gene transcription factor Pdx1 reduced the GFP/mCherry ratio by 80% as compared to a non-targeting siRNAs (Figure 1C and 1D) [6].
An RNAi library containing four independent siRNAs targeting the mouse GPCR-ome and selected GPCR related genes was transfected into the reporter cell line. The ratio of the GFP to mCherry fluorescence five days after transfection was calculated for each siRNA and the data were then analyzed using the redundant siRNA analysis (RSA) software [7]. To avoid off-target effects, each siRNA was transfected separately and only genes with more than one siRNA hit were selected for further analysis.
The top genes judged by RSA were then ranked by unsupervised clustering of each gene's RSA p value and its expression level in primary mouse islets, since only those genes expressed in primary cells are of biological interest (Figure 2A). Two publically available mouse islet mRNA-Seq data sets were used. One of these data sets has been previously published and consists of approximately four million mapped reads from islets isolated from female non-pregnant mice and approximately four million mapped reads from islets isolated from pregnant mice [8]. The second, submitted to the NCBI Short Read Archive by Merck, contains approximately 120 million reads from mouse islets (see methods). Because of these modest read numbers, some low abundance transcripts may be erroneously reported as being not expressed using this analysis [9].
siRNAs to the top six genes (Ffar2, Gpr27, Grk5, p2ry6, Gpr109a, Bdkrb2) that reduced insulin promoter activity and had detectable expression in primary mouse islets were transfected into the screening cell line for confirmation (Figure 2B). All six genes had at least 2 siRNAs confirm. For the siRNAs that increased GFP/mCherry, we retested the top three genes with high RSA scores and detectable expression in mouse primary islets – Adra2a, Cckar, and Aplnr. Of these three, only the known negative regulator of insulin secretion, Adra2a, confirmed with two independent siRNAs (Figure 2C).
Several of the positive regulators of the insulin promoter we identified were already known to stimulate insulin release in beta cells. Of particular interest was the orphan GPCR, Gpr27, which had no known role in insulin production but was previously found to be enriched in the mouse and human pancreatic islet [10], [11]. We subsequently tested all four siRNAs targeting Gpr27 in the library set on an independently generated MIN6 reporter line expressing stable GFP under the control of the insulin promoter and mCherry under the control of the RSV promoter. All four siRNAs reduced GFP/mCherry fluorescence (Figure S1A). Furthermore, all 4 siRNAs efficiently reduced expression of the Gpr27 mRNA (Figure S1B). We also confirmed that Gpr27 is enriched in beta cell lines (beta TC and MIN6) compared to an alpha cell line (alpha TC) (Figure S2A) and is expressed in primary mouse beta cells (Figure S2B).
Since the screen was based on a human insulin promoter fragment, we measured the effect of Gpr27 knockdown on the endogenous mouse Ins2 and Ins1 genes. Because mature insulin mRNAs have a half-life of nearly 80 hours, we measured insulin pre-mRNAs as previously described [6], [12]. MIN6 cells infected with a Gpr27 shRNA expressing adenovirus (Ad-shGpr27) had a 40–60% reduction in pre-ins2 and pre-ins1 levels compared to control adenovirus (Ad-control) (Figure 3A). To confirm these findings in primary beta cells, we infected intact primary mouse islets with these same adenoviruses. At a high MOI, we were only able to obtain 50% infection rates as measured by flow cytometry, presumably reflecting poor adenovirus penetration into the core of the mouse islet [13]. Therefore, intact islets were dissociated prior to adenovirus infection. Three days after infection, cells were isolated by flow cytometric sorting for GFP and RT-qPCR was performed. Knockdown of Gpr27 produced a significant ∼30% reduction of pre-ins2 (p = 0.03). Concomitantly, there was a nearly significant 30% reduction in the less abundant pre-ins1 message (p = 0.055) (Figure 3B).
While insulin production requires insulin promoter activity, minute-to-minute changes in plasma insulin levels are controlled by insulin secretion. Therefore, we asked if Gpr27 knockdown would affect glucose stimulated insulin secretion. Infection of MIN6 cells with Ad-control at an MOI necessary to get >90% infection inhibited glucose stimulated insulin secretion (data not shown). Therefore, we infected MIN6 cells at a lower MOI to achieve approximately 60% infection and measured glucose stimulated insulin secretion from this mixed population by batch incubation. Ad-shGpr27 infected MIN6 cells secreted ∼40% less insulin at 20 mM glucose compared to Ad-control infected cells (Figure 3C). There was no statistically significant difference at 2 mM glucose. Notably, we did not detect a difference in total insulin as normalized to total protein concentration (Ad-control = 27.9+/−1.1 mg insulin per g of total protein; Ad-shGpr27 = 29.4+/−0.94 mg insulin per g of total protein, p value = 0.13). This was not unexpected since the half-life of insulin mRNA is ∼80 hours and the knockdown of Gpr27 was limited to 72 hours due to adenovirus toxicity after that time point. We conclude that Gpr27 plays a measurable role in insulin secretion in addition to insulin promoter activity.
To define the mechanism of Gpr27 action, we measured transcript levels of selected regulators of the insulin promoter by RT-QPCR in MIN6 cells after Ad-shGpr27 infection. Glis3, Pax6, Nkx6.1, HNF4a, and Pdx1 were reduced after Gpr27 knockdown while others including MafA, NeuroD1, and Pax4 were unchanged (Figure 4A). Concordant with this expression data, Gpr27 knockdown reduced the transcriptional activity of mini-enhancers that bind to Glis3 and Pdx1 (Z, E1/A1, E2/A3) while Gpr27 knockdown had no effect on mini-enhancers that bind to MafA and NeuroD1 (C1/E1) (Figure 4B and 4C).
Since Pdx1 is required for insulin promoter activity and insulin secretion [14], [15], we asked if Pdx1 is required for the effect of Gpr27's on the insulin promoter. By luciferase assay, we found that the single knockdown of Pdx1 reduced insulin promoter activity by 90% and Gpr27 knockdown alone reduced insulin promoter activity by 40%. However, the knockdown of both Gpr27 and Pdx1 had no additional effect over the single knockdown of Pdx1, showing that Pdx1 is important for the effect of Gpr27 on the insulin promoter (Figure 4D). Importantly, double knockdown of both Gpr27 and Pdx1 was as efficient as single knockdown (Figure S3).
We then asked if Pdx1 was required for the effect of Gpr27 on insulin secretion. The knockdown of Pdx1 reduced fractional insulin secretion at 20 mM glucose and total insulin content (Figure 4E and Figure S4). As with the adenoviral knockdown of Gpr27, an siRNA to Gpr27 reduced glucose stimulated insulin secretion. However, the knockdown of Gpr27 in addition to Pdx1 did not further reduce insulin secretion at 20 mM glucose. We conclude that Gpr27 plays a measurable role in insulin secretion and insulin promoter activity via a mechanism involving Pdx1.
G protein coupling software analysis predicts that Gpr27 could function via Gi or Gq/11 signaling pathways [16]. Since Gpr27 is already expressed in MIN6 cells, we ectopically expressed mouse Gpr27 in HEK293T cells. Robust expression of FLAG-tagged Gpr27 was detected by 24 hours on the surface of the majority of cells (Figure 5B). We then measured cAMP and IP1 – higher cAMP would indicate Gs coupling, lower cAMP would indicate Gi coupling and higher IP1 would indicate Gq/11 coupling (Figure 5A). Gpr27 expression resulted in a 2-fold elevation of IP1 levels while leaving cAMP levels unchanged (Figure 5C and 5D) showing that in this heterologous cell type, Gpr27 may activate the Gq/11 pathway.
If Gpr27 activates Gq/11 in beta cells, then IP1 levels should be reduced in MIN6 cells after knockdown of Gpr27. Therefore, we measured IP1 levels and cAMP levels in MIN6 cells after Gpr27 knockdown. Indeed, knockdown of Gpr27 resulted in reduced IP1 levels while cAMP levels were not significantly changed (Figure 5E and 5F). Taken together, these data show that Gpr27 positively regulates inositol phosphate levels, supporting a role for Gpr27 in activating the Gq/11 pathway.
To identify new regulators of the insulin promoter, we developed a novel siRNA screening system in MIN6 cells that allows rapid measurement of insulin promoter activity. As an initial test of the system, an siRNA screen of the GPCR-ome was performed. The RSA algorithm was used to select hits in order to capitalize on the four fold redundancy of the siRNA library [7]. To further increase the specificity of the screen, at least 2 siRNAs must have been identified for a gene to be a hit. The top RSA hits were then prioritized by expression level in mouse primary islets. Besides filtering out genes expressed in MIN6 but not in primary islets, this step also eliminates off-target hits. On the other hand, hit genes with low expression may have been erroneously eliminated because they were below the limit of detection of the mRNA-seq data available at this time [17]. Nonetheless, this filtering step allowed us to focus on genes with reasonable expression in primary cells.
While the confirmation rate for siRNAs to positive regulators was 100%, the confirmation rate for negative regulators was only 33%. This is likely due, in part, to the more modest effect of these siRNAs (∼20–30% increase in GFP/mCherry ratio) as compared to the reconfirmed Adra2a(∼50%), a known negative regulator of insulin secretion.
We identified several other known regulators of insulin secretion as regulators of the insulin promoter. The bradykinin receptor 2 mediates increases in insulin secretion in beta cells [18], [19]. Pyrimidinergic receptor 6 (p2yr6) agonists augment insulin release and this receptor participates in an autocrine feedback loop that potentiates insulin secretion [20], [21]. The free fatty acid receptor 2, which has been hypothesized to play a role in beta cells, was also identified as a positive regulator of the insulin promoter in our screen [22]. Several other receptors were identified in the screen that have no known role in beta cells and these may merit further investigation. Of note, Glp1r was not identified in this screen for a trivial reason; siRNAs targeting this gene were not included in the commercial screening set. Given the nature of our screen, hits would be predicted to either have basal activity or have ligand present in the culture conditions as has been described for p2yr6 [21].
We were most intrigued by the orphan GPCR, Gpr27 [23]. Previous studies have shown that it is enriched in the pancreatic islets of both human and mouse [10], [11]. Detailed mouse tissue profiling of Gpr27 expression by RT-QPCR shows high expression in the mouse brain with lower expression in the islet and heart [24]. Furthermore, Gpr27 mRNA is up-regulated in Neurogenin3 positive endocrine precursors in the developing mouse pancreas [11]. Conversely, Gpr27 is 8-fold down regulated in the Neurogenin3 knockout pancreas [25], [26]. Taken together, these data suggest Gpr27 is an endocrine pancreas specific gene.
We confirmed that knockdown of Gpr27 reduces the activity of human insulin promoter reporters, levels of endogenous mouse Ins2 pre-mRNA, and glucose stimulated insulin secretion. Importantly, Gpr27 knockdown also reduces the levels of endogenous Ins2 pre-mRNA in dissociated primary mouse islets. We also found that the mRNAs for multiple transcription factors that activate the insulin promoter (Glis3, Pdx1, HNF4a) were reduced by Gpr27 knockdown. Other transcription factors critical for beta cell development were also reduced including Nkx6.1 and Pax6. In agreement with the reduction in their expression, only Pdx1 and Glis3 binding mini-enhancers were affected by Gpr27 knockdown (Figure 4C). Finally, there was no further reduction in insulin promoter activity when adding Gpr27 knockdown to Pdx1 knockdown. A limitation of this double knockdown experiment is that given the very strong effect of Pdx1 knockdown alone on insulin promoter activity, a further reduction with Gpr27/Pdx1 double knockdown may be either below our limit of detection or simply reflect no remaining insulin promoter activity.
How might Gpr27 affect both insulin transcription and glucose stimulated insulin secretion? The Gq/11 pathway was an obvious candidate as the expression of Gpr27 in HEK 293T cells increased IP1 levels while the knockdown of Gpr27 reduced IP1 levels in MIN6 cells. Furthermore, triggering of an engineered Gq/11-coupled GPCR in beta cells increases steady state insulin mRNA levels and insulin secretion [27]. However, Pdx1 levels did not change after triggering this Gq/11-coupled GPCR [27] and Gq/11 knockout beta cells have normal levels of Ins1 and beta cell transcription factor mRNAs [21]. Therefore, even if Gpr27 directly couples to Gq/11, Gpr27 may affect insulin secretion and insulin promoter activity independent of Gq/11 as has recently been demonstrated for the M3 receptor [28].
Another candidate for mediating the effects of Gpr27 on insulin promoter and insulin secretion was Pdx1 since it is known to positively regulate both insulin transcription and insulin secretion [14], [15]. We found that the double siRNA knockdown of Gpr27 and Pdx1 produced no further reduction in insulin secretion over Pdx1 knockdown alone, suggesting that Pdx1 is important for Gpr27's effect on insulin secretion. In combination with the reduction in Pdx1 mRNA by Gpr27 knockdown, these data suggest Gpr27 functions upstream of Pdx1. However, the double knockdown data do not exclude the possibility that a Pdx1 lies in a parallel pathway to Gpr27 and these two pathways intersect upstream of insulin secretion.
Taken together, these data suggest that a linear pathway connecting Gpr27 to a single G protein and a single regulatory element in the insulin promoter is overly simplistic. Indeed, a single GPCR can trigger multiple G proteins (reviewed by [29]), can trigger a combination of G protein dependent and independent pathways [30], and can function as heterodimers [31]. Likewise, the insulin promoter contains multiple elements that are both redundant and cooperative [32]. The complexity of these systems highlights the advantage of using a broad, unbiased approach to finding new and unexpected regulators of the insulin promoter. Here, we used such a system to identify a novel GPCR regulator of both insulin secretion and insulin promoter activity – Gpr27. Based on its islet expression and its positive effects on the insulin promoter and insulin secretion, we suggest that Gpr27 may be a novel target for diabetes therapies.
MIN6 cells were a gift from Dr. Miyazaki. Alpha TC and beta TC were a gift from Dr. Hanahan. Cells were maintained in high glucose DMEM with 10% fetal bovine serum, and 71.5 mM beta-mercaptoethanol. Sublines were isolated by limiting dilution. Original passage lines were used between passage 25–40. Sublines were used at passages 5–10.
Human insulin promoter deletions have been previously described [5]. Promoters were subcloned from pFoxCAT into pFoxLuc [33]. For the lentiviral reporter, the human −362 promoter region was cloned upstream of destabilized GFP or GFP and this cassette was used to replace the U6/CMV-EGFP in pSicoR. pSicoR-RSV-mCherry was created by replacing the U6/CMV of pSicoR mCherry with the RSV promoter [5]. Mini-enhancer reporter constructs have also been previously described [5], [34]. They were subcloned upstream of a minimal thymidine kinase promoter-firefly luciferase reporter.
Approximately 5,000 MIN6 cells were transfected in 96 well plates using HiPerfect (Qiagen) with a final siRNA concentration of 25 nM. Cells were analyzed by flow cytometry (LSRII, BD) 5 days after transfection and the geometric mean fluorescence intensity of GFP was normalized to that of mCherry. If the knockdown of GFP by an anti-GFP siRNA was not >80%, the transfection of that plate was considered to be a technical failure and the plate was discarded. This occurred on 1 out of 20 plates and for this reason some genes were only targeted by 3 siRNAs (including Gpr27). Each well was normalized to the negative control siRNA on that 96 well plate. For the confirmation assay for Gpr27 siRNAs, a distinct MIN6 human insulin promoter-GFP/RSV-mCherry reporter line was transfected with the indicated siRNAs with Lipofectamine RNAiMax for 5 days and GFP and mCherry were measured.
Mouse islet mRNA-seq data was downloaded from the NCBI Short Read Archive (SRP000752 and SRP002569) and FPKM values were calculated using the TopHat and Cufflinks software using the NCBI RefSeq as the reference. Log FPKM and negative log RSA p values were clustered using Cluster 3.0 and heat maps were plotted with JavaTreeView.
siRNAs were obtained from Qiagen. All custom Taqman probes had a confirmed PCR efficiency of between 95–110%. Samples without reverse transcriptase did not amplify. See Text S1 for sequences of custom probes. Taqman probes to mouse Glis3, MafA, Pdx1, NeuroD1, Pax4, Nkx6.1, HNF4a were obtained from Applied Biosystems. Negative control siRNAs for the reconfirmation assay were All-Stars Negative Control (1027280), Negative Control (1022076), Unspecific-Luciferase-1 (1022070), Unspecific-Luciferase-2 (1022073), Hs_LMNA_11 (1022050), Mm_Lmna_5 (SI02655450), Hs_GAPD_5 (SI0253266), Hs_ACTB_1 (1022168).
Total RNA was isolated by Trizol (Invitrogen). The RNA was DNase I treated (Turbo DNase, Ambion) and reverse transcription was performed (Superscript III, Invitrogen) using a combination of random hexamers and oligo dT primers. For cell line experiments, each qPCR reaction used between 10–30 ng of total RNA equivalent. To convert to arbitrary linear units, the following formula was used: (2∧15)*(2∧(deltaCT to beta-glucuronidase).
Islets from 12–30 week old MIP-GFP mice were isolated by the UCSF Islet Production Core. Islets were digested with trypsin until single cell suspensions were obtained. Cells were sorted by flow cytometry (Aria II, BD or MoFlo, DakoCytomation) into GFP positive and negative fractions and total RNA was isolated. 20 ng of total RNA equivalent was loaded per QRT-PCR reaction.
140,000 MIN6 cells were transiently transfected in 24 well plates with the relevant siRNA (5 pmoles), the indicated insulin reporter firefly luciferase plasmid (100 ng), and pRL-TK(Promega) (25 ng) using Lipofectamine 2000 (Invitrogen). For double siRNA knockdowns, 5 pmoles of each siRNA or 10 pmoles of control (anti-GFP) were used. Two days after transient transfection, firely and renilla luciferase were measured using the Dual Luciferase Assay (Promega).
Gpr27 was cloned by PCR from mouse genomic DNA downstream of a viral signal sequence and amino terminal FLAG epitope tag [35]. This cassette was used to replace the EGFP in pSicoR. HEK 293T cells were transiently transfected with either pSicoR-EGFP or pSicoR-FLAG-Gpr27 using LT1 (Mirus).
293T cells were dissociated with PBS without Ca or Mg, stained with M1 anti-FLAG antibody and a Goat anti-mouse secondary antibody coupled to Alexa-594 (Invitrogen).
For 293T, one day after transient transfection in 24 well plates, cells were placed in stimulation buffer (HTRF) for 30 minutes at 37 degrees. The stimulation buffer was then removed and the cells were lysed using the kit lysis buffer. IP1 and cAMP were then measured as directed by the protocol in 384 well plates (HTRF). IP1 and cAMP levels were normalized to live cells numbers counted from duplicate wells. Viable cells counts from Gpr27 transfection were within 20% of control plasmid transfection. For MIN6 cells, 125,000 cells were infected with Gpr27 shRNA or control adenovirus at an MOI of 200 and grown in 24 well dishes (resulting in nearly ∼95% infection). Three days after infection, the cells were placed in stimulation buffer (HTRF) for 30 minutes at 37 degrees. The stimulation buffer was then removed and the cells were lysed in 1% Triton-X100, 50 mM HEPES pH 7.0, NaF 15 mM. The lysate was pre-cleared by centrifugation at 14,000 rpm for 10 minutes. A fraction of the lysate was taken for protein quantitation(micro-BCA, Pierce), IP1 or cAMP measurement (HTRF). Data were normalized to total protein content.
The Gpr27 shRNA was cloned into a modified version of pSicoR with a BstXI site replacing the HpaI site. The mouse U6 promoter and Gpr27 shRNA were then subcloned from pSicoR and placed upstream of the CMV-GFP marker in pAdTrack [36], [37]. Adenovirus was prepared and tittered as previously described [38].
Islets were isolated by the UCSF Islet Production Core Facility from 8–12 week old C57Bl/6 male mice. After 24 hours of culture in RPMI and 10% FBS, islets were trypsinized until single cell suspensions were obtained. The dissociated islet cells were resuspended in RPMI+10% FBS and infected with adenovirus at multiplicity of infection (MOI) of 25. Three days after infection, the cells were sorted by flow cytometry (Aria II, BD) for GFP positive cells (50–75% of the live population) and RT-qPCR was performed. The knockdown of pre-ins2, pre-ins1 or Gpr27 from Gpr27 shRNA adenovirus infected cells was calculated by the delta-delta CT method compared to the control adenovirus infection.
For the adenovirus assays, approximately 500,000 MIN6 cells were infected with the indicated adenoviruses at an MOI of 100 in 6 cm dishes in complete media. Three days after infection, the infection rate was ∼60% by FACS for GFP. Cells were washed 5 times in KRBH buffer (10 mM HEPES pH 7.4, 130 mM NaCl, 5 mM KCl, 1.25 mM KH2PO4, 1.25 mM MgSO4, 2.68 mM CaCl2, 5.26 mM NaHCO3) with 2 mM glucose and rested for 2 hours at 37 degrees. Cells were then washed an additional 3 times with 2 mM glucose KRBH and incubated in 3 mL of 2 mM glucose KRBH for 1 hour at 37 degrees. This supernatant was collected and replaced with 20 mM glucose KRBH for 1 hour at 37 degrees. Cells were washed with PBS before lysis in 50 mM Tris-HCl pH 8.0, 150 mM NaCl, 1% Triton X-100 with protease inhibitors. Lysates were spun at 14,000 rpm for 10 minutes and supernatants were spun at 5000 rpm for 5 minutes before analysis by an Ultrasensitive Insulin ELISA (Mercodia). Total protein was measured by Micro-BCA (Pierce). Total insulin was normalized to total protein in the lysate. For the siRNA transfections, 20,000 MIN6 cells were transfected per well of a Corning CellBIND 96 well plate with 25 nM of each siRNA (or 50 nM of control siRNA) using Lipofectamine RNAiMax. 5 days after transfection, cells were washed in KRBH with 2 mM glucose twice, then incubated for 2 hours at 37 degrees, then washed again with KRBH 2 mM glucose twice, then incubated for one hour with KRBH 2 mM, then KRBH with 20 mM glucose for 1 hour. Lysates were prepared in 75 uL of lysis buffer as above. Due to the lower cell numbers in the 96 well plate assay, total insulin was normalized to total genomic DNA measured by Qubit High Sensitivity DNA kit (Life Technologies).
For siRNA primary confirmation assay, an independent, two sample, one tailed t-test was used. For the primary islet adenovirus knockdown of Gpr27 an independent, one sample, two tailed t-test was used. All other p values were calculated with an independent, two sample, two tailed t-test.
Animal experiments were approved by the UCSF Institutional Animal Care and Use Committee (Protocol AN082433-02) with care taken to avoid any unnecessary suffering. Animals were maintained in accordance with the applicable portions of the Animal Welfare act and the DHHS Guide for the Care and Use of Laboratory Animals.
|
10.1371/journal.pgen.1008212 | Longevity is determined by ETS transcription factors in multiple tissues and diverse species | Ageing populations pose one of the main public health crises of our time. Reprogramming gene expression by altering the activities of sequence-specific transcription factors (TFs) can ameliorate deleterious effects of age. Here we explore how a circuit of TFs coordinates pro-longevity transcriptional outcomes, which reveals a multi-tissue and multi-species role for an entire protein family: the E-twenty-six (ETS) TFs. In Drosophila, reduced insulin/IGF signalling (IIS) extends lifespan by coordinating activation of Aop, an ETS transcriptional repressor, and Foxo, a Forkhead transcriptional activator. Aop and Foxo bind the same genomic loci, and we show that, individually, they effect similar transcriptional programmes in vivo. In combination, Aop can both moderate or synergise with Foxo, dependent on promoter context. Moreover, Foxo and Aop oppose the gene-regulatory activity of Pnt, an ETS transcriptional activator. Directly knocking down Pnt recapitulates aspects of the Aop/Foxo transcriptional programme and is sufficient to extend lifespan. The lifespan-limiting role of Pnt appears to be balanced by a requirement for metabolic regulation in young flies, in which the Aop-Pnt-Foxo circuit determines expression of metabolic genes, and Pnt regulates lipolysis and responses to nutrient stress. Molecular functions are often conserved amongst ETS TFs, prompting us to examine whether other Drosophila ETS-coding genes may also affect ageing. We show that five out of eight Drosophila ETS TFs play a role in fly ageing, acting from a range of organs and cells including the intestine, adipose and neurons. We expand the repertoire of lifespan-limiting ETS TFs in C. elegans, confirming their conserved function in ageing and revealing that the roles of ETS TFs in physiology and lifespan are conserved throughout the family, both within and between species.
| Understanding the basic biology of ageing may help us to reduce the burden of ill-health that old age brings. Ageing is modulated by changes to gene expression, which are orchestrated by the coordinate activity of proteins called transcription factors (TFs). E-twenty six (ETS) TFs are a large family with cellular functions that are conserved across animal taxa. In this study, we examine a longevity-promoting transcriptional circuit composed of two ETS TFs, Pnt and Aop, and Foxo, a forkhead TF with evolutionarily-conserved pro-longevity functions. This leads us to demonstrate that the activity of the majority of ETS TFs in multiple tissues and even different animal taxa regulates lifespan, indicating that roles in ageing are a general feature of this family of transcriptional regulators.
| Ageing is characterised by a steady systematic decline in biological function, and increased likelihood of disease[1]. Understanding the basic biology of ageing therefore promises to help improve the overall health of older people, who constitute an ever-increasing proportion of our populations. In experimental systems, healthy lifespan can be extended by altered transcriptional regulation, coordinated by sequence-specific TFs[2–6]. Thus, understanding TFs’ functions can reveal how to promote health in late life. Forkhead family TFs, especially Forkhead Box O (Foxo) orthologues, have been studied extensively in this context. This effort has been driven by the association of Foxo3a alleles with human longevity[7]; and the findings that the activation of Foxos is necessary and sufficient to explain the extension of lifespan observed following reduced insulin/IGF signalling (IIS) in model organisms[8–11]. Foxos interact with additional TFs in regulatory circuits, and it is in this context that their function must be understood. For example, in Caenorhabditis elegans, the pro-longevity activity of Daf-16 is orchestrated with further TFs including Hsf, Elt-2, Skn-1, Pqm-1 and Hlh-30/Tfeb [3,12–15]. Examining regions bound by Foxos across animals has highlighted the conserved presence of sites to bind ETS family TFs[16]. In Drosophila, two members of this family, namely Aop (a.k.a. Yan) and Pnt, have been linked to ageing via genetic interactions with Foxo and IIS[4], and similar interactions are evident in C. elegans [17]. These findings raise questions of the overall roles of ETS factors in ageing, and their relationship to the activities of Foxos.
The ETS TFs are conserved across animals, including 28 representatives in humans[18,19]. Their shared, defining feature is a core helix-turn-helix DNA-binding domain, which binds DNA on 5’-GGA(A/T)-3’ ETS-binding motifs (EBMs). They are differentiated by tissue-specific expression, and variation in peripheral amino acid residues which, along with variation in nucleotides flanking the core EBM, confers DNA-binding specificity[20]. ETS TFs generally function as transcriptional activators, but a few repress transcription[21,22]. Aop is one such repressor in Drosophila. Aop and its human orthologue Tel are thought to repress transcription by competing with activators for binding sites[21,23], recruiting co-repressors[22,24,25], and forming homo-oligomers that limit activator access to euchromatin[26–30]. Consequently, Aop's role in physiology must be explored in the context of its interactions with additional TFs, especially activators. Foxo is one such activator[31]. Both Foxo and Aop are required for longevity by IIS inhibition[9], each is individually sufficient to extend lifespan[4], and both are recruited to the same genomic loci in vivo. Whilst activating either in the gut and fat body extends lifespan, the effect of activating both is not additive. Furthermore, if Aop is knocked down, activating Foxo not only ceases to extend lifespan, but even becomes deleterious for lifespan[4]. Overall, these findings suggest that gene expression downstream of IIS is orchestrated by the coordinated activity of Aop and Foxo, and that there is a redundancy in the function of the two TFs, even though Foxo is a transcriptional activator and Aop a transcriptional repressor. We started this study by characterising Aop and its relationship with relevant transcriptional activators, including Foxo. This led us to uncover that roles in ageing are widespread throughout the ETS TF family, extending across multiple fly tissues and diverse animal taxa.
How does the transcriptional programme triggered by Aop relate to that triggered by Foxo? We sought to identify genes that were differentially regulated in response to activation of either TF. We focused on adult female fly guts and fat bodies (equivalent to mammalian liver and adipose), since these are the organs from which Foxo and Aop promote longevity[4]. We induced expression of Foxo, AopACT (encoding a constitutively active form of AOP) or both under the control of the S1106 driver by feeding flies with the RU486 inducer. We profiled genome-wide transcriptional changes in dissected guts and abdominal fat bodies (as associated with the cuticle) with RNA-Seq and identified genes responding to RU486 within each genotype at a False Discovery Rate (FDR) of 10% (these and all subsequently mentioned gene set assignments are given in S1 Supplementary Information, along with full statistics for all genes in all genotypes; the key to the location of each sheet is contained within the S1 Supplementary Information). In both tissues, we found that the sets of genes regulated by either Foxo or AopACT overlapped significantly (gut p<10−19, fat body p<10−4, Fig 1A). To further assess whether Aop’s and Foxo’s transcriptional programmes were similar, we tested for correlated expression changes in response to the two TFs within the union of all 712 genes differentially regulated by either TF in the gut, or the equivalent 727 genes in the fat body. The transcriptional programmes triggered by Foxo or AopACT were significantly correlated within these unions (Fig 1B and 1C, Kendall's Tau rank-correlation test: gut tau = 0.17, p = 1e-14; fat body tau = 0.32, p<2.2e-16). Interestingly, the sets of differentially expressed genes were largely tissue-specific (S1 Fig), suggesting that this correlated response may be a general feature of the Aop and Foxo regulons and independent of the tissue-specificity of target promoters. Gene Ontology (GO) enrichment analysis suggested that, in the gut, this combined set of Aop- and Foxo-regulated genes tended to be involved in translation and energy metabolism, whilst the equivalent analysis in the fat body showed enrichment for regulators of gene expression (details of this GO analysis and all those subsequently mentioned are given in S1 Supplementary Information). We independently confirmed this correlated response to Aop and Foxo using qRT-PCR of two transcripts identified by transcriptomics: a characterised transcriptional target of IIS [32], tobi (Fig 1D, linear model: RU486 F1,13 = 26.04, p = 2e-4; no effect of genotype, full details of this and all subsequent linear models are contained in one sheet of the S1 Supplementary Information), and alcohol dehydrogenase (Adh, Fig 1E—linear model: RU486 F1,9 = 7.83, p = 0.02; no effect of genotype). Hence, Aop and Foxo not only promote longevity, but also individually effect equivalent transcriptional programmes.
What are the outcomes of combining Aop and Foxo activity? FOXO co-localises extensively with AOP in the genome, with 60% of FOXO-bound loci also bound by AOP in the adult gut and fat body[4]. Since AOP functions by repressive interactions with transcriptional activators, we hypothesised that FOXO activity would be modulated by AOP. We tested this hypothesis in vitro. Transcriptional reporters were constructed by combining the Adh basal promoter with FOXO-responsive elements (FREs: AACA), ETS-binding motifs (EBMs: GGAA) or both, and examined for their response to FOXO and AOPACT in Drosophila S2 cells (Fig 2A, S2 Fig). In the presence of EBMs, AOP prevents activation by ETS activators [e.g. 33]. In the presence of FREs, FOXO is known to activate transcription [31]. We confirmed published observations for individual TFs on the reporters that contained their individual binding elements: FOXO was sufficient to activate transcription from the FREs (t-test t = 6.64, p = 3.7e-5), while, as expected [23,26,28], AOPACT did not impact expression from EBMs (t-test t = -0.66, p = 0.26).
We conducted three replicate experiments to assess the interactive output of AOP and FOXO. Combining the FREs and EBMs allowed AOPACT to attenuate activation by FOXO, revealing that AOP can moderate FOXO’s activity when brought onto the same promoter. By striking contrast, in the absence of EBMs, AOPACT synergised with FOXO to stimulate induction to an order of magnitude greater than FOXO alone, indicating that AOPACT can indirectly accentuate FOXO’s ability to activate transcription (Fig 2A). While the magnitude of these effects varied, it was consistently present across three independent experiments (S2 Fig). To analyse these data we used linear modelling, testing how the complement of TF binding motifs altered the output of combining the TFs, both in each individual experiment, and across the three experiments. This analysis confirmed that the output of combining AOP and FOXO was promoter-dependent (Linear model: FOXO:AOP:FRE:EBM—data from all three replicates, F1,158 = 21.06, p = 9e-6; data from Fig 2A F1,48 = 15.34, p = 2e-4; see also statistical analysis section of S1 Supplementary Information). Since the synergistic interaction occurred in the absence of EBMs, this is most likely an indirect effect, occurring not via a direct interaction on the promoter but rather via AOP-induced transcriptional changes elsewhere in the genome. Note that in the presence of EBMs, any synergistic effect of AOP appears counteracted by the repression occurring from direct AOP binding to the promoter. Synergy may account in part for the similarity of AOP’s and FOXO’s transcriptional programmes in vivo (Fig 1). Hence, AOP is not only able to moderate the activity of other ETS activators, but also the Forkhead TF FOXO, with the presence or absence of EBMs in a promoter determining whether AOP enhances or moderates FOXO activity.
The in vitro analysis suggested that Foxo’s in vivo output should depend on Aop activity. To examine if synergy and antagonism of Foxo by Aop can be observed on native promoters in vivo, we looked at what happens when Aop and Foxo were combined. We used our above-described RNA-Seq experiment and sorted the union of differentially expressed genes by the direction of regulation upon induction of Foxo, AopACT, or both, paying attention to altered regulation when the TFs were co-induced. To visualise the groupings, we compared the fold-change values for each gene between different conditions by calculating per gene Z-score (number of standard deviations away from the mean fold-chage; Fig 2B and 2C). In this way, we could identify sets of genes that may be synergistically or antagonistically regulated by Aop and Foxo. We note that neither Aop nor Foxo were significantly down-regulated by the other in either tissue, indicating that their combined transcriptomic outputs result from interactive effects on promoters (S1 Supplementary Information). We selected specific candidates for validation by qRT-PCR, and used linear models to test for interactive effects of the TFs, indicated by differential effects of RU486 feeding on the study genotypes. Indeed, we found that Aop was able to antagonise Foxo’s induction of aay and 4ebp in the gut (Fig 2D and 2E; aay genotype:RU486 F2,17 = 15.43, p = 1e-4; 4ebp genotype:RU486 F2,17 = 8.38, p = 2e-3; full analysis in S1 Supplementary Information). On the other hand, Aop synergised with Foxo to modulate expression of PGRP-SC2 in the gut and dilp6 in the fat body (Fig 2F and 2G; PGRP-SC2 genotype:RU486 F2,15 = 4.06, p = 0.03; dilp6 genotype:RU486 F2,17 = 6.61, p = 8e-3; full analysis in S1 Supplementary Information). Thus, transcript profiling followed by qRT-PCR validation confirmed that the two modes of AOP:FOXO interaction observed on synthetic reporters can also occur in vivo. This simultaneous synergy and antagonism of AOP and FOXO may explain why, whilst activation of either TF is sufficient to promote longevity, their co-activation does not extend lifespan additively[4].
Whilst interactions with FOXO appear to account for some of the transcriptional outputs of AOP, 80% of AOP-bound genomic sites are not bound by FOXO in vivo[4]. Since AOP alone is insufficient to regulate transcription when brought onto a promoter (Fig 2A and references[21,23,26,28]), interactions with other transcriptional activators must account for the full breadth of Aop's physiological and transcriptomic effects. Pnt is one such transcriptional activator. Pnt and Aop have mutually antagonistic roles in development, which is presumed to occur by competition for binding sites since the two recognise the same DNA sequence[23,30,34]. We confirmed this interaction on reporters in S2 cells: Transcriptional induction by PNTP1 (a constitutively active isoform[35]) was completely blocked by AOPACT (Fig 3A, linear model AOP:PNT F1,16 = 41.8, p = 7.9e-6; also see references[23,28,36,37]), suggesting that PNT inhibition may be a key factor in Aop’s pro-longevity effect. Additionally, Pnt over-expression can block the longevity effects of both Foxo and IIS [4,9], suggesting that Pnt may also modify Foxo’s transcriptional output. To evaluate emergent interactions in vivo, the transcriptome-wide effects of co-expressing AopACT, PntP1 and Foxo in the gut and fat body were examined.
We assessed the transcriptomic outcomes of induction of PntP1 either alone or in combination with AopACT and Foxo (note that this is an extension of the above-described transcriptomic experiment, which was performed at the same time). For each of the gut and the fat body, we assembled sets of genes that were differentially regulated upon induction of any of the three TFs or their combinations (union of all genes differentially expressed at FDR 10%, set assignments per tissue in S1 Supplementary Information, noting that the preceding Foxo/Aop-regulated genes are a subset). This formed a union of 945 genes in the gut, and 1214 genes in the fat body. We sorted these genes by their pattern of regulation (i.e. set assignment) and visualised the groupings based on per-gene Z-score. This revealed a complex pattern in both tissues where each TF appeared able to influence the outcomes of the other two (Fig 3B and 3C). To distil these interactions, we tested explicitly for genes whose regulation is subject to a statistically significant three-way interaction of Foxo, AopACT and PntP1 induction. In the gut, 511 transcripts were subject to the combinatorial, interactive effects of the three TFs, as were 617 in the fat body (10% FDR, see results in S1 Supplementary Information). To reveal emergent transcriptional programmes in each tissue, principal components analysis (PCA) was performed over these sets of transcripts (Fig 3D and 3E). Remarkably, the first principal component (PC) of differentially expressed genes in the gut distinguished flies by published lifespan outcomes[4], with short-lived flies expressing PntP1 alone or in combination with Foxo at one end of the PC; long-lived flies expressing one or both Foxo and AopACT forming a distinct group at the other end of the PC; and AopACT countering the effect of PntP1 to form an intermediate group (Fig 3E). In the fat body, a similar grouping was apparent on the diagonal of PCs 1 and 2 (Fig 3D), despite more variability in the data, probably resulting from the difficulty of dissecting this organ. To infer functional consequences of these distinct transcriptional programmes, transcripts from the input set corresponding to the PCs were isolated and GO enrichment analysis performed. This revealed a strong enrichment of genes with roles in energy metabolism, whose expression was strongly correlated to the PCs (S3 Fig). Overall, a combined view of the PCA and GO analysis predicted that: (1) inhibiting Pnt may recapitulate the transcriptional programme of Aop and Foxo and promote longevity, and (2) that Pnt, alongside Aop and Foxo, may regulate metabolism in young flies.
Since Aop and Foxo appeared to drive a transcriptional programme opposed to that of Pnt, we hypothesised that directly limiting physiological levels of Pnt would be sufficient to recapitulate their effect on gene expression. We first assessed the transcriptome-wide changes in the gut and fat body induced by RNAi-mediated knockdown of Pnt. The sets of genes differentially regulated by Pnt knockdown (FDR 10%) significantly overlapped those regulated by AopACT or Foxo in both the gut and the fat body (Fig 4A). Additionally, in the union of the genes regulated by PntRNAi, AopACT or Foxo induction in the fat body, correlated effects of Pnt knockdown and Foxo or Aop activation were evident (Fig 4B and 4C). However, such broad correlations were not evident in the gut (Fig 4D and 4E). Hence, reducing the physiological levels of Pnt can recapitulate some aspects of the Aop/Foxo transcriptional programme. But is this sufficient to extend lifespan?
Inducing RNAi against Pnt from day three of adulthood in the gut and fat body was indeed sufficient to increase lifespan (Fig 4F, log-rank p = 7.2e-4). To further validate this finding, we backcrossed a loss-of-function p-element insertion in Pnt (PntKG04968, henceforth PntKG), into an outbred, wild-type background for ten generations. The mutation was homozygous lethal. However, heterozygote females exhibited a 20% increase in median lifespan (Fig 4G, log-rank p = 9.2e-11). We also tested whether expressing a transcriptional repressor of the Pnt locus extended lifespan. The HMG-box repressor capicua (cic) represses expression of Pnt[38] and, consistent with the effects of PntRNAi, overexpressing cicΔC2 (a cic mutant lacking a known MAPK phosphorylation site) in the gut and fat body also substantially extended lifespan (Fig 4H, log-rank p = 1.5e-7). These experiments demonstrated that countering Pnt is sufficient to recapitulate aspects of the Aop/Foxo transcriptional programme and extend lifespan, and corroborate the conclusions of transcriptomic analysis that Aop and Foxo act in part by countering Pnt.
Our data show that each member of the Foxo-Aop-Pnt circuit can be targeted in the gut and fat body to extend lifespan. What is the function of this circuit, and Pnt in particular, in young flies, before ageing occurs? The RNA-Seq data sets suggested metabolic regulation. Since the levels of Pnt appeared to dictate transcriptional and lifespan outcomes (Figs 3 and 4), we evaluated its metabolic role in more detail. The presence of genes including lipases and perilipin (Lsd-2) in the transcriptome data suggested that Pnt modulates lipid metabolism. Therefore, we applied nutritional stresses to alter triacylglyceride (TAG) storage, and assessed how PntP1 altered the response to these stresses. We quantified TAG after a week of PntP1 induction, and then after a subsequent six days of starvation. PntP1 accentuated the loss of TAG per unit weight induced by starvation (Fig 5A; linear model RU486:starvation F1,19 = 7.03, p = 0.02), but not overall weight loss (S4 Fig), suggesting that Pnt sensitises flies specifically to cues for lipolysis. The mobilisation of TAG stores was associated with decreased resistance to starvation, with flies over-expressing PntP1 dying 24% earlier on average (Fig 5B log-rank p = 1.3e-14). This ability of Pnt to promote catabolism of energy stores may be beneficial in the face of over-nutrition, and relevant to the Western human epidemic of metabolic disease associated with energy-rich diets. A Drosophila model of such energy-rich diets is feeding flies a high sugar diet. Flies fed 40% sugar die substantially earlier than controls fed a 5% sugar diet, and accumulate TAG[39,40]. However PntP1 overexpression restored TAG levels in flies on a high-sugar diet to those observed on a low-sugar diet (Fig 5C). Whilst there was no statistically significant interaction of sugar and PntP1 induction in a linear model (RU486:sugar F1,17 = 0.32, p = 0.57), the adipogenic effect of sugar was opposed by Pnt, such that TAG levels on a high-sugar diet with Pnt induction were not different from those on a low-sugar diet without Pnt induction (t-test: t = 0.01, p = 0.99). Moreover, PntP1 induction spared flies from the full extent of the early death induced by dietary sugar, increasing median survival time by 26%, despite having no effect on the low-sugar diet (Fig 5D; cox proportional hazards RU486:sugar p = 6.2e-3). Altogether, these results indicate that while Pnt activity is detrimental during ageing, in youth it predisposes flies to leanness, which correlates survival of nutritional stress. This may suggest that metabolic regulation is an adaptive function of the Foxo-Aop-Pnt circuit in early life, but that the configuration which is optimal for metabolism is deleterious for later survival.
Animal genomes encode multiple ETS factors: In Drosophila the ETS family comprises the repressor Aop and activators: Pnt, Eip74EF, Ets21C, Ets65A, Ets96B, Ets97D and Ets98B, each of which is expressed with its own unique tissue-specific pattern (Fig 6A). Finding lifespan-limiting roles of Pnt in addition to the previously described pro-longevity role of Aop, suggested that other ETS TFs with functions as transcriptional activators may also have the same lifespan-limiting effect. We examined the function of the other ETS TFs in Drosophila lifespan by knocking down their expression levels with RNAi in combination with inducible drivers. The data obtained in >40 lifespan assays are summarised in Fig 6B, including information on Aop, Pnt and Foxo. Summary statistics of each lifespan, along with associated genetic information, are presented in S1 Supplementary Information, while individual lifespan curves are presented in S5–S8 Figs. We identified each of Eip74EF, Ets21C and Ets97D as limiting lifespan in at least one tissue. For Ets21C, we confirmed the result using an available mutant (S5 Fig). Whilst some of the effects we observed were modest, overall the data pointed to roles in ageing for five of the eight Drosophila ETS TFs.
The effects were in general tissue specific. RNAi against Pnt, Ets21C and Ets97D restricted to the gut and the fat body with the S1106 driver extended lifespan (Fig 6B, S5 Fig), the same tissues where Foxo and Aop act [4]. Knockdown of Pnt but not of Ets21C in enterocytes (ECs), using the GS5966 driver, was sufficient to extend lifespan, as was activating either AopACT or Foxo (Fig 6B, S6 Fig). Pnt and Ets21C have been characterised as regulators of intestinal stem cell (ISC) proliferation[38], but activating cognate RNAi constructs with drivers that are active in ISCs (GS5961 and TIGS) did not consistently or substantially extend lifespan (Fig 6B, S6 Fig). Pnt functions in neurogenesis[41], and its continued expression in adult neurons suggested an ongoing, physiologically relevant role. However, neuronal PntRNAi induction, achieved with the ElavGS driver, did not affect lifespan, while over-expressing either AopACT or Foxo was deleterious and in contrast to their benefits in gut and fat body (Fig 6B, S7 Fig). Eip74EF is more highly expressed in the brain than other tissues (Fig 6A), indicating that neurons may mediate the beneficial effect of its ubiquitous knockdown. Indeed, expressing Eip74EFRNAi in neurons using the inducible, neuron-specific driver Elav-GS extended lifespan (Fig 6B, S7 Fig). Overall, these data show that members of the Drosophila ETS family, along with Foxo, have distinct effects on lifespan in distinct tissues.
The ETS TFs act downstream of receptor tyrosine kinase (RTK) pathways. The insulin receptor InR is an established regulator of Aop and Foxo[8], and reducing its activity promotes lifespan[9]. Whilst expressing InRDN (a dominant-negative form) in the gut and fat body enhanced lifespan, expressing the same construct in ECs did not (Fig 6B, S8 Fig), indicating that another RTK may function upstream of ETS TFs and Foxo in the ECs. The epidermal growth factor receptor EGFR can signal to Pnt and Ets21C via cic[38], suggesting EGFR in ECs may regulate lifespan. Indeed, inducing the dominant-negative form EGFRDN in the gut and fat body or ECs extended lifespan (Fig 6B, S8 Fig). Hence, different ETS factors may limit lifespan downstream of different RTK pathways in different tissues.
The evidence suggested that a role in ageing is shared amongst multiple ETS factors in Drosophila. ETS TFs are conserved throughout multicellular animals, and the extensive conservation of roles in lifespan amongst the ETS family in the fly suggested that this lifespan modulation may be a fundamental property of these TFs, that extends to other species. The genome of the nematode C. elegans encodes 11 ETS TFs in total. At least one of these, Ets-4, has been reported to limit lifespan in the worm intestine[17]. We screened the majority of the other C. elegans ETS TFs for roles in lifespan by feeding worms RNAi from egg or L4 onwards (S1 Supplementary Information). Expanding the repertoire of proteins that limit worm lifespan, we found that knockdown of Lin-1 (an orthologue of human ELK1, ELK3 and ELK4) consistently extended C. elegans lifespan, in multiple independent trials from L4 stage or egg (e.g. Fig 6C). Thus, multiple ETS factors limit lifespan in species separated by hundreds of millions of years of evolutionary divergence, hinting at a general role for this family of TFs in animal longevity.
Promoting healthy ageing by transcriptional control is an attractive prospect, because targeting one specific protein can restructure global gene expression to provide broad-scale benefits. This study suggests key roles for ETS TFs in such optimisation. The results show dual roles for Aop: balancing Foxo’s outputs, and opposing Pnt’s outputs. These functions coordinate transcriptional changes that correspond to lifespan. Repressing transcription from the ETS site appears to be the key longevity-promoting step, and indeed lifespan was extended by limiting multiple ETS TFs, in multiple fly tissues, and in multiple taxa. Altogether, these results show that inhibiting lifespan is a general feature of ETS transcriptional activators. Presumably the expression of these TFs is maintained, despite costs in late life, because of benefits in other contexts. For example, Pnt is important during development[23,34–36], and expression may simply run-on into adulthood. We have now shown that Pnt is also important for adults facing nutritional variation or stress, and genomic evidence suggests equivalent functions for Ets-4 in C. elegans[17]. In addition, Ets21C is required to mount an effective immune response[42], and both Ets21C and Pnt control gut homeostasis[38]. Tissue environment appears to be another important contextual factor that determines the lifespan effects of specific ETS TFs. Differences between tissues in chromatin architecture are likely to alter the capacity of a given TF to bind a given site, and our results show that a given TF, and also upstream RTKs, do not necessarily lead to the same lifespan effect across all tissues. The tissue-specific functions that we show for ETS TFs, Foxo and RTKs, suggests that transcription is locally coordinated by distinct receptors and TFs in distinct tissues, but that lifespan-regulatory signalling nevertheless converges on the ETS site. This differentiation makes it all the more remarkable that roles in lifespan appear to be conserved amongst ETS family TFs, even in diverse tissue contexts.
The structure of molecular networks and their integration amongst tissues underpins phenotype, including into old age. Unravelling the basics of these networks is a critical step in identifying precise anti-ageing molecular targets. Identifying the least disruptive perturbation of these networks, by targeting the “correct” effector, is a key goal in order to achieve desirable outcomes without undesirable trade-offs that may ensue from broader-scale perturbation. This targeting can be at the level of specific proteins, cell types, points in the life-course, or a combination of all three. The tissue-specific expression pattern of ETS TFs, and the apparent conservation of their roles in longevity, highlights them as important regulators of tissue-specific programs that may be useful in precise medical targeting of specific senescent pathologies.
All experiments were carried out in outbred, Wolbachia-free Dahomey flies, bearing the w1118 mutation and maintained at large population size since domestication in 1970. All transgenes (S1 Supplementary Information) were backcrossed into this background at least 6 times prior to experimentation, and stocks were maintained without bottlenecking. Cultures were maintained on 10% yeast (MP Biomedicals, OH, USA), 5% sucrose (Tate & Lyle, UK), 1.5% agar (Sigma-Aldrich, Dorset, UK), 3% nipagin (Chemlink Specialities, Dorset, UK), and 0.3% propionic acid (Sigma-Aldrich, Dorset, UK), at a constant 25°C and 60% humidity, on a 12:12 light cycle. Experimental flies were collected as embryos following 18h egg laying on grape juice agar, cultured at standardised density until adulthood, and allowed to mate for 48h before males were discarded and females assigned to experimental treatments at a density of 15 females/vial. To induce transgene expression using the GeneSwitch system, the inducer RU486 (Sigma M8046) was dissolved in absolute ethanol and added to the base medium to a final concentration of 200 μM. Ethanol was added as a vehicle control in RU-negative food. For lifespan experiments, flies were transferred to fresh food and survival was scored thrice weekly. Feeding RU486 to driver-only controls did not affect lifespan (S1 Supplementary Information). For starvation stress experiments, flies were fed RU486 or EtOH-supplemented media for one week, before switching to 1% agarose with the equivalent addition of RU486 or EtOH, with death scored daily until the end. For sugar stress experiments, sugar content was increased to 40% w/v sucrose[39,40].
Worms were maintained by Brenner’s protocol[43], at 20°C on NGM plates seeded with Escherichia coli OP50. For lifespan experiments, N2 (wildtype N2 male stock, N2 CGCM) were used at 20°C on NGM plates supplemented with 15μM FUDR to block progeny production. RNAi treatment was started from egg or late larval stages (details in Supplementary Materials). Animals that died from internal hatching were censored.
The pGL3Basic-4xFRE-pADH-Luc construct (called pGL4xFRE, reference [31]) was used as template to generate PCR products containing 6xETS-4xFRE-pADH, 4xFRE-pADH, 6xETS-pADH- or pADH (primers in S1 Supplementary Information, ETS sequence described in [44]), flanked by XhoI and HindIII sites, cloned into the corresponding sites in pGL3-Basic and confirmed by sequencing. PntP1 was amplified from UAS-PntP1 genomic DNA with Q5 High-Fidelity Polymerase (NEB M0491S - primers in S1 Supplementary Information), and AopACT was cloned from genomic DNA of UAS-AopACT flies[4]. PntP1 and AopACT sequences were then cloned into the pENTR-D-TOPO gateway vector (Thermo 450218) before recombination into the pAW expression vector.
Drosophila S2 cells were cultured in Schneider’s medium (Gibco/Thermo Scientific 21720024), supplemented with 10% FBS (Gibco/Thermo Scientific A3160801) and Penicillin/Streptomycin (Thermo 15070063). Cells were split into fresh media 24h before transfection, then resuspended to a density of 106 ml-1 and transfected using Effectene reagent (Qiagen 301425) in 96-well plates, according to the manufacturer’s instructions. Reporters and TF expression plasmids were co-transfected with pAFW-eGFP to visually confirm transfection, and pRL-TK-Renillaluc as an internal control for normalisation of reporter-produced Firefly luciferase. Reporters and pRL-TK-Renillaluc were transfected 1:1. When multiple TF expression plasmids were transfected, it was done 1:1. Each TF expression plasmid was transfected 4:1 relative to reporters or pRL-TK-Renillaluc (i.e. for every ng TF expression plasmid, 0.25 ng reporter and 0.25 ng pRL-TK-Renillaluc were transfected). The total amount of DNA transfected was then topped up to a standard quantity across all experimental conditions with pAFW-eGFP, in equal volumes of TE buffer. Reporter activity was measured 18h after transfection using Dual-Luciferase reagents (Promega E1960). pAHW-Foxo and/or pAW-AopACT were co-transfected with promoters bearing combinations of FREs and EBMs. pAW-AopACT and pAW-PntP1 were co-transfected with a promoter bearing EBMs.
Flies bearing combinations of UAS-Foxo, UAS-AopACT and UAS-PntP1, or UAS-PntRNAi in an S106-GS background were dissected after six days adult feeding on RU486. Tissues were dissected in ice-cold PBS. Guts were dissected by cutting off the head and last abdominal segment, pulling on the crop through an incision at the abdomenal-thoracic junction, then removing tubules. Reproductive anatomy was then removed from the abdomen and the remainder of the abdomen taken as fat body. Dissected tissues were placed directly into ice-cold Trizol (Ambion 15596026). In the Foxo-AopACT-PntP1 epistasis RNA-Seq experiment, four experimental replicates were sampled per condition, each comprising a pool of 12 fat bodies or guts. In the PntRNAi experiment three replicates were sampled per condition, also each comprising organs from 12 flies. RNA was extracted by Trizol-chloroform extraction, quantified on a NanoDrop, and integrity was assessed on an Agilent Bioanalyzer. Poly(A) RNA was pulled down using NextFlex Poly(A) beads (PerkinElmer NOVA-512981) and integrity was re-assessed. In the Foxo-AopACT-PntP1 epistasis RNA-Seq experiment, only samples with the highest RNA yields and integrity were included in library preparation, leaving 2–3 samples per experimental condition. All three replicates were prepped and sequenced in the PntRNAi RNA-Seq experiment. RNA fragments were given unique molecular identifiers and libraries were prepared for sequencing using NextFlex qRNAseq v2 reagents, (barcode sets C and D, PerkinElmer NOVA-5130-14 and NOVA-5130-15) and 16 cycles of PCR. Individual and pooled library quality was assessed on an Agilent Bioanalyzer and quantified with a Qubit spectrophotometer. Sequencing was performed by the UCL Cancer Institute, using an Illumina HiSeq 2500 instrument (paired-end 50bp) for the Foxo-AopACT-PntP1 epistasis experiment, and a NextSeq 500 (paired-end 75bp) for the PntRNAi experiment.
cDNAs were made from the polyA RNA preps that were prepared for sequencing, using SuperScript II Reverse Transcriptase (Thermo 18064014) and OligoDT. qRT-PCR was performed on an Applied Biosystems QuantStudio 6 Flex real-time PCR instrument with Fast SYBR Green PCR Master Mix (Thermo Fisher), with primers supplied by EuroFins Genomics (all oligo sequences in S1 Supplementary Information), relative to a standard curve comprising a pool of all samples and the instrument's standard PCR cycle.
TAG was measured as in [45] in whole adult S106; UAS-PntP1 flies following one week of RU486 feeding. Briefly, flies were CO2-anaesthetised, weighed on a microbalance, and immediately flash-frozen in liquid N2. Flies were thawed in ice-cold TEt buffer (10 mM Tris, 1 mM EDTA, 0.1% v/v Triton-X-100) and homogenised by shaking with glass beads (Sigma G8772) for 30s in a ribolyser at 6500 Hz. Aliquots of homogenates were heated to 72°C for 15m to neutralise enzymatic activity, then spun 1m at 4500g and 4°C to pellet debris. Triglyceride was measured by treating 5 μl sample with 200 μl Glycerol Reagent (Sigma F6428) for 15m at 37°C and measuring absorbance at 540 nm, then incubating with 50 μl Triglyceride Reagent (Sigma F2449) for 15m at 37°C and re-measuring absorbance at 540 nm, calculating glycerol content in each reading, then quantifying triglyceride content as the difference between the first and second measurement.
Sequence libraries were quality-checked by FastQC 0.11.3, duplicate reads were removed using Je 1.2, and reads were aligned to D. melanogaster genome 6.19 with HiSat2 2.1. Alignments were enumerated with featureCounts 1.6. All downstream analyses were performed in R 3.3.1. The gut and fat body were analysed in parallel. In the RNA-Seq experiment analysing Foxo-AopACT-PntP1 epistasis, genes with no counted transcripts were excluded (S1 Supplementary Information). In the subsequent PntRNAi experiment, genes were filtered by the same criteria and any genes that were not analysed in the first experiment were also excluded. Read counts are given in S1 Supplementary Information. The transcriptomic effect of RU486 feeding was established for each individual genotype in the experiment, using DESeq2 at a false discovery rate (IHW) of 10%. To identify correlated effects amongst genotypes, sets of shared targets were formed as unions of DE gene sets from individual genotypes. Log2 fold-change values (Figs 1–4) were plotted from the DESeq2 output. Three-way epistatic interaction amongst TFs were identified by fitting models of the form
yi∼genotype+RU486+block+genotype:RU486
where block represented experimental replicate. The tripartite interaction of Foxo, AopACT and PntP1 was identified by applying the model to all genes across all experimental conditions, and isolating genes with a significant genotype:RU486 term.
GO analysis was performed using the TopGO package, applying Fisher’s test with the weight01 algorithm. Principal Components Analysis was performed on read counts of these genes following a variance-stabilizing transformation. To characterise gene-expression correlates of principal components, loadings onto principal components were extracted using the dimdesc function from the FactoMineR library, and GO analysis performed as previously. Transcripts of genes annotated with enriched GO terms were then plotted per term by centring variance-stabilised reads to a mean of zero and plotting against PC values per sample. Heatmaps were plotted using the heatmap.2 function from the gplots library, ordering rows by hierarchical clustering by Ward’s method on Euclidian distance, and scaling to row.
Fly lifespan data were analysed using log-rank tests in Microsoft Excel or Cox Proportional Hazards in R for the interaction of sugar and PntP1 expression. Worm lifespan data were analysed by log-rank tests in JMP.
Luciferase reporter data were normalised by taking the ratio of firefly luciferase to renilla luciferase signal and, for each promoter, taking the median reporter signal in the absence of FOXO and AOPACT as the start value, then calculating fold-change (i.e. difference in start and end values, divided by start value) for each sample. To assess the interaction of FOXO and AOP with promoters’ complements of TF-binding motifs, these normalised data were analysed by fitting a linear model of the form
y∼FRE*EBM*FOXO*AOPACT
in which y was the natural log of fold-change+1, FRE and EBM represented the TF-binding complement, and FOXO and AOPACT represented co-transfection with pAHW-Foxo or pAW-AopACT. By the same approach, the interactive effect of PNTP1 and AOPACT were assessed by fitting a linear model of the form
y∼PNTP1*AOPACT
in which y represented the natural log of fold-change+1.
The effect of PntP1 overexpression on TAG and lifespan responses to nutrient stress (starvation or high-sugar diet) were analysed by a model of the form
y∼RU486*diet
where y represented TAG normalised to unit weight in a linear model, or survival in a Cox Proportional Hazards model (survival library).
|
10.1371/journal.pbio.1000604 | A High Precision Survey of the Molecular Dynamics of Mammalian Clathrin-Mediated Endocytosis | Dual colour total internal reflection fluorescence microscopy is a powerful tool for decoding the molecular dynamics of clathrin-mediated endocytosis (CME). Typically, the recruitment of a fluorescent protein–tagged endocytic protein was referenced to the disappearance of spot-like clathrin-coated structure (CCS), but the precision of spot-like CCS disappearance as a marker for canonical CME remained unknown. Here we have used an imaging assay based on total internal reflection fluorescence microscopy to detect scission events with a resolution of ∼2 s. We found that scission events engulfed comparable amounts of transferrin receptor cargo at CCSs of different sizes and CCS did not always disappear following scission. We measured the recruitment dynamics of 34 types of endocytic protein to scission events: Abp1, ACK1, amphiphysin1, APPL1, Arp3, BIN1, CALM, CIP4, clathrin light chain (Clc), cofilin, coronin1B, cortactin, dynamin1/2, endophilin2, Eps15, Eps8, epsin2, FBP17, FCHo1/2, GAK, Hip1R, lifeAct, mu2 subunit of the AP2 complex, myosin1E, myosin6, NECAP, N-WASP, OCRL1, Rab5, SNX9, synaptojanin2β1, and syndapin2. For each protein we aligned ∼1,000 recruitment profiles to their respective scission events and constructed characteristic “recruitment signatures” that were grouped, as for yeast, to reveal the modular organization of mammalian CME. A detailed analysis revealed the unanticipated recruitment dynamics of SNX9, FBP17, and CIP4 and showed that the same set of proteins was recruited, in the same order, to scission events at CCSs of different sizes and lifetimes. Collectively these data reveal the fine-grained temporal structure of CME and suggest a simplified canonical model of mammalian CME in which the same core mechanism of CME, involving actin, operates at CCSs of diverse sizes and lifetimes.
| The molecular machinery of clathrin-mediated endocytosis concentrates receptors at the cell surface in a patch of membrane that curves into a vesicle, pinches off, and internalizes membrane cargo and a tiny volume of extracellular fluid. We know that dozens of proteins are involved in this process, but precisely when and where they act remains poorly understood. Here we used a fluorescence imaging assay to detect the moment of scission in living cells and used this as a reference point from which to measure the characteristic recruitment signatures of 34 fluorescently tagged endocytic proteins. Pair-wise comparison of these recruitment signatures allowed us to identify seven modules of proteins that were recruited with similar kinetics. For the most part the recruitment signatures were consistent with what was previously known about the proteins' structure and their binding affinities; however, the recruitment signatures for some components (such as some BAR and F-BAR domain proteins) could not have been predicted from existing structural or biochemical data. This study provides a paradigm for mapping molecular dynamics in living cells and provides new insights into the mechanism of clathrin-mediated endocytosis.
| Clathrin-mediated endocytosis (CME) is the principal means by which mammalian cells internalize cell surface receptors (reviewed in [1]). Some 40 years of electron microscopy (EM), genetic, and biochemical studies are distilled in the canonical model of CME [2] (reviewed in Figure S1). Here, interaction of receptors with adaptor proteins stabilise nascent clathrin-coated pits (CCPs) at random sites on the plasma membrane [3]. Growing CCPs acquire cargo and invaginate via clathrin polymerization [4] and the coordinated action of curvature-inducing/sensing BAR [5] and F-BAR domain proteins [6],[7], ENTH domain proteins [8], and possibly actin [9]–[11]. The neck of the deeply invaginated CCP is severed in a mechanism involving the large GTPase dynamin [12],[13], and possibly a phosphoinositide (PI) phosphatase [14], to release a clathrin-coated vesicle (CCV), which uncoats through the action of GAK/auxilin [15],[16].
Understanding how the multiple components of CME are spatially and temporally organized is a challenging problem that has been tackled using live-cell fluorescence microscopy (reviewed in [2],[17]). In a typical experiment using dual colour total internal reflection fluorescence microscopy (TIR-FM), the recruitment dynamics of fluorescent protein (FP)–tagged endocytic proteins were measured relative to the disappearance of spot-like CCPs, which was used as a fiducial marker to indicate internalization [6],[18],[19]. Using this strategy the recruitment dynamics of endocytic proteins were coarsely grouped into “early” and “late” relative to CCP disappearance [20] (Figure S1), but finer temporal resolution was not possible because the moment of scission, the endpoint of the invagination process, was unknown. In addition to spot-like CCPs, larger clathrin patches were also observed at the substrate proximal surface of many cell types, where they were variously thought to participate in the canonical pathway of CME [4],[21] or cell adhesion [22],[23], or were thought to represent endocytic intermediates in an actin-dependent mode of endocytosis distinct from the canonical pathway of CME [23].
To circumvent the subjective classification of endocytically active clathrin-coated structures (CCSs), a TIR-FM assay was invented to detect single scission events directly by monitoring the accessibility of pH-sensitive fluorescent CCP cargo to rhythmically imposed changes in extracellular pH (the “pulsed pH” [ppH] assay [10], reviewed in Figure S2). Surprisingly, it was discovered that scission events were hosted by spot-like CCPs, as predicted from the canonical model, and also by larger clathrin patches (collectively referred to as CCSs [10]), thus raising questions about what characterises endocytically active CCS at optical resolution.
The following study was designed to explore the fine-grained temporal structure of late stages of the mammalian CME machinery using TIR-FM and the ppH assay. First, scission events were mapped to their host CCSs to determine what dynamic characteristics defined endocytically active CCSs. It was found that CCSs of diverse size and lifetimes hosted scission events that engulfed comparable amounts of receptor cargo, and CCSs could either disappear (“terminal events”) or persist (“non-terminal events”) following scission. Second, we assessed the accuracy of CCS disappearance as a fiducial marker for internalization and showed it introduced an error comparable to the time course of CCS invagination and CCV formation. It was thus necessary to use the ppH assay to obtain a precise measurement of recruitment dynamics. Third, we surveyed the recruitment dynamics of a representative set of 34 mammalian endocytic proteins to sites of scission and derived, for each protein, a characteristic “recruitment signature” by aligning and averaging ∼1,000 recruitment traces per protein. A cluster analysis of recruitment signatures revealed the modular organization of the CME machinery, similar to yeast [24], while closer inspection revealed unanticipated features of some signatures. Finally, scaling relationships between CCS size and lifetime and the cohort of endocytic proteins recruited to scission events were explored. It was found that the same set of proteins was recruited in the same order to scission events at diverse dynamic classes of CCSs, although subtle scaling relationships between CCS size and protein recruitment were identified.
Collectively these data provide, to our knowledge, the highest resolution temporal map of the late stages of mammalian CME to date. This map (1) suggests a simplified model of mammalian CME in which the same core mechanism can operate at both spot-like CCSs and larger clathrin patches observed with fluorescence microscopy, (2) illustrates the similar modular organization of mammalian and yeast endocytosis, and (3) proves that recruitment dynamics of endocytic proteins such as the F-BAR protein FBP17 and BAR domain protein SNX9 cannot always be predicted from biochemical or structural properties.
To detect CME scission events at CCSs, NIH-3T3 cells were transiently transfected with Clc-mCherry and TfR-phl and assayed using the ppH assay, as described previously [10]. A large-diameter perfusion tip was brought close to the target cell, and perfusate was cycled between buffer of pH 7.4 and pH 5.5 in synchrony with image acquisition at 0.5 Hz (see [10] and Figure S2). In an image acquired at arbitrary time point t, at pH 7.4, TfR-phl concentrated in spots and patches of Clc-mCherry and free in the plasma membrane fluoresced brightly (Figure 1A). When the perfusate was switched to pH 5.5 and an image was acquired 2 s later (at t+2 s), TfR-phl fluorescence at the plasma membrane was quenched and revealed bright punctae of pH-insulated TfR-phl sequestered in internal vesicles, while Clc-mCherry fluorescence remained unchanged (Figure 1A). The cycle of pH switching and image acquisition was repeated to generate an image series acquired at alternating high and low pH. Scission events manifested as the abrupt appearance of TfR-phl spots in images acquired at pH 5.5, colocalized with Clc-mCherry-labelled CCSs (Figure 1B; Video S1). Although it took 4 s to complete a cycle of pH change, the precision with which scission events were detected was ∼2 s because, for an event to be detected, scission had to occur in a ∼2-s time window at pH 7.4 prior to detection at pH 5.5 (see [10] and Figure S2). We could therefore align the red fluorescence traces, acquired at 0.5 Hz, with an accuracy of 2 s. Visual inspection revealed that scission events were associated with both punctate CCSs and also larger, pleiomorphic clathrin patches (Figure 1C; Video S1), and events could occur repeatedly at larger CCSs, as shown previously [10] (Figure 1D). Larger CCSs may represent flat clathrin lattices, with peripheral invaginations, or clusters of smaller CCSs too close to resolve by optical microscopy [25],[26]. Inspection of kymographs revealed that Clc-mCherry and TfR-phl patches waxed and waned in synchrony at both small and large CCSs, demonstrating the similarity of these two signals and suggesting that TfR7 fluorescence could be used as a surrogate signal to report the relative size or lifetime of CCSs (Figure 1E). Scission events were not always associated with the disappearance of the host CCS, and, similar to previous findings, events were either terminal (where the spot-like CCS disappeared following scission, red arrows in Figure 1E) or non-terminal (where CCS persisted following scission, yellow arrows in Figure 1E) [10].
To analyse large numbers of scission events we developed a semi-automated analysis pipeline to identify candidate events, screen for bona fide events, and quantify the fluorescence changes associated with these events in both the green and red channels. The purpose of this screening strategy was not to detect all scission events in an image series but to impose stringent selection criteria and automatically sample a large proportion of genuine scission events. The criteria for selection of bona fide scission events included persistence of the TfR5 spot, association with a “host” CCS, adequate signal-to-noise ratio (SNR), and slope of the TfR5 signal following appearance (Figure 1B, see Materials and Methods for details).
To quantitatively investigate the characteristics of endocytically active CCSs we detected scission events in seven cells expressing Clc-mCherry and TfR-phl and identified a set of 851 bona fide events. First we analysed the relationship between the relative amount of TfR-phl localized at a CCS (TfR7 fluorescence), the relative amount of clathrin (Clc fluorescence), and the relative amount of TfR-phl internalized by a scission event (TfR5 fluorescence). As expected, there was a significant correlation between TfR7 fluorescence and Clc7 fluorescence (Spearman's rho = 0.85, p<0.05; Figure S3), showing that larger CCSs contained more TfR-phl cargo, and indicating that CCS size could be estimated using TfR7 fluorescence. However, there was no significant correlation between Clc7 and TfR5 fluorescence (Spearman's rho = −0.0024, p>0.05) or between TfR7 and TfR5 fluorescence (Spearman's rho = −0.0022, p>0.05; 3). Therefore, and consistent with both visual inspection of the current data and previous results [10], the amount of cargo internalized by scission events was independent of the size of the host CCS, and endocytically active CCSs could be either spot-like structures or larger, pleiomorphic clathrin patches. In mechanistic terms, this is consistent with the relatively constant dimensions of coated invaginations viewed by EM whether they occurred in isolation, as part of a cluster, or as a peripheral invagination at a flat patch of clathrin [25],[26]. To check that extracellular acidification did not affect the size of clathrin-coated invaginations, we fixed cells under control conditions and after exposure to acidic buffer for 1 min or 10 min, and imaged them using thin section EM (Figure S3). Under both control and acidified conditions the clathrin-coated invaginations were of relatively uniform size, with a maximum dimension of ∼100 nm (Figure S3F–S3I).
Next we explored what dynamic characteristics defined endocytically active CCSs. All CCSs present in the Clc-mCherry dataset (seven cells) were tracked using a multi-particle tracking algorithm, similar to previous studies [27] (see Materials and Methods), yielding a set of 11,447 track histories. For each CCS track history the fluorescence of Clc-mCherry was quantified, and the CCS track histories were classified according to the presence or absence of scission events, wherein a track history was defined as scission detected if a bona fide scission event fell within five pixels, or 500 nm. The median normalised Clc-mCherry fluorescence of scission detected CCSs was significantly greater than for scission undetected CCSs (0.190 versus 0.078, p<0.05; Figure 1F and 1G), and the median lifetime of scission-detected CCSs was longer than the lifetime of scission-undetected CCSs (189 s versus 38 s, p<0.05; Figure 1H and 1I). Therefore, scission events defined a class of larger, longer-lived CCSs. The shorter-lived scission-undetected CCSs most likely correspond to the “abortive” CCSs described previously [3],[27],[28], although some of these structures may have represented endosomal clathrin.
For NIH-3T3 fibroblasts the average time between de novo appearance of a spot-like CCS and the first detected scission event was previously found to be ∼100 s [10]. This was similar to previous estimates in BSC1 cells, wherein productive CCSs were defined as spot-like CCSs having lifetimes anywhere from tens to hundreds of seconds (average 87 s) [3],[27],[28]. Because the size and lifetimes of scission-detected CCSs were so variable (Figure 1H and 1I), in our subsequent investigation of late events in CME we made measurements over a time window of ±80 s, centred on scission.
In previous analysis of the molecular dynamics of CME, the disappearance of spot-like CCSs was used as a fiducial marker to indicate endocytic events [6],[18],[19]. However, we discovered that CCS disappearance gave an inaccurate and imprecise estimate of scission, with a temporal uncertainty comparable to the time course of CCS invagination and CCV formation [10] (−7±22 s; n = 107; six cells) (Figure 1J and 1K). CCS disappearance most likely corresponded to CCV uncoating and/or movement, and if CCS disappearance was used as a fiducial marker for CME the waveform of aligned and averaged recruitment signatures would be significantly smeared. We hypothesized that measuring the recruitment of endocytic proteins with improved temporal accuracy might reveal otherwise hidden temporal structure in the CME mechanism, and so we measured the recruitment signatures of a representative set of 34 mammalian endocytic proteins relative to scission.
First, and to illustrate the experimental strategy and details of the analysis, we determined the kinetics of dynamin1 recruitment relative to scission. The Dyn1-mCherry signals acquired at pH 5.5 and pH 7.4 were corrected for bleed through and interlaced, and confidence intervals were calculated on the fluorescence recruitment signature using a randomization procedure (–S4D). As a negative control for protein recruitment we assayed caveolin1-mCherry, which forms spot-like structures at the plasma membrane but which is not enriched at sites of CME [29] (Figure S4E–S4H).
Dynamin is essential for scission [30], and it is thought to be recruited in the last steps of vesicle formation [18],[19]. Cells co-transfected with TfR-phl and Dyn1-mCherry and imaged with TIR-FM microscopy at pH 7.4 showed punctuate patterns that were partially colocalized (Figure 2A), and scission events, localized to patches of TfR-phl marking CCSs (Figure 2B), were frequently (75%) preceded by a transient burst of Dyn1-mCherry (Figure 2B and 2C). Examination of the average fluorescence traces revealed that the TfR7 signal dropped before scission, which might indicate progressive polarization of receptor cargo in the invaginating CCS similar to AP2 [31] (Figure 2D). The average recruitment signatures of Dyn1-mCherry showed a peak 2 to 4 s before vesicle detection (Figure 2D–2F), which corresponded to the time of vesicle creation. Before this transient burst Dyn1-mCherry was, on average, present at low levels on the CCS, as seen in the average and in individual examples, consistent with previous observations [32] (Figure 2B and 2D). Visual inspection revealed that pre-scission recruitment of Dyn1-mCherry manifested as low-amplitude “flickering”, which persisted following scission in non-terminal events, consistent with continued recruitment of Dyn1-mCherry to the remaining portion of CCSs at the plasma membrane (Figure 2E). Strikingly, the temporal spread of Dyn1-mCherry average fluorescence (∼8 s) and peak recruitment around scission (Figure 2D) was much narrower than when CCS disappearance was used as a reference for CCV creation (∼20 s) [18],[19]. Finally, the recruitment kinetics of Dyn2-mCherry was very similar to that of Dyn1-mCherry (Figure 2G).
Visual inspection revealed heterogeneity among individual Dyn1-mCherry fluorescence traces (Figure 3A and 3B). To explore whether there was any evidence for natural sub-classes of recruitment signature, the full set of Dyn1-mCherry recruitment traces was normalised and overlaid to generate a cloud plot (Figure 3C). The average fluorescence recruitment trace followed the highest data density, and there was no obvious evidence of bifurcations or the presence of “natural” sub-classes of Dyn1-mCherry recruitment traces (Figure 3C). Therefore, the heterogeneity apparent among individual traces was largely unstructured and most likely represented natural noise rather than mechanistic differences between scission events.
To further test the reproducibility of the Dyn1-mCherry average recruitment signature two datasets were generated using either human or mouse Dyn1-mCherry. For human Dyn1-mCherry seven cells were analysed (1,276 events), and for mouse Dyn1-mCherry 21 cells were analysed, arbitrarily divided into two pools of 10 cells (Pool 1, 2,126 events) and 11 cells (Pool 2, 2,622 events). The average recruitment signatures for human Dyn1-mCherry-transfected cells and either pool of mouse Dyn1-mCherry-transfected cells were very similar (correlation coefficient >0.95), with only minor differences in the pre-scission offset (Figure 3D). Therefore, although individual Dyn1-mCherry fluorescence recruitment traces were variable, the average Dyn1-mCherry recruitment signatures were reproducible and remarkably stable.
Next, we applied the ppH protocol and analysis to an additional set of 33 mammalian endocytic proteins fused to mCherry (Figure S5). To generate an overview of the molecular dynamics of CME we chose a range of proteins that included well-established players (e.g., dynamin and GAK), proteins with tentative or poorly understood links to CME (e.g., Eps8 [33]), and proteins with established links to endocytosis in yeast and which we hypothesized should be recruited to sites of endocytosis in mammalian cells (e.g., cofilin and coronin [34],[35]). Of the 34 endocytic proteins analysed, only the recruitment signature of cortactin had been previously measured with a temporal resolution of 2 s, and the recruitment dynamics of the other 33 proteins remained uncharacterised at this resolution. A reverse transcription PCR (RT-PCR) analysis revealed that all proteins except ACK1, amphiphysin1, CIP4, and FCHo1 were expressed in fibroblasts (Figure S6). It remains possible that the expression of such a diverse set of endocytic proteins is peculiar to cultured cells and would not normally be seen in native tissue. For example, dynamin1 is thought to be expressed predominantly in neurons [36], although low levels of dynamin1 expression have been detected in primary mouse fibroblasts, and the expression level in fibroblast cell lines was found to increase upon immortalization [9]. However, and as described previously [15],[23],[27],[37],[38], we expected that proteins expressed in fibroblasts and heterologously expressed proteins would still incorporate into the CME machinery and could thus reveal useful information.
For each protein we generated red FP (RFP) fusion constructs and assayed 5–7 cells per construct using the ppH protocol, yielding a dataset of ∼1,000 bona fide scission events per protein type (Table S1). Overexpression of mCherry-tagged proteins may perturb the recruitment dynamics of endocytic proteins or have other deleterious effects on the endocytic machinery. Therefore, to ameliorate the possible effects of overexpression cells were transiently co-transfected with TfR-phl and the relevant RFP chimera ∼48 h prior to the experiment, and cells with only the lowest 10%–20% levels of expression used for imaging experiments. In our experience this procedure gave the most consistent results, and target cells showed no overt changes in morphology. Although the incidence rate of scission events varied up to 5-fold between constructs (Table S1), variability between cells expressing the same construct was also high and cells expressing low levels of a selection of RFP fusion proteins still internalized Tfn-A647 (Figure S7). Moreover, and by definition, the ppH assay measured the dynamics of protein recruitment only to successful scission events.
The recruitment signatures of each protein were assessed, and the full set of traces compared pair-wise and organized in a dendrogram by hierarchical clustering (Figure 4; the full set of fluorescence recruitment signatures is shown in Figure S8 and peaks histograms in Figure S9). This analysis revealed, similar to previous results in yeast [24], that natural groups or clusters were formed based on the similarity of recruitment signatures. In each of the seven groups or modules there were proteins expected to show similar recruitment signatures on the basis of previous knowledge (i.e., previous imaging studies, known binding affinities, and known biochemical properties), while some patterns of recruitment were unexpected. A brief comparison of key predictions, based on a priori models, and actual observations follows below.
The clathrin recruitment signature, reported by Clc-mCherry, showed a slow build up that peaked at scission and dropped sharply thereafter, presumably as the newly formed vesicle uncoated (Figure 4A and 4B; although note the different signatures of terminal and non-terminal events, Figure S8). Most similar to clathrin were the PI(4,5)P2-binding epsin N-terminal homology domain (ENTH)/AP180 N-terminal homology domain (ANTH) adaptor proteins epsin and CALM [8],[39], both of which directly bind clathrin. Surprisingly, NECAP, which has a high affinity for the AP2α-ear [40], displayed a similar recruitment profile to clathrin rather than AP2.
Other adaptor proteins formed a distinct subgroup within the clathrin/adaptor protein module. Based on previous work it was predicted that AP2 fluorescence (marked by mu2-mCherry) should markedly decrease before scission, indicating the polarized segregation of AP2 in the nascent bud [31] and/or loss from developing buds before clathrin [41] (though see [42]). This was indeed the case, and, in addition, the adaptor protein Eps15, TfR7 (i.e., the receptor cargo), and the F-BAR domain proteins FCHo1 and FCHo2 showed similar signatures, suggesting that these proteins were also polarized and/or lost from the developing bud before clathrin (Figure 4A, 4B,and 4F). This latter observation may be consistent with a recently proposed role for FCHo proteins in CCP nucleation and the generation of curvature early in bud formation [43].
Dynamin was present at low levels on CCSs at all times, and a burst of recruitment preceded scission (Figures 2, 4A, and 4C). Other proteins showed a similar pattern of biphasic recruitment and thus defined a dynamin module. These included actin-binding proteins such as the actin- and clathrin-binding protein Hip1R [44] as well as the motor protein myosin6, which binds actin and the adaptor protein Dab2 [45]. Other proteins involved in actin dynamics and grouped in the dynamin module included the Arp2/3 activator N-WASP [46],[47], Eps8, an actin capping protein that forms a complex with Abi1 and binds N-WASP [33] and the motor protein myosin1E [48]. The PI(4,5)P2 phosphatase synaptojanin2β1, which binds to the NBAR domain protein amphiphysin1 [49], had recruitment kinetics similar to those of dynamin and peaked at scission but showed little recruitment at time points before −20 s (Figure 4A). Finally, the F-BAR protein syndapin2 [50],[51], which binds dynamin and N-WASP, was recruited early, peaking at −4 s before being quickly discarded following scission (Figure 4A and 4E). The rapid loss of syndapin2 signal may be due to collapse of the highly curved membrane neck at the moment of scission.
The improved temporal accuracy of the ppH assay allowed us to re-evaluate the temporal relationships between dynamin recruitment and actin dynamics. Earlier work suggested that dynamin and actin were recruited sequentially to sites of scission [18]. Here, a more accurate comparison of dynamin and actin recruitment revealed that dynamin and actin recruitment both peaked at scission and that the final burst of dynamin recruitment lagged the onset of actin polymerization by ∼20 s (Figure 4C).
It is generally accepted that actin polymerization plays a role in some (but not all, see [52],[53]) forms of CME [9],[11],[18],[44],[46],[47],[50],[54]. Here, a more accurate measurement of actin dynamics using the ppH assay revealed an ordered sequence of proteins involved in actin dynamics. After the Arp2/3 complex activator N-WASP, which peaked before all the other actin module proteins and groups with the dynamin module, the F-actin-binding proteins Arp3, Abp1, cortactin, and lifeAct were recruited (Figure 4D). Unique among tested proteins, the average lifeAct signal was significantly below random prior to scission (Figure 4A), probably because bright stress fibres adjacent to sites of scission artificially lowered the background subtracted fluorescence value (e.g., see Figure S4).
Peak recruitment of the actin-severing protein cofilin [55] and the Arp2/3 suppressor coronin [56] were both significantly skewed post-scission, suggesting an ordered shut down of the actin polymerization machinery and disassembly of scission-associated actin (Figure 4D).
Based on contemporary models of CME [57] we predicted that recruitment of BAR and F-BAR domain proteins should follow patterns consistent with the differing curvatures of their respective membrane-binding domains, since purified proteins induce different degrees of curvature in membrane tubulation assays in vitro and membrane curvature increases as CCSs invaginate [6],[7]. The sequential recruitment of the F-BAR domain protein syndapin2 and a group of NBAR domain proteins (endophilin2, BIN1, and amphiphysin1) followed by scission matched this prediction (Figure 4E). Similar to syndapin2, NBAR proteins were also rapidly discarded following scission, presumably because of the collapse of the highly curved membrane neck at the moment of scission. However, the recruitment of the BAR domain protein SNX9 differed from prediction. SNX9 recruitment began before scission, peaking ∼12 s after scission, similar to coronin and cofilin rather than to its binding partner, dynamin (Figure 4E). Similarly, the recruitment of the F-BAR domain proteins CIP4 and FBP17 also differed from prediction [6] (Figure 4F). Both proteins showed complex recruitment dynamics, with components of recruitment both before and after scission and, strikingly, FBP17 recruitment peaked markedly post-scission, at a time similar to that of GAK (Figure 4F).
It was shown previously that the kinase GAK, which is necessary for CCV uncoating, was recruited shortly after the large GTPase dynamin to sites of CME [15],[16]. Here, we found that GAK recruitment commenced at scission and peaked on average ∼8 s thereafter, as predicted. The recruitment profile of GAK was the same for both terminal and non-terminal scission events (Figure 4G). In the canonical model of CME, bona fide endocytic structures were represented as spot-like CCSs that formed de novo [23]. We therefore analysed a subset of 100 scission events associated with spot-like CCSs that formed de novo and found that the first detected scission event occurred, on average, 93 s following CCS inception (minimum = 20 s) and similar to the 100 s calculated in a previous study [10]. The GAK recruitment signature was again similar (Figure 4G, inset), and therefore, irrespective of the behaviour of the host CCS, the dynamics of the uncoating reaction associated with scission events were comparable.
The recruitment signature of GAK defined a module including ACK1, a serine threonine kinase implicated in tumorigenesis [58] and OCRL1, a 5′ phosphatase and Rab5a effector [59]. Interestingly, GAK and OCRL1 were recruited only after scission, whereas ACK1 was gradually recruited as CCSs matured (Figure 4A). The last module to be recruited consisted of the Rab5a effector APPL1 [60] and Rab5a itself (Figure 4A). The Rab5a signal was small and temporally spread, but significantly raised above baseline. Most likely this marks the outer limits of recruitment detection using the ppH protocol.
Having accurately measured the recruitment signatures of a representative set of endocytic proteins we next asked whether the same set of proteins was recruited to scission events at different dynamic classes of CCSs.
Previous studies defined different populations of CCSs on the basis of size (i.e., spot-like CCSs versus larger CCSs) and lifetime or whether CCSs disappeared following scission (terminal events) or persisted (non-terminal events) [10],[23],[28],[61]. Detailed mechanistic inferences have been based on these types of dynamic classification [23]. Therefore, we explored whether the set of endocytic proteins recruited differed between terminal and non-terminal scission events or between scission events at CCSs of different size or lifetime.
First we analysed whether the same set of proteins was recruited to scission events at terminal and non-terminal scission events. Terminal and non terminal events were sorted by computing the ratio of average FTfR7 before and after scission (see Materials and Methods). For all constructs tested, there was approximately the same number of events in each category (Table S1). For all proteins tested the average fluorescence profiles were strikingly similar between terminal and non-terminal events before scission, with occasional shifts towards higher values for non-terminal events (Figure S8). This strongly suggests that the mechanisms of protein recruitment were the same for both classes of events. By contrast, the recruitment signatures after scission differed markedly for proteins that were significantly recruited at time points well removed from scission such as clathrin module proteins or some dynamin/myosin module proteins. Interestingly, in many recruitment signatures (e.g., Eps15, mu2, myosin6, or CALM), the average fluorescence trace of non-terminal events increased steadily after scission to a maximum around 40 s post-scission, suggesting a characteristic time course of CCS maturation between successive scission events, and similar to findings in a previous study [10].
We established that there was a good correlation between Clc-mCherry fluorescence and TfR7 fluorescence and that, by inference, TfR7 fluorescence could be used to confidently predict the relative size or lifetime of CCSs (Figures 1E, S3, 5A, and 5B). Therefore, to investigate the relationship between CCS size and patterns of protein recruitment, TfR7 patch fluorescence was normalised by cell, and, for each trace, the average fluorescence (FTfR7) was calculated over the time interval −18 s to −10 s relative to scission (Figure 5C). For each cell the FTfR7 values formed a continuous distribution (Figure 5C) that was divided into three equally populated groups representing “small” CCSs (blue fluorescence traces), “medium” CCSs (green fluorescence traces), and “large” CCSs (red fluorescence traces, Figure 5). As expected, when the normalised fluorescence recruitment traces for Clc-mCherry were assigned to CCS size groups 1–3, the group average recruitment signatures were well separated (Figure 5Di). This simply reflected the fact that larger CCSs had more clathrin and confirmed that TfR7 fluorescence could be used to predict CCS size (see also Figure S3). However, and as a control, when Clc-mCherry recruitment traces were randomly assigned to three groups, the average traces for groups 1–3 were almost identical (Figure 5Dii). Thus, we can be confident that Clc-mCherry fluorescence scaled strongly with TfR7 fluorescence, as expected. Similarly, when TfR7 fluorescence traces were assigned to CCS size groups 1–3, the fluorescence signatures were (by definition) well separated (Figure 5Ei), but when TfR5 fluorescence traces were assigned to CCS size groups 1–3 and averaged, the TfR5 class averages were virtually identical (Figure 5Eii). Therefore (and similar to Figure S3), the amount of TfR internalized did not scale strongly with the size of the host CCS, consistent with the idea that quantized scission events occurred at CCSs of apparently different sizes.
The analysis was repeated for the 34 endocytic proteins of this study to assess how different recruitment signatures scaled with CCS size. Sample classified recruitment signatures are shown in Figure 5F, and, for each protein, the relative strength of the scaling relationship between CCS size and protein recruitment was visualised by calculating the summed absolute difference between the group averages and overall average (Figure 5G, 95% bootstrapped confidence interval in grey). Thus, for example, the average FCHo1 fluorescence traces for the small and large groups of CCSs (Figure 5G, blue and red, respectively) were well separated from the pooled FCHo1 average fluorescence trace, indicating a strong scaling relationship between CCS size and the amount of FCHo1 at the CCS. The scaling relationship was significant because it exceeded the boundaries of the confidence interval in grey (Figure 5G).
In general, the group averages for structural proteins such as clathrin and the adaptor proteins (mu2, Eps15, and FCHo1/2) scaled strongly with CCS size. The relationship between Dyn1-mCherry recruitment and TfR7 cluster size was more complex. As noted earlier, low amplitude flickering of Dyn1-mCherry was noted at CCSs before the final recruitment burst that marked scission (Figures 2B, 2C, 3A, and 3B). The overall amplitude of Dyn1-mCherry recruitment did scale with CCS size, but this could be explained by the difference in offset of the “pre-scission” signal, consistent with two components to the Dyn1-mCherry signal: pre-scission recruitment scaled with CCS size, suggesting a link with clathrin in the host CCS, but the burst of dynamin associated with scission was of relatively constant amplitude, consistent with recruitment to budding structure of constant dimensions. Other transiently recruited proteins, such as endophilin2, showed similar behaviour (Figure 5F). A notable exception was synaptojanin2β1, which showed robust recruitment to large CCSs but lower amplitude recruitment to smaller CCSs (Figure 5F). Finally, the amplitudes of Arp3 and lifeAct recruitment signatures were independent of CCS size (Figure 5F and 5G). In general, proteins of the actin module were among the proteins least dependent on the size of the host CCS.
A second characteristic that has been used to define dynamic groups of CCSs is lifetime [27],[28]. To test whether TfR7 patches could be used as indicators of CCS lifetime, TfR7 patches from cells expressing Clc-mCherry were tracked, and the set of 11,091 track histories was classified according to the presence or absence of scission events. The estimate of scission-undetected TfR7 patch lifetime was 33.8 s, which was ∼11% lower than the 38 s estimated using Clc-mCherry as a marker for CCSs. This slightly lower lifetime is because the TfR-phl signal tended to drop slightly before the Clc-mCherry signal in the run-up to scission (Figures 2D and 4B). The estimate of scission-detected TfR7 patch lifetime was found to be 178 s, which is within 6% of the 189 s estimated using Clc-mRFP as a CCS marker. Therefore, TfR7 patch lifetime could be used to estimate CCS lifetime.
Similar to CCS size, scaling relationships were found between CCS lifetime and the relative amount of protein recruited (Figure 6A–6E). Longer lived CCSs tended to have more clathrin and adaptor proteins while, by contrast, GAK and lifeAct showed the weakest dependence on CCS lifetime (Figure 6E and 6F). This is trivially explained if larger CCSs tended to have longer lifetimes, and indeed TfR7 patch fluorescence and lifetime had a positive (though modest) correlation of 0.29 (p<0.05, full set of events used), similar to previous observations [62].
Collectively, these analyses demonstrate that the same set of proteins was recruited to scission events at different dynamic groups of CCSs, with subtle scaling relationships between CCS size, lifetime, and the relative amount of different proteins recruited. However, this analysis did not reveal whether the same set of proteins was recruited to each scission event.
The physical properties of CCSs were not predictive of which endocytic proteins were recruited to scission events (Figures 5 and 6). However, there is evidence that CCSs with different complements of adaptor proteins and receptor cargo coexist in the same cell [63], and it has been shown that the dependence of CME on actin differs between the apical and basolateral domains in epithelial cells [53]. Therefore, there may be differences in the set of proteins recruited to individual scission events, even though they internalized similar amounts of the same cargo (TfnR-phl).
First, we checked whether the automated selection criteria were biased towards a mechanistically distinct subtype of CME. For five example cells expressing mCherry chimeras of Clc, Hip1R, N-WASP, dynamin1, or GAK we visually inspected the set of events rejected by our selection criteria and “recalled” events judged to be bona fide by a human operator (see Materials and Methods; Figure S10). There was no significant difference in the kinetics of protein recruitment to the subset of “recalled” events when compared to events automatically selected (Figure S10). Therefore, no measurable bias was introduced by the parameters set for automatic detection.
Second, we determined how many scission events scored positive for recruitment of any given protein (Figure 7). The probability of detecting protein recruitment is dependent on multiple physical factors including signal and detector limitations, the kinetics of protein recruitment, and the magnitude and texture of background fluorescence (see Materials and Methods). Two strategies were used to detect recruitment (Figure 7). The first strategy was biased towards the detection of proteins recruited with slower kinetics and used image segmentation to determine the maximum probability of detection relative to scission (Figure 7A–7C). The second strategy was biased towards the detection of more transient signals and identified significant peaks in the quantified fluorescence traces (Figure 7D and 7E).
Of the 34 proteins analysed 25 proteins from six modules (clathrin, actin, dynamin, GAK, FBP17, and Rab5 modules) were detected at more than 50% of scission events using either detection strategy (Figure 7). It seems unlikely that these 25 proteins were recruited to distinct and mutually exclusive variants of CME, and there was most probably some overlap between any given pair. Of these proteins high-abundance structural proteins such as clathrin, adaptor proteins, and other members of the clathrin module were most readily detected (gold bars, Figure 7C and 7E). Proteins of the dynamin module were the next most frequently detected (pale blue bars, Figure 7C and 7E). Proteins of the actin module (red bars, Figure 7C and 7E) were detected less frequently, with the notable exception of Abp1 (maximum probability of detection = 0.97, Figure 7C). The clathrin- and F-actin-binding protein Hip1R was also detected with high frequency (maximum probability of detection = 0.99, Figure 7C). Detection of the F-actin-binding protein Abp1 was facilitated by the proteins' punctate distribution and low background fluorescence (Figure 7F). By contrast, an alternative F-actin marker, lifeAct, was recruited promiscuously to all F-actin structures at the cell cortex, which gave a bright and highly textured background, and most likely contributed to the lower probability of detecting lifeAct at scission events (Figure 7G).
The set of proteins that were detected at ∼50% of scission events or fewer using either detection strategy included the NBAR module (BIN1, Endo2, and Amph1, pink bars, Figure 7C and 7E). However, because NBAR proteins are thought to be essential components of the CME machinery [9],[54], this most likely represents limitations of detection, as found previously [37], rather than core mechanistic differences between scission events. The low incidence of detection of other proteins is less easy to interpret. For instance, CIP4 and Rab5 were detected with low incidence, but the significance of this currently remains unclear (Figure 7).
Early EM studies revealed clathrin-coated invaginations at the substrate proximal surface of adherent cells as discrete entities, in clusters or at the edges of large, flat lattices of clathrin [25],[26]. Subsequent live-cell imaging studies using TIR-FM described, for a variety of cell types, corresponding heterogeneity among CCSs labelled with clathrin-FP at the substrate proximal surface of adherent cells [3],[10],[23],[28],[31],[62]. It was shown that both transient spot-like CCSs (average lifetime = ∼40–60 s) and larger, longer lived CCSs (average lifetime = ∼60 s to 10 min [or more]) coexisted in NIH-3T3 fibroblasts, HeLa, and COS cells [10],[23], while transient spot-like CCSs (average lifetime = ∼40 s) predominated in freshly plated BSC1 cells [3],[28],[31],[62]. Larger and longer lived CCSs were triggered by specific receptor/adaptor combinations [62], and cell adhesion could also play a role [22]. Faced with such natural ultrastructural and dynamic heterogeneity, it was important to establish which CCS characteristics, measured in live-cell TIR-FM experiments, defined CCS intermediates in CME. The detection of individual scission events presented here and previously [10] helps achieve this by quantifying the relationships between scission and CCS dynamics and size in an unbiased manner.
We can make four main conclusions from our study of CCS characteristics relative to scission. First, the lifetimes of scission-detected CCSs followed a left-skewed distribution ranging from a few tens of seconds through to hundreds of seconds, as predicted by earlier studies [3],[28]. The shorter lived population of scission-undetected CCSs identified most likely corresponded to abortive CCSs described previously [3],[28], although intracellular CCSs may have contributed. The average time between CCS inception and the first detected scission event was ∼100 s (minimum lifetime of 20 s), which reflected the time required to construct a productive CCS. However, CCS lifetimes should be interpreted with caution since CCSs can host multiple scission events (see also [10] and the third point below).
Second, the size of scission-detected CCSs followed a left-skewed distribution without obvious quantization. No correlation was detected between overall CCS size and the amount of TfR-phl cargo internalized by scission events, consistent with an earlier study [10].
Third, the disappearance of spot-like CCSs, which has been widely used as a fiducial marker for CME [18],[19], coincided with scission events with the predicted frequency but it was found to be an imprecise marker for scission (Δt between scission and spot-like CCS disappearance = 7±22 s; Figures 1 and S2). Moreover, CCS disappearance did not report all scission events, and approximately ∼50% of scission events were classified as non-terminal because the host CCS did not completely disappear following scission. Indeed, CCSs could host multiple scission events before disappearing (see also [10]).
Fourth, evidence that the scission events detected at different dynamic groups of CCSs proceeded through to completion (i.e., CCV uncoating) was provided by the remarkable invariance of the GAK recruitment signature. The kinase GAK is an established marker for CCV uncoating [15],[16], and the GAK recruitment signature was the same for terminal and non-terminal scission events, for scission events at spot-like CCSs that formed de novo, and for scission events at different size and lifetime classes of CCSs.
The most parsimonious explanation for these findings is that CCVs, of similar size, could either bud in isolation or from larger, heterogeneous CCSs (Figure 8). This is consistent with the relatively constant dimensions of clathrin-coated invaginations previously observed by EM, irrespective of whether the invaginations were isolated or part of larger CCSs [25],[26]. Based on these results we conclude that the classification of endocytically active CCSs, observed at optical resolution using TIR-FM, should be broad to encompass the heterogeneity of scission-competent CCS sizes and lifetimes. As a practical guide, any CCS that colocalizes with acid-accessible TfR-phl and that exists for more than 20 s could be considered scission competent [10] and potentially capable of hosting multiple scission events.
In a further exploration of the organization of CME in fibroblasts we analysed the recruitment of 34 types of endocytic protein to scission events, 30 of which were native to NIH-3T3 fibroblasts. To appreciate this analysis properly it is important to consider what physical factors contribute to the observed dynamics of protein recruitment and the resulting shapes of ensemble recruitment signatures. First, the fluorescence signals measured at single scission events using TIR-FM occur in a volume of ∼1 al, illuminated by an evanescent field in which the intensity of the electromagnetic field decreases exponentially as a function of distance in the z-axis [64]. Due to the small depth constant of the illuminating evanescent field (∼100 nm) and the comparable dimension of an invaginating CCP (∼100 nm diameter), for two proteins to show a similar average recruitment signature they must be recruited to the detection volume over a similar time course and must share a similar spatial distribution at the developing CCP as it projects into the evanescent field along the z-axis [10],[31]. Second, a recruitment signature reflects the average concentration of an FP-labelled protein at the site of endocytosis relative to the cytoplasm. Labelled protein, expressed at low levels, must compete with endogenous proteins for recruitment, and this, combined with detector limitations and the relatively low quantum efficiency of mCherry [65], most likely contributes to noise among individual recruitment profiles and influences the probability of detecting protein recruitment. We established that for one example protein, dynamin1, the noise appeared to be unstructured and that the trajectories of the averaged recruitment signatures for dynamin1 in NIH-3T3 cells were remarkably stable.
A detailed analysis suggested involvement of the core clathrin, actin, and dynamin modules in the majority of scission events since all coat components (clathrin, AP2, epsin2, FCHo, CALM, and NECAP) and both Hip1R (which binds clathrin and F-actin [44]) and Abp1 (which binds dynamin, F-actin, and Arp2/3 [66]) were detected at >90% of scission events, and dynamin1/2, synaptojanin2β1, myosin6, and Eps15 were detected at >75% of events. These findings agree with the widely accepted view that TfR internalizes via a clathrin- and dynamin-dependent pathway [67],[68] and are in agreement with earlier studies that demonstrated an important, though nonessential, role for actin in CME in fibroblasts ([9]–[11], but see [23]). The fact that other proteins such as the BAR domain proteins endophilin2 or BIN1 were detected at only a subset of scission events suggests that there were inherent limitations of recruitment detection, since these proteins are thought to be essential for scission [9],[54]. However, it remains possible that there were genuine molecular differences between scission events, perhaps through the influence of other types of (unlabelled) receptor cargo [63], in response to changes in physical parameters, such as membrane tension [69] or because of genuine underlying variability in the core mechanism of CME [53]. Nonetheless, and based on the data presented here, at optical resolution potential molecular differences between scission events in NIH-3T3 cells did not correlate with obvious differences in CCS behaviour.
Next we explored scaling relationships between CCS size and lifetime and the set of proteins recruited to scission events. As shown previously [62], CCS lifetime and size were moderately correlated, with longer lifetimes for larger CCSs, and, as predicted, the recruitment signatures of some proteins such as the coat protein clathrin and adaptor proteins scaled with overall CCS size. A set of core components (e.g., dynamin and endophilin2) showed more complex scaling relationships with CCS size, perhaps reflecting variable degrees of recruitment to the budding and non-budding portions of larger CCSs. However, the recruitment signatures of a core set of proteins including GAK (a kinase essential for the uncoating reaction [16],[70]), and most notably actin and actin-binding proteins, were independent of CCS size. This is consistent with our central thesis that CCVs of relatively constant size budded at host CCSs of diverse size and lifetime via a common core mechanism, and supports a role for actin in CME in NIH-3T3 fibroblasts [10],[11].
Seminal imaging studies from the Drubin lab and other groups revealed the modular organization of yeast endocytosis [24]. Here it was shown that at least four modules or groups of proteins showed similar recruitment dynamics to sites of endocytosis at yeast actin patches [24]. More recently, comparisons were drawn between the modular organization of yeast and mammalian endocytosis, with an emphasis on the conserved role of actin [9],[71]. However, earlier TIR-FM studies of the late stages of mammalian CME used the disappearance of spot-like CCSs as a fiducial marker, which could not sample endocytic events from all dynamic classes of CCSs nor yield a temporally precise estimate of scission. Consequently, the recruitment dynamics of endocytic proteins could only be broadly classified as “early” and “late” (Figure S1). The data presented here, based on the comparison of accurately measured recruitment signatures derived from large datasets (∼1,000 events), give a more detailed overview of the modular organization of mammalian CME. The modules identified here comprise the following (Figure 8): (1) the coat module, divided into (i) a clathrin sub-module (epsin2, CALM, clathrin light chain, and NECAP) and (ii) an adaptor/F-BAR sub-module (FCHo1/2, Eps15, AP2); (2) the NBAR domain module (endophilin2, amphiphysin2, and BIN1); (3) the actin module, divided into (i) actin polymerization sub-module (Abp1, cortactin, and Arp3) and (ii) actin depolymerization/suppression (cofilin, coronin1B, and SNX9); (4) the dynamin/myosin/N-WASP module (dynamin1, dynamin2, synaptojanin2β1, myosin1E, N-WASP, Eps8, Hip1R, myosin6, and syndapin2); (5) the GAK/post-scission module (GAK, ACK1, and OCRL1); (6) the Rab5a module (Rab5a and APPL1); and (7) the FBP17/CIP4 module, based on the unique recruitment signatures of these two proteins and dissimilarity to any other recruitment signatures.
The shapes and relative timing of many of the recruitment signatures are broadly consistent with measurements made in previous imaging studies in yeast [24] and in mammalian cells [6],[18],[19]. In addition, many recruitment signatures provided new information as a consequence of improved accuracy. First, and as predicted from a previous study [31], the recruitment signatures of members of the adaptor sub-module decreased before scission because of polarization in the developing invagination. In addition, the F-BAR domain proteins FCHo1 and FCHo2 showed similar recruitment signatures, suggesting these curvature-inducing proteins were also polarized and consistent with a proposed role for FCHo proteins in the early stages of the invagination process [43]. Second, it was predicted that actin recruitment should begin before dynamin recruitment at sites of scission, although time-locked measurements with the required accuracy to test this hypothesis had not previously been made [9],[18]. Here, we showed that the onset of actin polymerization did indeed precede the final burst of dynamin recruitment by ∼20 s, consistent with a role for actin polymerization early in the invagination stage of CME and the later recruitment of dynamin to the deeply invaginated CCP, where it executed scission [9] (Figure 8). We also discovered that coronin1B and cofilin, proteins involved in the down-regulation of actin polymerization and F-actin severing, respectively, were recruited at later time points, again similar to yeast endocytosis [24],[72],[73]. Third, it was proposed that scission of endocytic invaginations in yeast is triggered by a PI-phosphatase that dephosphorylates PiP2 and thus induces a line tension in the membrane neck [74]. In mammalian cells the large GTPase dynamin is thought to execute scission [9],[30], but, intriguingly, the recruitment of the PI-phosphatase synaptojaninβ1 showed a recruitment trajectory similar to that of dynamin (and proteins of the NBAR module) and also peaked at scission (Figure 4). Therefore, it is plausible that induction of a line tension also contributes to the mechanochemistry of scission in mammalian cells [74]. Finally, it was predicted that recruitment of F-BAR and BAR domain proteins should follow an ordered sequence dictated by their preference for different-curvature membrane tubules in vitro [75] and that recruitment should occur over a trajectory similar to that of actin polymerization [6],[9],[76]. The ordered recruitment of syndapin2 and the NBAR module (endophilin2, BIN1, and amphiphysin1) did indeed match this prediction. However the post-scission peak recruitment of SNX9 and the complex, biphasic recruitment of FBP17 and CIP4 did not. These findings illustrate that the recruitment sequence of these BAR and F-BAR domain proteins could not be predicted purely on the basis of either structural information or biochemical properties. The possible function(s) of SNX9 and FBP17/CIP4 post-scission remain to be elucidated, although it is possible that these proteins may act as relays to recruit additional binding partners to the newly formed endosome (Figure 8).
The study presented here employed the detection of scission events to construct what is to our knowledge the highest resolution temporal map of mammalian CME to date. The map (1) suggests a simplified canonical model of mammalian CME in which the same core mechanism operates at both spot-like CCSs and larger CCSs observed with fluorescence microscopy, (2) illustrates the similar modular organization of mammalian and yeast endocytosis, and (3) proves that recruitment dynamics of endocytic proteins such as the F-BAR protein FBP17 and BAR domain protein SNX9 cannot always be predicted from biochemical or structural properties.
NIH-3T3 cells were cultured as described previously [10]. Cells were co-transfected using Lipofectamine 2000 (Invitrogen) with human transferrin receptor fused to super-ecliptic phluorin (hTfnR-phl [10]) and the relevant endocytic protein open reading frame (ORF) fused to a RFP. Freshly transfected cells were replated onto pre-cleaned number 1 borosilicate glass coverslips (VWR International) and imaged 24–48 h later as described previously [10].
ORFs of endocytic proteins were amplified by PCR (Phusion PCR kit; Finnzyme) from IMAGE clones (Geneservice), or directly amplified from cDNA libraries (see Table S2 for details of primers and cDNA sources for the expression constructs used). Each pair of PCR primers was engineered with the appropriate 3′ and 5′ restriction sites for cloning and sequence for either a 9-, 12-, or 13-amino-acid linker between the target protein and FP, as described previously [65]. The amplified cDNAs were cloned into mammalian expression vectors in frame with a RFP (in the case of Hip1R, tDimer [65]; in the case of myosin1E, mApple [77]; and for all other proteins mCherry [78]; see Table S2) to generate either N- or C-terminal fusion proteins upon expression.
Primers were designed to PCR a ∼700-bp fragment that was specific to the protein isoforms used in this study. Total cell RNA was purified from NIH-3T3 cells using the RNAeasy Mini Kit (Qiagen). RT-PCR reactions were run using the OneStep RT-PCR kit from (Qiagen) using the manufacturer's protocol. The QIAxcel capillary gel electrophoresis system (Qiagen) was used to visualise RT-PCR products. Samples were run using the DNA screening cartridge using the AM420 run settings (5 kV sample injection voltage for 10 s, 5 kV separation voltage for 420 s; suitable for DNA concentrations of 10–100 ng/µl). A photomultiplier detector converted the emission signal into a gel image and an electropherogram that allows visualisation and quantification, respectively, of each PCR product. The Biocalculator software package (Qiagen) was used to analyse the peaks for each sample. Aligment marker of 50 bp to 1.5 kb was used to align run samples.
The TIR-FM and ppH perfusion system have been described previously [10].
Cells were transfected with plasmid encoding TfR-phl and RFP-tagged endocytic protein and plated onto coverslips 24 h before imaging. Transfected cells were located and imaged using a spinning disk UltraVIEW ERS confocal (PerkinElmer) using a ×40/1.4 NA oil immersion PlanApo objective (Olympus). After acquiring an initial image (denoted t = 0 min) transferrin conjugated to Alexa 647 (Tfn-A647; Invitrogen) was added to the chamber at 10 mg/ml in 10 mM HEPES buffer saline solution (pH 7.4). After 30 min at room temperature the cells were washed twice in 10 mM MES (pH 4.0) to strip away surface-bound Tfn-A647 and returned to HEPES buffer saline (pH 7.4). The cells were then imaged to determine uptake of transferrin (image denoted as t = 30 min).
Movies of cells during the alternate pH protocol were divided in four parts, TfR-phl at pH 5.5 (TfR5 movie) and at pH 7.4 (TfR7 movie), and the RFP fusion protein at the two pH values (movies RFP5 and RFP7). To detect protein clusters (CCSs or TfR7 clusters as in Figure S4, CCVs in the TfR5 movies) images were subjected to segmentation based on wavelet transform (Multidimensional Image Analysis [MIA] add-on to Metamorph 6, written by V. Racine and J.-B. Sibarita, Curie Institute, Paris, France). The objects detected were then tracked using a simulated annealing algorithm [79] to identify endocytic events. The output of this tracking was a series of coordinates corresponding to the centre of mass of the objects, with unique identifiers (event numbers).
To determine the lifetimes of CCSs using either Clc7 or TfR7, a different tracking algorithm was used to account for transient breaks in track histories of 1–2 frames (i.e., gap closing was incorporated). The coordinate lists generated by MIA were reassigned in Matlab using a nearest-neighbour algorithm (“track.pro”, John C. Crocker and Eric R.Weeks, http://www.physics.emory.edu/~weeks/idl/index.html). For Clc data, independent track histories generated by MIA from Clc7 and Clc5 data were combined and reassigned, while for TfR only the TfR7 data were used. To verify tracking fidelity the reassigned tracks were overlaid on the original image series in Matlab and inspected visually. Although the tracking was perhaps not as robust as more recently published techniques [27], it was sufficiently robust to differentiate between long-lived CCSs and shorter lived CCSs (Figures 1 and 6).
All the tracked objects in the TfR5 movies were screened to identify genuine endocytic events using routines programmed in Matlab 7.4 (Mathworks). To qualify as bona fide events each candidate event required a TfR5 vesicle (i) that persisted for at least three frames (i.e., 8 s) following appearance, (ii) that appeared at least 20 frames after the start of the movie, or 20 frames before its end, to ensure quantification of signals for 80 s before and after the vesicle's appearance, (iii) that appeared and remained at more than seven pixels (0.7 µm) from the edge of the image, to ensure proper quantification (see below), (iv) that appeared de novo, and was not produced by the fusion of two objects or the dissociation of an object into two, (v) that overlapped, on appearance, with a pre-existing cluster detected in the segmented TfR7 movie, (vi) whose fluorescence was bigger than a defined SNR of 5 wherein SNR = (F0−av)/std, where F0 is the fluorescence at time 0, and av and std are the average fluorescence and standard deviation, respectively, in the five frames before vesicle appearance, and (vii) that was close to maximal fluorescence at the time of appearance. We calculated the slope of the fluorescence change in the first three frames of vesicle appearance (Figure 1) and discarded the events where this slope was greater than 0.1, which corresponds to a 10% increase in fluorescence.
The purpose of this screening was not to detect all events in a recording, but to have stringent criteria to select automatically a large proportion of events that were genuine scission events, to test a large number of candidate proteins in a manageable analysis time. Among the events that occurred at suitable times and locations (criteria i–iv), only 18.5%±0.8% of events (n = 191 cells) passed the last three criteria (v–vii), for a total of 239±11 candidate events per cell. Nevertheless, some false-positive events remained, so we reviewed our dataset visually by watching each event individually (the portion of the TfR5 and RFP5 movies around the 0 frame, and an average of five frames of movie TfR7 before the event) to assess if there were tracking errors, poor signal, simultaneous events nearby, or other problems. On average, 82.3%±0.9% (n = 191 cells) of events were confirmed by this second, manual screen, for a total of 191±8 confirmed scission events per cell.
To check for bias in the screening procedure we performed a visual screen on all tracked objects for five cells, each transfected with different mCherry-tagged proteins (1,400±360 tracked objects per cell). Of the events rejected by the automated screen (1,100±318 objects), a total of 10.3%±2.2% were visually identified as genuine scission events (104±24 events). Importantly, when the fluorescence from these “recalled” events was quantified and averaged, the RFP recruitment signatures were the same as the signatures obtained from “semi-automatically selected” events (Figure S10). The sum of absolute differences between average fluorescence traces of semi-automatically selected and recalled events was not statistically significant. This shows that our semi-automated procedure did not select a particular category of scission events.
Images in the green fluorescent protein (GFP) and RFP channels were acquired simultaneously with a Dual View (Optical Insights) beam-splitter that was adjusted with an image of beads that fluoresce in the two channels (yellow fluorescein carboxylate beads, 0.2-µm diameter, Invitrogen) to minimize distortion from one channel to another. However, small (0–5 pixels) shifts remained in the two channels that needed to be corrected digitally for optimal colocalization. We used a third-order polynomial spatial transform that interpolates between ten bead pairs to make the correction. When we quantified experimental data we did not transform the raw images (i.e., interpolate and reassign pixel fluorescence values) but instead used the spatial transform to recalculate the vesicle centre coordinates in the RFP channel. This works well, since the difference between the coordinates of a pixel (x,y) in the green channel and its transformed coordinates (u,v) in the red channel is only ever a fraction of a pixel.
We quantified the fluorescence 20 frames before and 20 frames after the appearance of a vesicle for all four movies in a three-pixel-radius circle centred on the object coordinates at the time of appearance (frame 0) for this frame and the 20 preceding frames, then centred at the tracked vesicle coordinates during tracking, and then on the last known coordinates after tracking was lost. Local background was estimated in an annulus (three pixels inner radius, six pixels outer radius) by taking an average of pixel values between the 20th and 80th percentiles to avoid contributions from neighbouring brightly fluorescent patches. This quantification is similar up to this point to other quantifications performed by us in previous studies [10],[18].
To correct for bleed through from the GFP to RFP channels we introduced a bleed-through coefficient (BT) for each cell to correct the fluorescence values with the formula FRFPx,corr = FRFPx−BT·FTfRx, where x is 5 or 7. Such corrections are acceptable as they involve only linear combinations of fluorescence values. BT was determined for each cell by minimizing the summed squared difference for values of BT taken between 0 and 0.05 in 0.001 increments (Figure S4). Values of BT were on average 3.00%±0.07% (n = 191). Differences in BT values could arise from small differences in background fluorescence, non-linearity in the camera, or changes over months of the optical properties of the various parts of the system (filters, mirrors, or camera). With this correction, fluorescence values from RFP5 and RFP7 could be combined to achieve a time resolution of 2 s (Figure S4).
To determine when the recruitment of a labelled protein became significant, we generated randomized datasets by shifting the event coordinates in a random manner within the cell footprint (Figure S4), and calculated fluorescence for all four movies as described above. We generated 200 randomized datasets for each cell, and then combined the average fluorescence measures to determine, for each data point, 95% upper and lower intervals (Figure S4).
To sort events into terminal and non-terminal events, we measured the average FTfR7 for four frames before scission and nine frames (36 s) after scission. The ratio between these two values was used to determine whether the event was terminal (ratio <0.4) or non-terminal (ratio >0.6). Events with ratios close to 0.5 were not sorted. To determine the time of peak RFP recruitment, we estimated a noise level with standard deviation of the last six FRFP values (12 s) of the recording. If the maximum is bigger than a threshold (six times noise above average), the time of the maximum FRFP value is taken as the maximum RFP recruitment time and used to construct the histograms in Figure 1F and others. The proportion of events with significant peak recruitment is given in Table S2 for each tested protein.
The goal was to visualise the overall structure of the Dyn1-mCherry set of fluorescence recruitment traces and determine whether there were “natural” (as opposed to analyst-imposed) classes. First the amplitudes of fluorescence traces were normalised by cell over the range [0,1], and the mean of each fluorescence trace was subtracted to reduce dispersion in the y-axis. Each normalised, offset fluorescence trace was projected into an image matrix, and at those points where the fluorescence trace overlaid a pixel a “1” was added to the pixel value, “0” otherwise. The resulting density plot was log-transformed to visualise both high- and low-density features.
We compared the average recruitment signatures by computing the correlation coefficients for each pair of curves corr(RFPa,RFPb). Correlation coefficients were 0.45±0.43 (average ± standard deviation, n = 561). We then used the correlation distance, dist(RFPa,RFPb) = 1−corr(RFPa,RFPb), to perform a hierarchical clustering using an average linkage algorithm that generated the dendrogram in Figure 4. This hierarchical cluster tree reflected the actual correlations between RFP curves, with a correlation coefficient between the cophenetic distances (the distances represented as horizontal bars in the tree) and the correlation distances of 0.81. Other linkage algorithms yielded lower correlation coefficients.
To perform these comparisons, we used the full range of measurements, from −82 s to +76 s relative to the time of vesicle detection. Away from time 0, the differences between the curves would be less significant than close to the moment of vesicle formation and so similarity measurements could be affected by the choice of time interval around vesicle creation. We performed the same clustering procedure using RFP measures only between −44 and +36 s relative to vesicle formation. The cluster tree generated was very similar to the one shown in Figure 4. There were only three minor differences between these two trees: (i) N-WASP grouped first with syndapin instead of with Eps8, (ii) dynamin2 grouped first with dynamin1 instead of with Hip1R, and (iii) coronin grouped first with Arp3 and cortactin instead of with SNX9 and cofilin.
Finally, for many proteins the non-terminal fluorescence traces showed little variation before and after scission (Figure S8, see definition below of these two types of events). The clustering could be different in an analysis using only the terminal fluorescence traces, wherein most proteins reach random values 80 s after scission. Therefore, we performed the clustering on non-terminal events only. Again, the resulting dendrogram was very similar to the one shown in Figure 4, with the same number of modules defined by a distance threshold of 0.2, and only minor differences: (i) ACK1 leaves the GAK cluster to be weakly (distance >0.2) attached to the dynamin cluster, (ii) endophilin groups first with syndapin within the dynamin cluster, and (iii) four other different groupings occur between proteins within the same cluster.
Overall, these tests suggest that the clusters defined in Figure 2 correspond to genuine similarities between the different RFP recruitment signatures that would correspond to functional units involved in CCV formation.
To explore the relationship between scission and CCS disappearance NIH-3T3 cells were transfected with Clc-mCherry and TfR-phl and assayed using the ppH protocol. All CCSs were tracked as described above. The end of each track history was extended by 20 frames (40 s) by padding with the last detected CCS location, and the Clc-mCherry fluorescence and TfR5 fluorescence were quantified for each candidate CCS. To identify CCS disappearance, abrupt drops in Clc-mCherry fluorescence were detected by convolving each Clc-mCherry fluorescence trace with a one-dimensional kernel appropriately tuned for negative edge detection (a negative step function kernel, convolved with a Gaussian, σ = 36 s). Step decreases in Clc-mCherry fluorescence manifest as spikes in the convolved signal, and the maximum response was used to define a t0 for each CCP fluorescence trace. By definition, this algorithm aligns the Clc-mCherry fluorescence traces to their respective maximal negative derivatives (i.e., maximal rate of fluorescence decrease). Although this differs slightly to the algorithm used previously [18], the temporal alignment is more robust. The resulting candidate CCS disappearance events were screened by comparing the average fluorescence of the first nine time points (t = −80 s to t = −44 s) and the last nine time points (t = 44 s to t = 80 s) of the fluorescence trace. Only those traces showing a decrease in average fluorescence with a magnitude at least 2.5-fold greater than the standard deviation of the first nine values were deemed bona fide. This removed false-positive disappearance events (i.e., abrupt but incomplete drops in Clc-mCherry fluorescence). To detect scission events associated with disappearing CCSs the TfR5 trace associated with each candidate CCP was screened for step increases in fluorescence of at least 25 fluorescence units between a given time point t and the average fluorescence over of the previous four time points. This is a less stringent criterion for detecting scission events than used in the main analysis but it was less prone to discarding dim or noisy scission events. Of 197 disappearing CCSs, 107 (54%) were associated with a scission event (Figure 1), close to the prediction that50% of events would be detected when the cell is bathed in pH 7.4 solution, the other 50% being invisible as they occur when the cell is under pH 5.5 solution.
NIH-3T3 cells expressing hTfR-SEpHl were isolated by FACS 24 h post-transfection, replated, and allowed to adhere overnight. To examine potential effects of acidic buffer on CCS morphology NIH-3T3 cells were incubated with MES buffered saline (pH 5.5) for 1 or 10 min before being washed briefly in PBS and fixed at room temperature in a solution of paraformaldehyde (2%) and glutaraldehyde (2.5%) in sodium cacodylate (0.1 M at pH 7.2). Fixed cells were harvested by scraping and centrifuged in a horizontal rotor (1,000 g, 5 min). The resulting cell pellet was placed in fresh fixative and stored at 4°C. In preparation for EM, fixed samples were washed thoroughly in sodium cacodylate buffer (0.1 M), post-fixed in OsO4 (1% in 0.1 M sodium cacodylate) for 1 h, and then washed with distilled water. Samples were stained en bloc with uranyl acetate (2%) in ethanol (30%) before dehydration in a graded ethanol series followed by 1,2-epoxypropane (propylene oxide) and then infiltrated and embedded in CY212 resin (Agar Scientific). Ultrathin (50–70 nm) sections were cut on a Reichert Ultracut E microtome and collected on uncoated 200-mesh grids. Sections were post-stained with saturated uranyl acetate before staining with Reynolds lead citrate. Images were acquired using a Philips EM208 microscope, with an operating voltage of 80 kV, and a CCD camera.
Graphics for protein structures were downloaded from the Research Collaboratory for Structural Bioinformatics (RCSB) consortium Protein Data Bank (PDB) website, where the original citations are also listed.
|
10.1371/journal.pgen.1000633 | Regulation of Ubx Expression by Epigenetic Enhancer Silencing in Response to Ubx Levels and Genetic Variation | For gene products that must be present in cells at defined concentrations, expression levels must be tightly controlled to ensure robustness against environmental, genetic, and developmental noise. By studying the regulation of the concentration-sensitive Drosophila melanogaster Hox gene Ultrabithorax (Ubx), we found that Ubx enhancer activities respond to both increases in Ubx levels and genetic background. Large, transient increases in Ubx levels are capable of silencing all enhancer input into Ubx transcription, resulting in the complete silencing of this gene. Small increases in Ubx levels, brought about by duplications of the Ubx locus, cause sporadic silencing of subsets of Ubx enhancers. Ubx enhancer silencing can also be induced by outcrossing laboratory stocks to D. melanogaster strains established from wild flies from around the world. These results suggest that enhancer activities are not rigidly determined, but instead are sensitive to genetic background. Together, these findings suggest that enhancer silencing may be used to maintain gene product levels within the correct range in response to natural genetic variation.
| Gene expression is generally governed by cis-regulatory elements, also called enhancers. For genes whose expression levels must be tightly controlled, enhancer activities must be tightly regulated. In this work, we show that enhancers that control the expression of the Hox gene Ultrabithorax (Ubx) in Drosophila are regulated by a negative autoregulatory feedback mechanism. Negative autoregulation can be triggered by less than a two-fold increase in Ubx levels or by varying the genetic background. Together, these data reveal that enhancer activities are not always hardwired, but instead may be sensitive to genetic and environmental variation and, in some cases, to the amount of gene product they regulate. The finding that enhancers are sensitive to genetic background suggests that the regulation of gene expression is more plastic than previously thought and has important implications for how transcription is controlled in vivo.
| The transcriptional control of gene expression in eukaryotes is governed by cis-regulatory elements, also known as enhancers, that integrate cell-type and temporal information by binding combinations of transcription factors. Genes that exhibit complex expression patterns are typically controlled by multiple cis-regulatory elements, some of which have overlapping, partially redundant activities [1],[2],[3],[4]. Current estimates suggest that from 10 to 80% of the non-coding DNA of higher eukaryotes is devoted to gene regulation [5],[6],[7], raising the question of how all of this regulatory information is integrated to generate accurate and stereotyped patterns of gene expression in space and time. A third dimension of gene regulation is quantity, which is especially relevant for genes that must be expressed within a narrow range of levels. One possible solution is that enhancers are precisely tuned to generate the appropriate level of transcription that is required in each cell. However, the precision that this type of mechanism demands seems difficult to achieve and especially vulnerable to genetic, environmental, and developmental noise. An alternative solution is that feedback or other regulatory mechanisms exist that modulate enhancer activities in response to the levels of gene product. Although feedback autoregulation is a well-known motif in transcriptional networks [8], mechanisms that might be used to tune expression levels are not well understood. This problem is particularly challenging for genes that have multiple, partially redundant regulatory inputs.
We have begun to study this problem in the fruit fly, Drosophila melanogaster, by analyzing the mechanisms that control the expression of the Hox gene Ultrabithorax (Ubx) in the haltere–a dorsal appendage on the third thoracic segment (T3) that helps the fly balance during flight [9]. Although Ubx protein is detected in all cells of the developing haltere imaginal disc, its pattern of expression is not uniform [10] (Figure 1A). Subsets of the complex regulatory input into the Ubx locus can be monitored by examining the expression patterns of Ubx enhancer traps, which exhibit different, overlapping subsets of the Ubx expression pattern (Figure 1). Ubx-Gal4lac1, for example, (monitored with UAS-GFP) is expressed uniformly throughout the anterior (A) compartment of the haltere disc, but only in the distal portion of the posterior (P) compartment (Figure 1B). In contrast, Ubx-Gal4LDN is expressed in distal regions (in both the A and P compartments) but is not expressed proximally (Figure 1D).
Somewhat paradoxically, transient ectopic expression of Ubx, induced either by heat shock or Gal4-mediated expression, resulted in Ubx loss-of-function transfomations that can be visualized both in the adult (as haltere to wing transformations; [11]) and in 3rd instar haltere imaginal discs (as groups of cells that showed a reduction or complete loss of Ubx protein) [12] (Figure 2). Thus, a transient pulse of high Ubx protein levels can lead to the complete and heritable silencing of all Ubx expression, implying that Ubx is being silenced by its own gene product.
Transient pulses of ectopic Ubx also resulted in the stable silencing of Ubx enhancer traps, including Ubx-Gal4lac1, Ubx-Gal4M1, Ubx-Gal4LDN, and Ubx-lacZ166 (Figure 2 and Table S1). When the absence of Ubx protein was observed, these cells also had no enhancer trap expression (Figure 2). However, in many cases enhancer trap silencing was observed in cells that had normal Ubx protein levels (Figure 2). In these cases we suggest that only the enhancers captured by the enhancer trap were silenced, and that other, partially redundant, enhancers in the Ubx locus remained active, resulting in an apparently normal pattern of Ubx expression. We also find, consistent with previous results [12], that the patches of Ubx-silenced cells in the haltere are clonal events and that the Polycomb system of epigenetic regulators is required for silencing (Figure S1 and Figure S2).
To obtain initial mechanistic insights into Ubx autoregulatory silencing, we carried out experiments that suggest it requires specific DNA binding by Ubx. For these experiments, we monitored the ability of chimeric Hox proteins to induce haltere-to-wing transformations when expressed via the vg-Gal4 driver. Although the more anterior Hox protein Antennapedia (Antp) was unable to induce Ubx silencing, transient overexpression of Antp-Ubx chimeric proteins revealed that the Ubx homeodomain and adjacent C-terminal sequences were both necessary and together sufficient to induce robust Ubx silencing (Figure 3). These findings suggest that Ubx protein, and not Ubx mRNA, is responsible for the induction of silencing. Further, as both the homeodomain and adjacent sequences are implicated in Ubx specificity and DNA binding [13],[14],[15], these results suggest that Ubx triggers silencing by binding to Ubx-specific cis-regulatory elements. Consistently, the Hox protein Abdominal-A (Abd-A), which is very similar to Ubx in both domains, also induced Ubx silencing when transiently expressed during haltere development (Figure 3).
We next tested whether more subtle increases in Ubx levels could also induce silencing. For these experiments, we monitored the expression of Ubx lacZ or Gal4 enhancer traps in flies that had extra copies of the wild type Ubx locus. Ubx-Gal4lac1 and Ubx-Gal4LDN were silenced in groups of haltere cells of 3x Ubx+ and 4x Ubx+ flies (100% of 4x Ubx+ haltere discs had at least one group of silenced cells) (Figure 4A–4D; Table S1). In these haltere discs, probably because the flies had multiple copies of Ubx+, the pattern of Ubx protein was invariably wild type (Figure 4A, 4B, 4D). Interestingly, the amount of silencing induced by 4 copies of Ubx was significantly decreased when one of these copies encoded a non-functional Ubx protein (the Ubx9–22 allele; data not shown). This result supports the idea that Ubx protein, not Ubx mRNA, is the inducer of silencing in response to extra copies of the Ubx locus.
Ubx-Gal4M1 and Ubx-lacZlac1 responded differently to 4x Ubx+: instead of being silenced in clones, these enhancer traps were no longer expressed in proximal regions of the haltere disc, but distal expression remained unchanged (Figure 4E, 4F). For Ubx-lacZ166, the levels were strongly reduced in 4x Ubx+ flies compared to 2x Ubx+ flies (Table S1). Note, however, that Ubx-lacZ166 can be completely silenced in clones in response to hs-Ubx (Figure S3 and Table S1). Finally, the expression of Ubx-Gal4M3 did not change in the presence of four copies of the Ubx+ locus (Figure 4G and Table S1). Taken together, these results allow us to make three important conclusions. First, silencing is occurring at the level of Ubx enhancers, not entire Ubx alleles, because different Ubx enhancer traps respond in different ways. Second, silencing can be triggered by the presence of only one or two additional Ubx+ loci, suggesting that less than doubling Ubx levels is sufficient to silence some enhancers. Third, although all Ubx enhancers can be silenced by high Ubx levels, lower Ubx levels result in a range of responses that depend on which enhancer trap, and therefore which subset of Ubx enhancers, is being monitored. Thus, we conclude that different Ubx enhancers are sensitive to different levels of Ubx protein. We also generated flies to monitor two different enhancer trap insertions into the Ubx locus (Ubx-lacZ166 and Ubx-Gal4lac1) at the same time. When silencing was triggered by heat shock-induced Ubx, we observed silencing of both enhancer traps, but at different frequencies: Ubx-Gal4lac1 was silenced to a greater extent than Ubx-lacZ166 (Figure S3). This finding provides additional support for the idea that individual enhancer traps, and thus different subsets of Ubx enhancers, respond differently to the same increase in Ubx levels.
The above results show that epigenetic autoregulatory silencing of Ubx enhancers occurs in response to elevated Ubx levels. Interestingly, increasing the dose of Ubx+ results in smaller halteres [16], but this size change does not scale linearly with the number of Ubx+ genes. Haltere size is similar to wild type in flies with 3x Ubx+ or 4x Ubx+, while in flies with 6 copies of Ubx+, haltere size is greatly reduced (Figure 5A and Figure S4A). These results suggest that haltere size is buffered against increasing doses of the Ubx+ gene. A similar buffering can be observed when Ubx protein levels are quantified in haltere discs from animals with different numbers of Ubx+ genes. When one copy of Ubx is inactivated (1x Ubx+), Ubx protein levels are nearly halved (Figure S4A). However, when the Ubx+ complement is doubled (4x Ubx+) or tripled (6x Ubx+) only 39% and 60% increases in Ubx protein levels were detected, respectively (Figure S4A). The less-than-expected increases in Ubx levels seen in Ubx duplications is not because they fail to express wild type levels, as they are sufficient to fully rescue a Ubx null mutation, both phenotypically [17],[18] and with respect to Ubx protein levels (data not shown). Together with the results described above, we suggest that the buffering of Ubx levels and haltere size is due, at least in part, to the epigenetic silencing of Ubx enhancers in response to higher than normal doses of Ubx+.
In wild type animals, we hypothesized that enhancer silencing may be used to ensure uniform Ubx levels in response to naturally occurring genetic variation in the cis- and trans-regulation of Ubx expression. We tested this idea by out-crossing our laboratory Ubx-Gal4lac1 flies to 32 D. melanogaster strains established from wild populations around the world. In our lab stock, less than 5% of haltere discs showed any evidence of Ubx-Gal4lac1 silencing. However, when outcrossed to wild D. melanogaster strains, we frequently observed silencing of Ubx-Gal4lac1 in haltere discs of the F1 generations (Figure 5 and Table S2). Although the frequency of silencing varied between wild stocks, it was consistent for each wild stock in a statistically significant manner (Figure 6). Of the 32 stocks crossed to Ubx-Gal4lac1, 14 resulted in no detectable silencing in the F1 generation, 6 showed weak silencing in the F1 generation, and 12 showed strong silencing in the F1 generation (Figure 5 and Table S2). Because the amount of silencing can, in some cases, approach 100% (e.g. Tw2 F1), while 4x Ubx+ resulted in ∼20–30% silencing (Figure 6), we suggest that differences beyond Ubx levels contribute to silencing in these F1 outcrosses. Genetic variation may, for example, result in differences in the levels or activities of the trans-regulators of Ubx. Silencing was also observed when Ubx-lacZlac1 and Ubx-Gal4LDN were outcrossed to wild populations, demonstrating that this effect is not limited to Ubx-Gal4lac1 (Figure 5R–5U and Table S1). Despite the silencing of Ubx enhancer traps, the pattern and levels of Ubx protein were similar in the wild stocks, our laboratory stocks, and in their F1 progeny (Figure S4B). We ruled out that the lack of enhancer trap expression in these outcrosses was due to a failure to initiate expression by carrying out a lineage tracing experiment, which demonstrates that Ubx-Gal4lac1 was expressed prior to silencing (see Materials and Methods). We also ruled out that transposon instability (e.g. hybrid dysgenesis [19]) was responsible for the loss of enhancer trap expression using several criteria (see Materials and Methods). Most importantly, silencing occurred at the same frequency when the male or female parent was from the wild (non-laboratory) stock and the amount of enhancer trap DNA, measured by qPCR, was unchanged between the parental and F2 generations. Further, silencing of enhancer traps in other genes, including Distalless-Gal4, homothorax-lacZ, and teashirt-lacZ was not observed by crossing these insertions to the same wild strains (data not shown).
We postulate that silencing induced in these outcrosses may be due to an incompatibility between the trans-acting factors (largely derived from the wild stocks) and cis-regulatory elements (linked to the monitored Ubx locus of the laboratory stock) controlling Ubx expression. In support of this idea, when Ubx-Gal4lac1 was further introgressed into weakly or strongly silencing wild stocks, which effectively increases the genetic complement from the wild strain background, an increase in the severity of silencing was observed when compared to the F1 generation (Figure 6 and Figure S5). We also never observed the complete absence of Ubx protein or haltere-to-wing transformations in any of these outcrosses, arguing that only a subset of enhancer inputs into Ubx is silenced in response to genetic variation. Consistently, individual enhancer traps responded differently when crossed to the same wild strains (Table S1).
Together, these results demonstrate that Ubx enhancer silencing is triggered when Ubx is present at higher than normal levels. When Ubx concentration is especially high (when Ubx is ectopically expressed via Gal4 or heat-shock promoters) all enhancer input into Ubx can be silenced, resulting in the complete absence of Ubx expression and haltere-to-wing transformations. Although such high levels of Ubx are not physiological, we also find that Ubx enhancer silencing can be triggered by additional copies of Ubx+, which in principle results in less than double the amount of Ubx protein. In this case, we find that the expression of some Ubx enhancer traps is clonally silenced (e.g. Ubx-Gal4lac1), while the expression of other enhancer traps (e.g. Ubx-lacZ166) is reduced. Thus, different Ubx enhancers are differentially sensitive to negative autoregulation; some are shut off by relatively low Ubx levels, while others require high Ubx levels to be silenced.
Most remarkably, we found that enhancer silencing can occur simply by varying the genetic background. In Drosophila melanogaster, due in part to its large population size, the frequency of DNA polymorphisms between individuals in the wild is estimated to be as high as 1 in 100 basepairs [20]. Due to these polymorphisms, we imagine that different strains of D. melanogaster, when kept in isolation from each other, may have subtly different ways of regulating Ubx. These may be due to strain-specific differences in the Ubx cis-regulatory elements, in the trans regulators of Ubx expression, or both. Consistent with this idea, it is of interest that gene expression levels, when assayed across entire genomes, show a lot of variability in natural populations [21],[22],[23],[24],[25]. Although we find that the final Ubx expression pattern and levels are very similar between lab and wild D. melanogaster strains, when two strains are bred together genetic differences may result in fluctuations in the initial Ubx levels. The silencing system described here may function to compensate for these fluctuations and thus ensure that the correct Ubx levels are produced throughout the haltere.
In the crosses to wild D. melanogaster strains, we found that the expression of genetically marked Ubx alleles varied tremendously, depending on the genetic background. Extrapolating from these results suggests that there is a lot of previously undetected variability in enhancer activities at the Ubx locus in wild files that would not have been detected using traditional assays. Thus, these results challenge the standard view that a given transcriptional enhancer integrates the same inputs and produces the same outputs, regardless of genetic background. Instead, due to natural genetic variation, the activity of a particular enhancer may vary widely between individuals in wild populations. Additionally, our results show that the activity of an enhancer can even vary among the cells within its expression domain (e.g. the haltere) in a single individual. We suggest that plasticity in enhancer activities is essential to compensate for genetic and perhaps environmental variation. Moreover, given that many genes may have multiple, partially redundant enhancers, enhancer silencing may be essential to buffer gene expression levels so that they remain within a narrow, biologically tolerable range. On the other hand, small differences in enhancer activities in flies in the wild may serve as a potential source of phenotypic variation that can be acted upon by natural selection. Since population genetic theory predicts that selection differentials of a small fraction of a percent are seen in natural populations with the effective population size of Drosophila [20], it is plausible that this variation is functionally significant, perhaps through a subtle influence of haltere morphology on flight performance.
The NC2 stocks were obtained from Greg Gibson (N.C. State University); all other wild stocks were obtained from the Bloomington Stock Center (Table S2).
To show that the lack of expression in these outcrosses was not due to a failure to initiate enhancer trap expression in the wild backgrounds, we carried out a lineage tracing experiment. The genotype of the stock was: Ubx-Gal4lac1 UAS-flp; actin>stop>GFP. The combination of UAS-flp and actin>stop>GFP records the history (i.e. marks the lineage) of Gal4 expression. When outcrossed to wild backgrounds, GFP expression was not silenced (in contrast to when the direct UAS-GFP readout was monitored). Together, these results suggest that Ubx-Gal4lac1 was initially activated but then silenced.
Hybrid dysgenesis was ruled out as a reason for loss of expression from P transposons by the following tests: 1) silencing occurs equally well, regardless of the direction the cross was set up, 2) silencing occurs equally well at 18° and 25°C (while hybrid dysgenesis is suppressed at 18°C), 3) silencing was not observed for some other transposon insertions (inside or outside of the Ubx locus) when crossed to the same wild stocks, 4) the miniwhite gene associated with the P element insertions did not lead to a variegated eye phenotype as would be expected for somatic transposon excision, and 5) quantitative PCR analysis confirmed that the amount of transposon DNA was the same in the parent (unsilenced) and F2 (silenced) generations. Finally, enhancer trap expression can be recovered when back-crossed into the laboratory stock background.
To measure Ubx protein levels in different genetic backgrounds, we stained haltere discs obtained from uncrowded yw (2x Ubx+), yw;If/Cyo;TM2/TM6B (1x Ubx+), yw;If/Cyo;DpP5/TM6B (3x Ubx+), yw;DpP10x2/CyoGFP;MKRS/TM6B (4x Ubx+), yw; DpP10x2/CyoGFP;DpP5/DpP5 (6x Ubx+),Hikone-R, Berlin-K, NC2-76, NC2-80, yw x NC2-76 F1s, Tw2, yw x Tw2 F1s, Florida-9, Reids-2, and Harwich wandering larvae with anti-Ubx (FP3.38) and a fluorescent secondary antibody. Stainings and confocal imaging were done identically and in parallel for ≥8 haltere discs from each genotype. The pixel intensities in identically sized regions of the distal anterior compartments were measured using Adobe Photoshop. This region was quantified because it is a relatively large area that expresses Ubx at uniform levels and gives rise to the main body of the haltere (the same portion measured in Figure 5A and Figure S4A). Similar trends were observed when average pixel intensities for the entire distal haltere were measured. The average intensities for each wild population differed by no more than 16%, suggesting that final Ubx levels are very similar despite differences in genetic background and silencing.
To quantify the extent of silencing of the Ubx-Gal4lac1 reporter in response to Ubx+ copy number and outcrosses to wild populations, third instar haltere discs were dissected from wandering larvae of yw122; DpP10x2/CyoGFP; Ubx-Gal4lac1UAS-GFP/TM6B (4xUbx+), and the GFP positive, F1 progeny of yw122; If/Cyo; Ubx-Gal4lac1UAS-GFP/TM6B crossed with NC2-80, NC2-76, Ber-2, Tw-2, and Harwich. GFP positive F3 progeny of yw122; If/Cyo; Ubx-Gal4lac1UAS-GFP/TM6B crossed with NC2-80 and NC2-76 were also dissected. For the outcrosses, we always used females from the wild populations. Haltere discs were fixed, mounted, and imaged for GFP and DAPI on a confocal microscope. Images were made binary in ImageJ. The GFP expressing area relative to the total disc area was measured for each disc, and this value was subtracted from the average GFP expressing area (relative to total disc size) of yw122; If/Cyo; Ubx-Gal4lac1UAS-GFP/TM6B haltere discs to yield a ‘% silencing’ value for each disc.
Larvae bearing the hs-UbxIa22 transgene [26] were heat-shocked at 37°C for 15–20 minutes 3 or 4 days after egg laying. Larvae were dissected at least 48 hours after heat shock to allow for total dissipation of exogenous Ubx. hs-UbxIa22 larvae that were not heat shocked showed no Ubx silencing. Neutral clones were induced using the same heat shock regime in flies of the genotype yw hsflp; FRT 42D Ub-GFP/FRT 42D; hs-UbxIa22/+.
Ubx-Gal4lac1 [27]; Ubx-lacZlac1 [28]; Ubx-Gal4LDN [29]; Ubx-Gal4M1 [29]; Ubx-lacZ166 [30]; and Ubx-Gal4M3 [29]. Although these lines are hypomorphic mutations of the Ubx locus, this is unlikely to contribute to our results because decreased production of Ubx would, if anything, cause an underestimate of the amount of silencing that occurs at the Ubx locus.
3x Ubx+ flies contain a tandem duplication of the Ubx locus (Dp(3;3)P5).
4x Ubx+ flies contain a tandem duplication of a transpositon of the Ubx locus onto the 2nd chromosome (Dp(3;2)P10). Further increases in Ubx+ copy number were created by combining these duplications [16]. Ubx9–22 expresses a non-functional Ubx protein due to a ∼1500 bp deletion that removes a splice acceptor site and part of the Ubx homeodomain-encoding exon [31].
Before crossing to enhancer traps, Ubx duplications were introduced into stocks containing marked chromosomes that do not cause silencing (yw hsflp; If/cyo; Dp(P5)/Tm6B and yw hsflp; Dp(3;2)P10x2/CyoGFP; MKRS/Tm6B).
To monitor silencing of Ubx-lacZ166 and Ubx-Gal4lac1 simultaneously (Figure S3), flies of the genotype, Dp(3;2)P10x2/heat shock-Ubx; Ubx-lacZ166/Ubx-Gal4lac1 UAS-GFP were given a 15 min. heat shock at 37°C 48 to 96 hrs after egg laying. Imaginal discs were dissected at wandering stage and stained for Ubx, βgal, and GFP. Silencing was not observed in flies of the same genotype without heat shock.
FRT101 ph504
FRT2A PcXT109
FRT42D Su(Z)2l.b8
FRT82B ScmD1
FRT42D PclD5
Of these mutations, when analyzed in loss-of-function clones, all but Pcl resulted in repression of Ubx in the haltere (due to derepression of more posterior Hox genes; data not shown) and therefore could not be used to assess their role in silencing.
UAS-GFP Ubx-Gal4lac1/TM6B
UAS-GFP (X); Ubx-Gal4LDN/TM6B
UAS-GFP (X); Ubx-Gal4M1/TM6B
FRT 82B UbxDf(109)/TM6B
hs-UbxIa22/TM6B [26]
Ubx9–22/TM6B
vg-Gal4 UAS-GFP
vg-Gal4 UAS-GFP UAS-flp act>cd2>Gal4
UAS-UbxHA
FRT42D Ub-GFP
FRT42D Ub-GFP; hs-UbxIa22/Tm6B
FRT42D
UAS-GFP; FRT42D arm-lacZ; Ubx-Gal4Lac1
hs-Gal4
(Previously described by [14]
UAS-Antp
UAS-AUA
UAS-UU* (* refers to a stop codon inserted immediately following the homeodomain)
UAS-AAU
UAS-AUU
Whole-fly genomic DNA was isolated from the lab stock containing the Ubx-Gal4lac1 enhancer trap (yw122; If/CyoGFP; Ubx-Gal4lac1 UAS-GFP/TM6B) and the GFP+ F2 progeny of the Ubx-Gal4lac1 stock crossed to strains Tw2, NC2-76, and NC2-80. Silencing was confirmed to be occurring in these crosses. The F2 progeny were generated by crossing Gal4lac1UAS-GFP F1 males to wild population females, precluding the possibility of recombination between chromosomes of the lab and wild genotypes. Primers were designed to amplify ∼200 bp in the Gal4 and UAS transgenes to determine their relative abundance in each genotype. A ∼200 bp sequence in the 5′UTR of homothorax was amplified to normalize for different amounts of template DNA. PCR amplification was performed in triplicate using Applied Biosystems 7300 Real Time PCR System, and SYBR Green PCR Master Mix. Product dissociation curves were examined to ensure that each primer set only amplified a single product. CT values and amplification curves were consistent with an equal abundance of the Gal4 and UAS sequences in all genotypes.
Standard protocols were used with the following primary antibodies:
Rabbit anti-β-Gal 1:10,000 (Cappel)
Mouse anti-En 1:10 (Hybridoma Bank)
Mouse anti-Ubx 1:20
Rat anti-HA 1:100
|
10.1371/journal.pntd.0002604 | Analysis of the Vaccine Potential of Plasmid DNA Encoding Nine Mycolactone Polyketide Synthase Domains in Mycobacterium ulcerans Infected Mice | There is no effective vaccine against Buruli ulcer. In experimental footpad infection of C57BL/6 mice with M. ulcerans, a prime-boost vaccination protocol using plasmid DNA encoding mycolyltransferase Ag85A of M. ulcerans and a homologous protein boost has shown significant, albeit transient protection, comparable to the one induced by M. bovis BCG. The mycolactone toxin is an obvious candidate for a vaccine, but by virtue of its chemical structure, this toxin is not immunogenic in itself. However, antibodies against some of the polyketide synthase domains involved in mycolactone synthesis, were found in Buruli ulcer patients and healthy controls from the same endemic region, suggesting that these domains are indeed immunogenic. Here we have analyzed the vaccine potential of nine polyketide synthase domains using a DNA prime/protein boost strategy. C57BL/6 mice were vaccinated against the following domains: acyl carrier protein 1, 2, and 3, acyltransferase (acetate) 1 and 2, acyltransferase (propionate), enoylreductase, ketoreductase A, and ketosynthase load module. As positive controls, mice were vaccinated with DNA encoding Ag85A or with M. bovis BCG. Strongest antigen specific antibodies could be detected in response to acyltransferase (propionate) and enoylreductase. Antigen-specific Th1 type cytokine responses (IL-2 or IFN-γ) were induced by vaccination against all antigens, and were strongest against acyltransferase (propionate). Finally, vaccination against acyltransferase (propionate) and enoylreductase conferred some protection against challenge with virulent M. ulcerans 1615. However, protection was weaker than the one conferred by vaccination with Ag85A or M. bovis BCG. Combinations of these polyketide synthase domains with the vaccine targeting Ag85A, of which the latter is involved in the integrity of the cell wall of the pathogen, and/or with live attenuated M. bovis BCG or mycolactone negative M. ulcerans may eventually lead to the development of an efficacious BU vaccine.
| Buruli ulcer (BU) is an infectious disease, characterized by deep, ulcerating skin lesions, particularly on arms and legs, which are provoked by a toxin. BU is caused by a microbe of the genus that also cause tuberculosis and leprosy. The 33 countries where Buruli ulcer has been detected, especially in West Africa, have mainly tropical and subtropical climates, although the disease is also present in temperate areas of Australia and Japan. There is no effective vaccine against BU and it is still not fully understood which immune defence mechanisms (antibodies and/or T cells) are needed to control the infection. The identification of microbial components that are involved in immune control is an essential step in the development of an effective vaccine. In this paper, we used an experimental mouse model to demonstrate the immunogenicity and the vaccine potential of enzymes involved in the toxin synthesis. Combinations with other vaccine candidates, such as a subunit vaccine against Ag85A targeting cell wall synthesis or with live, attenuated M. bovis BCG or mycolactone negative Mycobacterium ulcerans remain to be tested.
| Buruli ulcer (BU) is a necrotizing bacterial skin disease caused by Mycobacterium ulcerans. M. ulcerans produces a diffusible macrolide toxin, called mycolactone (ML) which is essential for bacterial virulence [1]. BU has been documented in 33 countries worldwide, although most of the cases occur in West Africa, primarily Benin, Côte d'Ivoire, Ghana and more recently Gabon. According to the World Health Organization, about 5000 cases annually are reported from 15/33 countries. The incidence in endemic regions of Ghana has been estimated at 150 cases/100 000 inhabitants. However, as the disease is not notifiable in many countries and most patients live in remote, rural areas with little medical infrastructure, the actual number of cases is likely to be much higher. Regardless, as the disease burden is mostly localized to certain geographical areas, the impact of vaccination and treatment efforts can be very high [2].
Prevention of BU is complicated by the fact that while M. ulcerans is present in the environment in disease endemic areas [3], [4], the route of transmission is largely unknown. In Australia, infection following contamination of a golf course irrigation system was reported [5] while many cases elsewhere are related to disruption of the environment, e.g. due to deforestation and building of dams [4]. Possible sources of infection include aquatic insects, mosquitoes and mammals [6], [7]. In temperate south-eastern Australia (State of Victoria) ringtail and brushtail opposums presenting typical ulcerative lesions have been identified and M. ulcerans DNA was detected at high level by real-time qPCR in faeces of these animals [8]. Person-to-person transmission appears to be extremely rare [9].
M. ulcerans is distinct from other mycobacteria in that it produces a lipid toxin (ML), which is synthesized by three large polyketide synthases encoded by mlsA1, mlsA2 and mlsB localized on the 174 kb pMUM001 virulence plasmid [10]. These synthases are composed of different modules, which each have a particular sequence of enzymatic domains. ML locally suppresses T cell responses at non-toxic levels [11]. This T cell suppression induced by ML is not completely understood, but it is clear that ML can alter both early signaling at the T cell receptor level by activation of the Src-family kinase Lck as well as blocking cytokine responses at a post-transcriptional level [11]. At higher concentrations, the toxin is cytotoxic Using a semiquantitative reverse transcription-PCR analysis of mRNA isolated from BU lesions, we have shown that production of IL-10 rather than production of IL-4 or IL-13 by Th2-type T cells may be involved in the low M. ulcerans-specific IFN-gamma response in Buruli disease patients [12]. A more in depth study by R. Phillips et al on serum from 37 BU patients from Ahafo Ano North District of Ghana demonstrated by use of Luminex technology that patients with active ulcers display a distinctive profile of immune suppression, marked by the down-modulation of four inflammatory chemokines: macrophage inflammatory protein (MIP) 1β, IL-8, monocyte chemoattractant protein (MCP) 1, and (to a lesser extent) fractalkine [13]. These immunological defects were induced early in the disease and resolved after anti-BU therapy [13]. An impaired capacity to produce Th1, Th2, and Th17 cytokines on stimulation with the mitogen Phytohaemagglutinin PHA was also observed in the Phillips' study (be it on a limited number of 4 patients with BUD and 4 healthy control participants) [13]. Interestingly, some of the defects in cytokine and chemokine response could be mimicked in vitro by incubation of CD4+ peripheral blood lymphocytes with ML [13].
ML is an obvious candidate for a BU vaccine, but by virtue of its chemical composition and possibly because of its immunosuppressive properties, the toxin is not immunogenic and in neither infected mice nor humans ML-specific antibodies have been found. However, antibodies against some of the polyketide synthase domains involved in ML synthesis, were found in BU patients and healthy controls from the same endemic region, suggesting that these domains are indeed immunogenic [14].
Aiming to interfere with ML synthesis, we have used a DNA prime/protein boost strategy targeting nine of these polyketide synthase domains. C57BL/6 mice were vaccinated against three variations of the acyl carrier protein domain (ACP1, ACP2, ACP3), against three acyltransferase domains (ATac1, ATac2, and ATp), against the enoylreductase domain (ER), against one of the ketoreductase domains (KR A) and against the ketosynthase load module domain (KS).
C57BL/6 mice were bred in the Animal Facilities of the WIV-ISP (Site Ukkel), from breeding couples originally obtained from JANVIER SAS in Le Genest Saint Isle, France. Mice were 8–10 weeks old at the start of the experiments. Female mice were used for immune analysis and male mice for the protection studies.
Virulent M. ulcerans 1615 strain (Malaysia) [10] was kindly given to us by Dr. P. Small (University of Tennessee). Bacteria were maintained and amplified in vivo in mouse footpad [15]. M. bovis BCG strain GL2 was grown for 2 weeks as a surface pellicle at 37°C on synthetic Sauton medium and homogenized by ball mill as described before and kept at −80°C in 20% of glycerol until used [16].
Bacterial expression vector pET-DEST42 encoding the genes of 8 enzymatic modules, ACP1, ACP2, ACP3, ATac1, ATac2, ATp, ER and KS or pDEST17 vector encoding KR A (all as C-terminally Histidine-tagged proteins), were constructed at the University of Melbourne, Australia and used for transformation and selection in E. coli BL-21. Following induction with IPTG for 2–4 hours, cells were lysed and recombinant proteins were purified according to standard protocol on immobilized metal affinity chromatography (IMAC) using gravity flow. Recombinant Ag85A protein from M. ulcerans (MUL 4987) was kindly given to us by Dr. G. Pluschke (Swiss Tropical and Public Health Institute, Basel, Switzerland).
Figure 1 shows the IMAC purified Pks domains and MUL4987 separated in 15% (left figure) or 12.5% SDS-PAGE (right figure) and stained with Protein Staining Solution (Thermo Scientific, Rockford, Illinois, USA) .
The genes encoding the nine enzymatic modules of the polyketide synthases were cloned in the eucaryotic expression vector pV1.Jns-tPA [17]. In this plasmid, the genes are expressed under the control of the promoter of IE1 antigen from cytomegalovirus, including intron A, preceded by the signal sequence of human tissue plasminogen activator
Briefly, sequences were amplified by PCR (Expand High Fidelity PCR System, Roche), from the corresponding pET-DEST42 and pDEST17 constructs. Primers used for cloning are shown in Table 1.
The amplified sequences were digested with Bgl II, Bcl I, or BamH I, purified on agarose (QIAkit PCR Purification kit, Qiagen) and T4 ligated into pV1.Jns-tPA vector digested with Bgl II. After ligation and transformation into DH5-α chemically competent E. coli cells (Invitrogen), clones were screened on LB-kanamycin medium (50 µg/mL) and plasmid was checked by restriction digestion and sequencing.
Plasmid DNA encoding the mature 32 kD Ag85A from M.ulcerans in V1J.ns-tPA vector was prepared as described before [18].
C57BL/6 were anesthesized by intraperitoneal injection of ketamine-xylazine and injected intramuscularly (i.m) in both quadriceps muscles with 2×50 µg plasmid V1-Jns-tPA encoding one of the nine polyketide synthase domains, empty vector as negative control and V1-Jns-tPA-Ag85A (MUL4987) as positive control on day 0 and day 21. On day 42, mice were injected subcutaneously (s.c.) in the back with 10 µg of corresponding, recombinant protein emulsified in Gerbu adjuvant, i.e. water miscible, lipid cationic biodegradable nanoparticles, completed with immunomodulators and GMDP glycopeptide (GERBU Biochemicals).
C57BL/6 mice were vaccinated intradermally with 1×105 colony forming units (CFU) of M. bovis BCG strain GL2 on day 0.
Vaccinated mice were sacrificed 3 or 6 weeks after the third immunization. Spleens were removed aseptically and homogenized in a loosely fitting Dounce homogenizer and cells were adjusted to 4×106 white blood cells/ml in RPMI-1640 medium (Gibco, Grand Island, NY) supplemented with 10% fetal calf serum (FCS), 5×10−5 M 2-mercapto-ethanol, penicillin, streptomycin and Polymyxin B sulphate (30 µg/ml, Sigma). Cells were cultivated at 37°C in a humidified CO2 incubator in round-bottom microwell plates individually and analyzed for Th1 type cytokine response to corresponding recombinant protein (5 µg/ml). Supernatants from at least three wells were pooled and stored frozen at −20°C. Cytokines were harvested after 24 h (IL-2) and 72 h (IFN-γ), when peak values of the respective cytokines can be measured.
IL-2 activity was quantified by sandwich ELISA using coating antibody anti-mouse interleukine-2 (14-7022, eBioscience) and biotinylated detection antibody anti-mouse IL-2 (JES6-5H4, 13-7021, eBioscience). The detection limit of the IL-2 ELISA is 5 pg/ml.
IFN-γ activity was quantified by sandwich ELISA using coating antibody R4-6A2 and biotinylated detection antibody XMG1.2 (both BD Pharmingen). The detection limit of the IFN-γ ELISA is 5 pg/ml.
Antigen-specific spleen cell IFN-γ secretion was also assayed by ELISPOT as described earlier. Briefly, 96-well flat-bottomed nitrocellulose plates (MAHA S4510, Millipore, Billerica, MA) were incubated overnight at 4°C with 50 µl of capture purified anti-mouse IFN-γ (15 µg/ml; BD Pharmingen, Erembodegem, Belgium) in phosphate-buffered saline (PBS) and then saturated with 200 µl/well of RPMI-complete medium 2 h at 37°C. 180 µl of spleen lymphocytes (pool of four mice per group) were added at a cell concentration of 4.106 cells/ml in the presence or absence of 20 µl proteins (5 µg/ml) and plates were incubated for 48 h at 37°C, 5% CO2. After extensive washing, plates were incubated 2 h at 37°C, 5% CO2 with 50 µl of biotinylated rat anti-mouse IFN-γ (2 µg/ml) (BD Pharmingen), washed and incubated for 45 min at 37°C, 5% CO2 with alkaline phosphatase labelled ExtrAvidine (Sigma-Aldrich, Bornem, Belgium). After washing, spots were revealed with Bio-Rad (Hercules, CA) alkaline phosphatase conjugate substrate kit, following the manufacturer's instructions and plates were analysed on a Bioreader-3000 LC (BioSys, Germany). Results are shown as mean spot-forming cells (SFC) per million lymphocytes.
Sera from C57BL/6 mice were collected by tail bleeding three and six weeks after the protein boost or six weeks after M. ulcerans challenge. Antigen-specific total immunoglobulin G (IgG) was determined by an enzyme-linked immunosorbent assay (ELISA) on serial dilutions of individual sera. The corresponding recombinant protein was used for coating (500 ng/well). Total antibody was detected using peroxidase-labeled rat anti-mouse immunoglobulin IgG (Experimental Immunology Unit, Université Catholique de Louvain, Brussels, Belgium) and orthophenylenediamine (Sigma) for revelation. Data are presented as the mean optical density at 490 nm (O.D490 nm) for 3–5 vaccinated mice tested individually for serum diluted 1∶50 and for 9 serial twofold dilutions thereof.
Six weeks after the protein boost (12 weeks after BCG), 15 mice/group were challenged with M. ulcerans 1615. 105 acid fast bacilli (AFB) obtained by in vivo passage in footpad, were injected in the right footpad of the vaccinated mice. The number of bacilli injected, suspended in Dubos Broth Base medium (Difco), was determined by counting under a microscope after Ziehl-Neelsen staining. Viability of the M. ulcerans inoculum was checked by plating on 7H11 Middlebrook agar, supplemented with oleic-acid-albumin-dextrose-catalase enrichment medium. Yellow colonies were counted after 8 weeks of incubation at 32°C. The number CFU equaled the number of AFB.
Five mice per group were sacrificed for enumeration of AFB six weeks after M. ulcerans challenge in the footpad. Briefly, the skin and bones were removed from infected footpad. Tissues were homogenized in a Dounce homogenizer and suspended in 2 ml of Dubos broth based medium containing glass beads. The number of AFB was counted on microscope slides after Ziehl-Neelsen staining.
Protection was also evaluated in ten mice/group by monitoring footpad swelling after M. ulcerans 1615 infection. The swelling was measured with a calibrated Oditest apparatus with a resolution of 0.01 mm as described previously [19]. Animals were euthanized when footpad swelling exceeded 4 mm according to the rules of the local ethical commission and survival curves were established.
For cytokine production analysis, antibody production and AFB counting, statistical analysis was made according to one-way ANOVA test. Subsequent multiple comparisons between the different groups of animals and the antigens used was made by a Tukey's correction test. Statistical results are represented in the figure by *** (p<0.001), ** (p<0.01) and * (p<0.05). Median survival time was calculated using GraphPad, Log-rank (Mantel-Cox) test.
As shown in Figure 2, vaccination against some of the Pks domains induced significant IgG antibodies. In particular, strong responses were found at three weeks after the protein boost in mice vaccinated against ATac2 and ATp. Vaccination against ACP1 and ER induced a weak IgG response (only 1/4 mice reactive), whereas IgG levels induced by vaccination against ACP2, ACP3, ATac1 KR A and KS were not different from IgG levels in naïve mice. Confirming previous findings [18], vaccination against Ag85A also induced strong antibody levels. At six weeks post vaccination (Supplementary Figure S1), IgG antibodies directed against ATac2 and ATp were still present, but lower than at week 3. ACP1 and Ag85A specific antibody levels remained at the same level, and ER specific antibodies were clearly higher at week 6 than at week 3 post protein boost (albeit with more variation between the 6 mice, antibody levels in 2/6 mice being lower than in the other 4 mice). IgG levels induced by vaccination against ACP2, ACP3, ATac1, KR A and KS remained negative.
Production of two Th1 type cytokines was analyzed in spleen cell culture supernatant of vaccinated mice stimulated with their respective antigens: Interleukin-2 and IFN-γ. IL-2 is a pleiotropic cytokine produced after antigen activation that plays pivotal roles in the immune response. Discovered as a T cell growth factor, IL-2 additionally promotes CD8+ T cell and natural killer cell cytolytic activity and modulates T cell differentiation programs in response to antigen, promoting naïve CD4+ T cell differentiation into T helper 1 (Th1) and T helper 2 (Th2) cells while inhibiting T helper 17 (Th17) and T follicular helper (Tfh) cell differentiation. Moreover, IL-2 is essential for the development and maintenance of T regulatory cells and for activation-induced cell death, thereby mediating tolerance and limiting inappropriate immune reactions [20]. The macrophage-activating cytokine IFN-γ on the other hand together with TNF-α is a well known pivotal cytokine in the control of mycobacterial infections, as illustrated by the increased susceptibility to tuberculosis in IFN-γ gene disrupted mice [21], [22]. Whereas IL-2 is produced exclusively by CD4+ T cells, IFN-γ can be produced by both CD4+ and CD8+ T cells, and therefore analysis of both cytokines may give complementary information.
Vaccination against ATac2, ATp, KR A, KS and Ag85A resulted in significant spleen cell IL-2 production in 24 hr culture supernatant, ranging between 400 and 1000 pg/ml when cells were stimulated in vitro with the corresponding antigen (Figure 3). The same two Pks domains ATac2 and ATp, that induced strong antibodies, were also good inducers of IL-2. In contrast, vaccination against KR A (which did not induce an antibody response) also induced a good IL-2 response. Vaccination against ACP1, ACP2, ACP3, ATac1 and ER induced only very modest IL-2 levels between 100 and 200 pg/ml. Stimulation of cells from unvaccinated mice with the recombinant proteins induced IL-2 levels were close to the detection limit (5 pg/ml).
Cytokine levels of the other Th1 cytokine IFN-γ were analyzed in 72 h spleen cell culture supernatants (Figure 4). Vaccination against all nine Pks domains induced some antigen-specific IFN-γ responses. Whereas vaccination against ACP1, ACP2, ATac1, ATac2, ER and KS resulted in mean IFN-γ levels of 2.500 pg/ml at most, responses against KR A and ACP3 mounted to 5.000 pg/ml and 7.500 pg/ml respectively. Finally vaccination against ATp and Ag85A resulted in mean IFN-γ levels of more than 10.000 pg/ml.
The number of IFN-γ producing cells was also examined by ELISPOT (Figure 5). Some IFN-γ producing cells could be detected after vaccination against all Pks domains, except ACP1. High numbers (between 150 and 200 SFC/106 cells) were measured in response to KS and Ag85A and highest numbers were observed in response to ATp (350 SFC/106 cells).
Mice were challenged in the footpad with virulent M. ulcerans 1615 six weeks after the protein boost and the number of AFB was enumerated 6 weeks later.
Vaccination against the ER domain, encoding an enoyl reductase, conferred significant protection at this early time point after challenge. Confirming previous findings, vaccination against Ag85A and vaccination with BCG also resulted in significantly reduced AFB numbers in footpad as compared to AFB numbers in naive mice (Figure 6).
Protection was also evaluated in ten mice/group by monitoring footpad swelling after M. ulcerans 1615 infection. The swelling was measured with a calibrated Oditest apparatus and animals were euthanized when footpad swelling exceeded 4 mm according to the rules of the local animal ethics committee.
Of all the PKS domains, only vaccination against ATp conferred a modest, but significant protection as measured by a delay in footpad swelling and median survival time increased from 47 days in the control group to 58 days in mice that received the ATp vaccine. Vaccination against Ag85A (MST 66 days) and vaccination with BCG (MST 99 days) significantly prolonged the survival time (Figure 7 and Table S1).
Buruli ulcer is a neglected tropical disease [23] for which there is no effective vaccine [2]. The M. bovis BCG vaccine, used for the prevention of tuberculosis, has been reported to offer a short-lived protection against the development of skin ulcers [24], [25] and to confer significant protection against disseminated cases of BU, e.g. osteomyelitis, both in children and in adults [26], [27]. Also in mice, BCG vaccine protects to some extent against infection with M. ulcerans [15] although a booster vaccination with the same BCG vaccine cannot increase the protective effect and mice finally succumb to the infection [19]. We have previously shown that vaccination with plasmid DNA encoding Ag85A from M. bovis BCG can protect, albeit transiently, C57BL/6 mice against footpad challenge with M. ulcerans [15]. Antigen 85 is a major secreted component in the culture filtrate of many mycobacteria such as M. bovis BCG, M. tuberculosis and M. avium subsp. paratuberculosis [28]. The antigen 85 complex (Ag85) of M. tuberculosis is actually a family of three proteins, Ag85A, Ag85B and Ag85C, which are encoded by three distinct but highly paralogous genes and that display an enzymatic mycolyl-transferase activity, involved in cell wall synthesis [29], [30]. Using a DNA prime/protein boost regimen, we have reported that a species specific vaccine composed of Ag85A from M. ulcerans was more effective than a vaccine composed of Ag85A of M. bovis BCG, conferring a protection, comparable to the protection conferred by the BCG vaccine [18].
Mycolactone is poorly immunogenic, but some of the polyketide synthase domains involved in its synthesis do induce antibodies in BU patients and healthy controls living in endemic regions of Buruli ulcer [31] [14]. Using a plasmid DNA prime/recombinant protein boost protocol, we have confirmed their immunogenicity in an experimental mouse model and have shown that vaccination can induce strong antigen-specific antibodies and Th1 type cytokine responses. Furthermore, a modest protection against a challenge infection with virulent M. ulcerans 1615 could be observed in mice vaccinated against ER (reduced AFB numbers early after challenge) and ATp (delayed footpad swelling and increased median survival time). Interestingly, ER was the only M. ulcerans-specific antigen leading to an IgG response discriminating ulcerative patients from endemic controls and antibodies against ATp could distinguish healthy controls living in endemic regions of Buruli ulcer from healthy controls living in a non-endemic region [14]. To what extent antibodies directed against these PKS domains can actually inhibit ML synthesis is not clear, but this question certainly warrants further analysis. Two studies have reported that the mycolactone PKS multienzymes are associated with the mycobacterial cell wall [31], [32]. Each mycolactone PKS domain is a component of a large, contiguous polypeptide that makes up the complete multienzyme. The structure of this enzyme is not known, but it is possible that some of the component domains might be orientated such that they are more readily accessed by host immune cells than others.
Based on biopsy specimens, M. ulcerans was originally described as an extracellular bacillus. However, the pathogen has an initial intracellular growth phase in macrophages and therefore, recognition of this early intracellular stage by an effective Th1 type immune response, could be an effective means to control the initial infection [33]. Following this proliferation phase within macrophages, M. ulcerans induces the lysis of the infected host cells and mycolactone-associated cytotoxicity is responsible for its subsequent extracellular localization [34]–[36]. [37], [38]. This extracellular phase suggests that humoral responses might also be important for protection against M. ulcerans. However, as for M. tuberculosis infection, the correlates of protection against M. ulcerans infection are also unknown. In our study, strongest protection was conferred by vaccines [perhaps use ‘antigens’ instead of ‘vaccines’??] that induced consistently strong cellular and humoral responses (see also Table S2). However, to formally proove the importance of these two cell compartments, transfer studies with specific antibodies or T cell populations would be needed. Finally, even the most immunogenic vaccine can exert its protective effect only when its cognate epitopes are generated and presented to the sensitized immune system upon infection. We speculate that the temporal differences in protection conferred by the ER and ATp vaccine might be related to variations in this antigenic presentation. In favour of this hypothesis is the finding that at the early time point after M. ulcerans infection, infected control mice produce stronger IFN-γ responses to the ER than to ATp domain (data not shown).
Hence, a combination vaccine targeting both the early intracellular and the subsequent extracellular stage, through the induction of strong Th1 T cells and antibodies respectively, may be needed to control the infection effectively. A vaccine composed of the mycolyl transferase Ag85A combined with the most immunogenic polyketide synthase domains is an interesting possibility that requires further study. Also priming with pDNA followed by boosting with M. bovis BCG or live, attenuated mycolactone-negative M. ulcerans mutants [33] could be envisaged, as has been reported in tuberculosis vaccine development [39], [40].
|
10.1371/journal.pgen.1001022 | Widespread Presence of Human BOULE Homologs among Animals and Conservation of Their Ancient Reproductive Function | Sex-specific traits that lead to the production of dimorphic gametes, sperm in males and eggs in females, are fundamental for sexual reproduction and accordingly widespread among animals. Yet the sex-biased genes that underlie these sex-specific traits are under strong selective pressure, and as a result of adaptive evolution they often become divergent. Indeed out of hundreds of male or female fertility genes identified in diverse organisms, only a very small number of them are implicated specifically in reproduction in more than one lineage. Few genes have exhibited a sex-biased, reproductive-specific requirement beyond a given phylum, raising the question of whether any sex-specific gametogenesis factors could be conserved and whether gametogenesis might have evolved multiple times. Here we describe a metazoan origin of a conserved human reproductive protein, BOULE, and its prevalence from primitive basal metazoans to chordates. We found that BOULE homologs are present in the genomes of representative species of each of the major lineages of metazoans and exhibit reproductive-specific expression in all species examined, with a preponderance of male-biased expression. Examination of Boule evolution within insect and mammalian lineages revealed little evidence for accelerated evolution, unlike most reproductive genes. Instead, purifying selection was the major force behind Boule evolution. Furthermore, loss of function of mammalian Boule resulted in male-specific infertility and a global arrest of sperm development remarkably similar to the phenotype in an insect boule mutation. This work demonstrates the conservation of a reproductive protein throughout eumetazoa, its predominant testis-biased expression in diverse bilaterian species, and conservation of a male gametogenic requirement in mice. This shows an ancient gametogenesis requirement for Boule among Bilateria and supports a model of a common origin of spermatogenesis.
| While sexual reproduction is widespread among animals, it remains enigmatic to what extent sexual reproduction is conserved and when sex-specific gametogenesis (spermatogenesis and oogenesis) originated in animals. Here we demonstrate the presence of the reproductive-specific protein Boule throughout bilaterally-symmetric animals (Bilateria) and the conservation of its male reproductive function in mice. Examination of Boule evolution in insect and mammalian lineages, representing the Protostome and Deuterostome clades of bilateral animals, failed to detect any evidence for accelerated evolution. Instead, purifying selection is the major force behind Boule evolution. Further investigation of Boule homologs among Deuterostome species revealed reproduction-specific expression, with a strong prevalence of testis-biased expression. We further determined the function of a deuterostomian Boule homolog by inactivating Boule in mice (a representative mammal, a class of Deuterostomes). Like its counterpart in Drosophila (a representative of the opposing Protostome clade), mouse Boule is also required only for male reproduction. Loss of mouse Boule prevents sperm production, resulting in a global arrest of spermatogenesis in remarkable similarity to that of Drosophila boule mutants. Our findings are consistent with a common origin for male gametogenesis among metazoans and reveal the high conservation of a reproduction-specific protein among bilaterian animals.
| Evolution of sexual reproduction, consisting of the origin and maintenance of sex, has been a central focus of evolutionary biology since the time of Darwin. The origin of sex has generally been simplified to the question of the origin of meiosis, which is known to have a single origin among all eukaryotes [1], [2]. However, sexual reproduction in higher eukaryotes is more complex than meiosis alone, and has evolved independently in plants and animals. The fundamental component of animal sexual reproduction is gametogenesis, the differentiation of sexually dimorphic male sperm and female eggs. Unlike meiosis, which is required in both males and females, most other components of gametogenesis are sex-specific or sex-biased, such as sperm tail formation. These traits are subject to strong selective pressures from natural selection, sexual selection, and/or sexual antagonism [3], [4]. Because of these selective forces, sex-biased reproductive-specific traits are known to diverge rapidly. Such patterns of rapid divergence are not only prevalent among morphological traits like male genitalia, but also extend to the molecular level, including DNA sequences, the expression profiles of sex-biased reproductive genes and regulatory pathways underlying sex determination. [5]–[17].
What remains unknown is to what extent features of sexual reproduction can be conserved. Most animals produce sex-specific gametes distinct in size, morphology, motility and development. Animal sperm are predominantly small and motile, with compact nuclei and often a beating flagellum, which are produced through a series of male-specific developmental steps called spermatogenesis. Eggs are usually large in size and immotile, and are produced through a distinct developmental process called oogenesis. The evolutionary origin of such dimorphic features of animal sexual reproduction is intriguing yet difficult to address experimentally since they left no trace in the fossil record. However, the identification of conserved male or female-specific gametogenic proteins across large evolutionary distances could uncover molecular traces of any ancient gametogenic machinery, providing evidence for a common origin of sexually dimorphic traits among animals. While reproductive proteins with conserved homologs in different phyla are not uncommon, most of them are involved in general cellular functions, and are hence also required for other non-reproductive processes [18], [19]. Their sequence conservation likely results from their pleiotropic functional constraints (i.e. additional functions in non-reproductive tissues) rather than their reproductive functions. This is consistent with a model in which the pre-existing somatic cell machinery underwent reproductive specialization during the evolution of gametogenesis in multi-cellular ancestors. A few reproductive-specific proteins have more restricted roles in sexual reproduction in distant species. However they are either required in different sexes or different stages of sexual reproduction in distant lineages of metazoans, or functional information is only available for one metazoan lineage [20]–[25]. Furthermore, most of these proteins appear to be associated with common features of germ cells in both sexes, suggesting that the unique functions that differentiate germ cells from somatic cells are likely to be conserved in animals [26]–[29]. Like meiosis, germ cells are common to both sexes and are therefore not subjected to as strong selective pressures as sex-specific or sex-biased processes are.
Besides meiosis and the specification of germ cells, most of the other components of gametogenesis appear to be sex-biased or sex-specific. Despite a large number of sex-specific gametogenesis proteins uncovered in the model organisms Drosophila, C. elegans and mice, no conserved male- or female-specific gametogenic factors common to all lineages of animals have been clearly demonstrated [30]–[33]. However, the major steps of dimorphic gametogenesis among animals are very similar. For example, the major steps in spermatogenesis consist of the development of germline stem cells, mitotic proliferation of spermatogonial cells, preparation for and entry into meiosis, meiotic divisions, and finally differentiation of haploid spermatids into highly specialized motile sperm. Most, if not all, of the developmental steps and the developmental sequence of those steps are similar among animals from different phyla [32], [34]. These similarities raise the question of whether they evolved independently in different lineages by convergent evolution, or evolved from a single ancestral prototype. While the absence of a universal male- or female-specific reproductive factor is predictable due to the fast divergence of reproductive proteins and hence is compatible with the hypothesis of multiple origins of spermatogenesis and oogenesis, it does not exclude the alternative single-origin hypothesis. It remains possible that spermatogenesis and oogenesis each evolved from a single prototype, followed by the rapid divergence of most components of the reproductive machinery. However, a few core components of the ancient prototype critical for sperm or egg production could remain conserved.
Identification of such conserved core components is the key to distinguishing between these two possibilities. Such ancient core reproductive components should fulfill the following criteria: present in most, if not all, of the major lineages of animals undergoing sexual reproduction; originated around the time when gametogenesis likely evolved; demonstrated conservation at the sequence, expression and functional levels in species from diverse animal phyla. The most stringent criteria for a conserved male or female gametogenesis factor requires a clear demonstration that these conserved components are only required for gametogenesis in one sex among animals, and not for any other processes, thus excluding any possibility that such factors are conserved due to essential functions outside of gametogenesis. A strong candidate for a conserved male gametogenic factor in animals appears to be BOULE, the ancestral member of the human DAZ gene family. The human DAZ gene family consists of a Y-linked DAZ gene and the autosomal DAZ-Like (DAZL) and BOULE genes, all of which share a conserved RNA recognition motif (RRM) and a more divergent DAZ repeat consisting of 24 amino acids rich in N, Y, and Q residues [35]. All DAZ family proteins studied so far appear to be restricted to reproduction [35], [36], and the DAZ gene is commonly deleted in men with few or no sperm [37], [38]. Although no mutations in DAZL or BOULE have been shown to be responsible for infertility in men or women, homologs of DAZL and BOULE are required for fertility in other species, and over-expression of DAZ family proteins promote the differentiation of human embryonic stem cells towards the germ cell lineage [35], [39]–[43]. Furthermore, a human BOULE transgene rescued partial testicular defects of fly boule mutants, suggesting functional similarity between these two distant homologs [44].
However, the only two metazoan Boule homologs whose function has been characterized in depth revealed opposite gametogenic requirements [39], [40]. Loss-of-function phenotypes of Boule homologs in Drosophila and Caenorhabditis elegans reveal their divergent roles in reproduction, with boule required for male reproduction in flies and for oogenesis in the worm [39], [40]. The prevalence of Boule homologs among other metazoan phyla remains unexplored, raising the possibility that Boule may have undergone adaptive evolution like many other reproductive genes and subsequently diverged at the functional level among different metazoan branches. While partial rescue of the fly boule mutant defect by a human BOULE transgene suggests functional conservation, this may only reflect similar biochemical properties of all DAZ family members [45]. Indeed, frog Dazl is able to partially rescue the Drosophila boule mutant, despite the fact that frog Dazl performs a reproductive function distinct from fly boule [46], [47]. We sought to gain further insight into the metazoan evolution of Boule and to determine if it has a general conserved reproductive function, or also a conserved sex-specific function. We systematically examined the prevalence of Boule homologs in major animal phyla and also the molecular evolution of Boule in two distant bilaterian classes. To understand the functional evolution of bilaterian Boule, we surveyed the expression of Boule homologs in representative bilaterian species and determined the functional conservation of deuterostomian Boule through expression and genetic analyses of the mouse Boule homolog.
We first asked what other animal lineages might have Boule homologs besides insects, mammals, and nematodes. In order to distinguish Boule from homologs of other DAZ family members as well as other general RNA binding proteins, we established the signature features of Boule that would allow us to identify Boule homologs in distant lineages with confidence. We separately aligned the protein sequences of known Boule homologs among two distant metazoan groups, mammals and insects, and established a consensus sequence for the RNA recognition motif (RRM) in each group (Figure S1). To determine general features of the Boule RRM we aligned the mammalian and insect consensus sequences to each other and found a 92-amino acid consensus sequence. The most conserved residues were in the two RNP motifs (PNRI(V)FVGG for RNP2 and DRAGV(I)SKGYGFV(I) for RNP1) that are known to be important for RNA binding in RRM proteins [48]. We also established a consensus sequence for the closely related Dazl RRM, and found it to be distinct from the Boule consensus sequence (Figure S2). Dazl homologs contain slightly different consensus sequences for both RNP2 (VFVGGI) and RNP1 (KGYGFVSF), have distinct sequences surrounding the RNPs, and have a conserved deletion of two amino acids (Figure S2) [35]. Interestingly, the mammalian Boule proteins appeared to share higher sequence similarity than insect homologs despite the fact that mammals have an additional Boule-like protein, Dazl, suggesting that the presence of Dazl did not relieve the selective pressure on Boule in any significant way. Not only is the sequence of the RRM highly conserved, but the proteins are similar in size, usually around 30 kDa, and contain a single RRM domain near the N-terminus [48]. While it is impossible to align all the exon-intron boundaries due to the extensive genomic divergence between distant species, we found that exon-intron structures spanning the region of the highly conserved RRM (exons 2, 3, 4, and 5) are conserved, except that Drosophila exons 3 and 4 are fused into a single exon (Figure S1C). Thus, comparison of the mammalian and insect Boule genes reveals conservation not only in specific protein sequences, but also in aspects of the genomic structure underlying these sequences.
Since sex-biased genes often undergo lineage-specific loss during evolution [13], we assessed the prevalence of Boule homologs in each branch of metazoan evolution (Figure 1). Starting with the Boule RRM consensus sequence, we used Tblastn to search the genomes of species from major phyla representing the two clades of Bilaterians, deuterostomes and protostomes, for Boule homologs. Among deuterostomes, Boule homologs were identified in at least one species of every phylum (Figure 1): in Chordata (human, Homo sapiens; mouse, Mus musculus; chicken, Gallus gallus; rainbow trout, Oncorhynchus mykiss; elephant shark, Callorhinchus milii; lamprey, Petromyzon marinus;), Tunicata (sea squirt, Ciona intestinalis), Cephalochordata (lancelet or amphioxus, Branchiostoma floridae), Echinodermata (sea urchin, Strongylocentrotus purpuratus), and Hemichordata (Acorn worm, Saccoglossus kowalevskii). Boule homologs were present in many protostomian species of the Ecdysozoa and Lophotrochozoa superphyla (Figure 1, ESZ and LTZ, Figure 2A). Boule was found in fruit flies (D. melanogaster), mosquitoes (Anopheles gambiae), lobster (Homarus americanus), green shore crab (Carcinus Maenas), wasp (Nasonia vitripennis) and nematodes (C. elegans), representing the Arthropoda and Nematoda phyla (Figure 1, ESZ), and also in each phylum of the Lophotrochozoans such as Platyhelmintha (flatworm, Schistosoma japonicum), Annelida (leech, Helobdella robusta), and Mollusca (snail, Biomphalaria glabrata) (Figure 1, LTZ). Therefore, homologs of a known gametogenic protein—Boule—are present throughout both deuterostomes and protostomes.
Next, we asked when Boule arose during evolution by determining whether Boule homologs are present in basal, non-bilaterian metazoans or beyond the animal kingdom in plants or fungi.
Based on the consensus Boule features, we determined that Boule homologs are absent in fungi and plants, suggesting that Boule is restricted to the animal lineage (Figure S3A, Figure 1). We then explored the genomes of basal metazoan animal species and found that there is no Boule homolog in the most primitive animal, Trichoplax (Figure S3B, Figure 1). However, we identified a Boule homolog in the sea anemone, a species from the primitive Cnidaria phylum. Comparison of the consensus Boule sequence against the sea anemone genome (Nematostella vectensis) reveals two proteins with high similarity [49]. Surprisingly, the RRMs of both proteins contain characteristics of the Boule consensus sequence, while one of the sea anemone proteins has identical signature RNP1 and RNP2 motifs (PNRIFVGG and GVSKGYGSVT) to those of the Boule consensus domain (Figure 3C). Furthermore, unlike the fused exons 3 and 4 in the Drosophila boule genomic structure, the sea anemone boule gene has separate exons 3 and 4 as in humans, suggesting that the ancestral Boule gene contained separate exons 2, 3, 4 and 5 that encoded the RRM. This gene (XM_001637198) is predicted to encode a protein around 22 kDa, close to the typical size of Boule proteins. Hence, a Boule homolog is present in the sea anemone, a representative of Cnidaria (Figure S3C, Figure 1). A second sea anemone protein also has some similarity to the characteristic Boule RNP1 and RNP2, but there are multiple differences in critical positions and a greater divergence from the Boule consensus sequence (Figure S3C). Furthermore, the gene itself does not possess two conserved exon/intron junctions in the second half of the RRM domain that are present in all other species examined, including Drosophila. Therefore, the second protein is likely a more divergent duplicate of the ancient Boule gene, specific to the Cnidarian lineage.
The sea anemone is one of the most primitive metazoan species that undergoes sexual reproduction. It has separate sexes, inducible spawning and external fertilization [49]. Our finding places the origin of the Boule gene prior to the divergence of Bilateria from Cnidaria, but likely after Trichoplax branched from the common ancestor of eumetazoans, making Boule one of the few ancient animal gametogenic proteins known so far. Further analysis of Boule homologs in other basal metazoan lineages could better pinpoint the origin of metazoan Boule.
Dazl arose through a duplication of Boule, likely after protostomian and deuterostomian splitting, but the exact point of Dazl origin within deuterostome evolution has not been defined [35]. Homologs of Dazl have been identified in mammals, birds, reptiles, amphibians and fish [35], [50], [51], but whether Dazl is present in other non-vertebrate deuterostomes is unknown. Using the Dazl RRM domain, we searched for Dazl homologs in the genomes of acorn worm from Hemichordata (Saccoglossus kowalevskii), sea urchin from Echinodermata (Strongylocentrotus purpuratus), lancelet from Cephalochordata (Branchiostoma floridae), and sea squirt from Tunicata (Ciona intestinalis). We could not detect any canonical Dazl homologs (Figure 1). The highest BLAST hit from those genomes were Boule homologs, suggesting that Dazl is not present in either non-chordate deuterostomes or primitive chordates, and is likely restricted to the vertebrate lineage. To further determine the origin of Dazl in vertebrate evolution, we searched the genomes of the jawless fish, lamprey (Petromyzon marinus) and could not identify a Dazl homolog (http://genome.wustl.edu/genomes/view/petromyzon_marinus/) (Figure 1). Given that Dazl is present in bony fish such as zebrafish and medaka [52], [53], we then asked if Dazl is present in the cartilaginous fish, phylogenetically the oldest group of living jawed vertebrates. We searched the genome of the elephant shark (Callorhinchus milii) and found no evidence of a Dazl homolog, though a shark Boule homolog is present [54]. This analysis suggests that Dazl originated around the time of vertebrate radiation, likely in the ancestral lineage of bony fish (Figure 1).
To further determine the evolutionary relationship of metazoan Boule homologs, we performed phylogenetic analysis of Boule homologs from the major animal branches, together with homologs of the other members of the DAZ family, Dazl and DAZ (Figure 2B and 2C). Boule clearly represents the most ancient and widespread clade among the DAZ family members, present from sea anemone to human, whereas all Dazl and DAZ homologs can be clustered together in one branch. This is consistent with the distinct reproductive functions of DAZ and Dazl homologs, and the late arrival of Dazl in vertebrate evolution and DAZ in primate evolution (Figure 2B and 2C) [35], [38], [41], [47], [55]. Conservation of Boule homologs in major lineages of animals and their evolutionary relationship throughout animal evolution suggests that Boule is a fundamental component of eumetazoan reproductive machinery essential for the survival of most animal species.
The ancient origin and widespread presence of such a reproductive gene is in stark contrast with the pervasive rapid evolution usually associated with reproductive genes, especially male reproductive genes [5], [12]. The presence of Boule in various animals provided the rare opportunity to examine how selective forces shaped the molecular evolution of a reproduction-specific gene in distant lineages. We therefore examined Boule homologs that recently diverged from each other for any signs of adaptive evolution. We analyzed two separate groups of homologs to determine if Boule is under different selective pressure when Dazl homologs are present. We compared the entire boule coding sequences among seven Drosophila species (D. melanogaster, D. sechellia, D. yakuba, D. virilis, D. erecta, D. willistoni and D. ananassae) as well as eight representative mammalian species [56].
To determine if positive selection has played a role in Boule evolution, we compared the ratio of the rate of nucleotide changes that result in a non-synonymous amino acid substitution (Ka) to the rate of nucleotide changes that cause a synonymous amino acid substitution (Ks). Positive selection is a process that favors the retention of mutations that are beneficial to the reproductive success of an individual. Neutral theory predicts that the rate of non-synonymous substitutions (that by definition affect protein sequence) is equal to the rate of synonymous substitutions. If a protein has evolved under positive selection, there are more non-synonymous substitutions (Ka) than synonymous substitutions (Ks), and an accordingly high Ka/Ks ratio. If the protein evolved under purifying selection or negative selection, a process that removes deleterious alleles, there is a decrease or absence of non-synonymous substitutions, and therefore Ka/Ks is much smaller than that expected under neutral theory. A Ka/Ks greater than 1 is a strong indication of positive selection whereas only a Ka/Ks smaller than 0.1 usually suggests a role of purifying selection.
Among all pairwise comparisons among Drosophila species, we found that non-synonymous substitutions (Ka) were not in excess of synonymous substitution (Ks). Instead, Ka/Ks ratios for all pairwise comparisons were below 0.1 (Table S1), significantly lower than the ratio reported for rapidly diverging proteins [10], [12]. Similarly, all pairwise comparisons among mammalian species revealed Ka/Ks ratios below 0.1 (Table S2), indicating that the presence of Dazl homologs in mammals had little impact on the selective pressure on Boule homologs. Furthermore, this suggests that positive selection was not the major force driving the evolution of Boule either in Drosophila or mammals. Instead, the low Ka/Ks ratio suggests that purifying selection was responsible for the strong functional constraint on the entire protein, making Boule an exception to the rapid evolution commonly seen in reproductive genes [5], [9].
The prevalence and strong functional constraint of Boule throughout protostomes and deuterostomes suggests that Boule is likely a common reproductive factor with a critical function essential for the survival of bilaterian species. However the only Boule homologs functionally characterized exhibit divergent roles in reproduction, with Drosophila boule necessary for male reproduction and the C. elegans boule homolog, daz-1, required for egg production [39], [40]. Recently, the Boule homolog in the fish Medaka was reported to be expressed in both testes and ovaries [53]. Such divergent roles and expression during gametogenesis raised the question of what the ancestral function of Boule was, and whether the expression and function of Boule homologs might have diverged despite the high conservation of the functional motif. Since Boule function has only been examined in protostomes (C. elegans and Drosophila), we reasoned that by determining Boule expression patterns in deuterostomes we could ascertain whether or not the expression or function of Boule is conserved among bilaterians.
We chose two deuterostome species (chicken and sea urchin) from separate phyla and asked if Boule homologs are preferentially expressed in the testis or ovary. Like Drosophila and C. elegans, the sea urchin is also an invertebrate and has only the Boule gene, whereas chicken is a vertebrate with both Boule and Dazl. We identified homologs of Boule in chicken (G. gallus) and purple sea urchin (S. purpuratus) (Figure 1, Figure 2A) [51], [57], and found that chicken Boule is expressed specifically in the testis, and is not present in ovaries or any other organs we examined (Figure 3A, Figure S4). However, chicken Dazl is expressed in both testes and ovaries, similar to mammalian Dazl [41], [58]. The expression of the Boule homolog in sea urchin, a primitive deuterostomian species from the Echinodermata phylum, is also testis-biased and not expressed in any non-gonadal tissue (Figure 3B). A transcript that lacks a complete RRM domain was detectable at low levels in ovary (not shown). However, this ovarian transcript may not be functional and is likely the isoform previously reported in sea urchin ovary and eggs by in situ hybridization [57]. Together these results show that deuterostome homologs of Boule are also reproduction specific, like their protostome counterparts, but with a tendency toward testis-biased expression.
Since the nematode Boule homolog is only required for ovarian function but not male gametogenesis, and Boule transcripts have been detected in the ovaries as well as in the testes of some other species [53], [57], [59], we wondered if such ovarian expression in sporadic species is a lineage-specific phenomenon or if it is a common feature. Thus, we turned to the laboratory animal model, the mouse, for an in-depth gene expression and functional analysis. Although mammalian Boule is highly expressed in the adult mouse testis but not ovary, it is not known if Boule is expressed in the ovary during development [35]. In view of the different timing of meiotic initiation in female and male mammals, we determined the developmental expression profile of mouse Boule during both male and female embryonic gonadal development (embryonic day 10.5, E12.5, E16.5 and E17.5) in comparison with adult gonads.
We first characterized the entire mouse Boule genomic region and identified alternatively spliced isoforms (Text S1 and Figure 3C). Using primers spanning all 12 exons, we found two major Boule transcripts, both of which were most highly expressed in the adult testis. The primary Boule transcript contained all 12 exons (Bol1) and was expressed only in the adult testis, whereas a second transcript lacking exon 11 (Bol2) was highly expressed in adult testes but also detectable at low levels in early embryonic gonads of both sexes and the adult ovary (Figure 3C). We thus confirmed that the predominant expression of Boule during reproductive development is in the adult testis, including a testis-specific isoform, and also identified previously unreported low levels of Boule RNA in mouse ovaries. Together with previous findings, a total of seven out of eight bilaterian species examined (human, mouse, cattle, chicken, fish medaka, sea urchin and fruit fly) representing three different phyla express Boule in the adult testis [35], [53], [60]. Expression is in the same cell types (spermatocytes and spermatids) in the testes of the human, mouse and fish medaka, suggesting conservation of developmentally-regulated testicular expression of Boule in vertebrate animals [35], [53].
The observation that Boule homologs show predominantly testis-biased expression in diverse species is consistent with a conserved male gametogenic function in bilateral animals. However, the oogenic requirement seen in C. elegans taken together with detectable levels of ovarian expression in several species suggests the possibility that an additional oogenic function is also conserved. Alternative Boule transcripts detected in mouse ovaries or embryonic gonads, albeit at much lower levels, could still play an important physiological function and therefore contribute to its sequence conservation. To ascertain if Boule is functionally conserved in deuterostomes and if ovarian expression of Boule is physiologically significant, it is necessary to examine the physiological function of Boule in deuterostomes.
To determine if the male-specific requirement of Drosophila boule is functionally conserved among Bilateria, we set out to generate a mutation of a deuterostomian Boule homolog to investigate its physiological requirement. We used mice as a representative species of deuterostomes, and used gene targeting to delete the RNA binding domain, and thus disrupt the critical function of mouse Boule (Figure 4). We replaced exon 3, which encodes a part of the RNA binding domain and is present in all Boule isoforms (Figure 3C), with a lacZ-neo vector through homologous recombination in embryonic stem cells. The removal of exon 3 resulted in a deletion of the RNA binding domain and a frame-shift in the remaining transcript. This transcript is expected to produce a truncated BOULE protein missing both its RNA binding domain and the remaining C-terminal portion of the protein. Four chimeric mice were derived and correct homologous recombination as well as germline transmission of the Boule mutation was confirmed. Homozygote mice recovered from matings among heterozygotes were identified by genotyping and confirmed by Southern hybridization (Figure 4B).
We next determined if the mouse Boule mutation was a complete loss-of-function mutation. We performed Northern blot hybridization on RNA from the testes of wild-type, heterozygote, and homozygote mice with Boule cDNA as a probe and found that there are three Boule transcripts present in wildtype testes, all of which are absent in homozygous Boule mutants. Instead, a single novel transcript corresponding to the size of the predicted chimeric transcript consisting of the truncated Boule and beta-geo (a transcript containing exon 1, 2 and lacZ, Figure 4C) is present. We further confirmed the absence of wild type Boule transcripts by the more sensitive RT-PCR and did not detect exon 3 in the mutant transcript. Instead we only detected a much larger PCR product spanning exon 2 and 4 in homozygotes (Figure 4D). This large PCR product is absent in wildtype and contains the lacZ gene from the knockout vector. Hence we conclude that Boule expression is completely disrupted in the mutant, and we have established a loss-of-function allele in the mouse Boule homolog.
Homozygote Boule mutants exhibited normal viability, growth and mating behavior (Figure 5A). We recovered the expected number of homozygotes from heterozygote matings (wild type∶ heterozygotes∶ homozygotes = 44∶79∶43), indicating that there was no effect on survival. We next tested the fertility of mice homozygous for the Boule mutation to determine whether the Boule mutation affected male and/or female reproduction. Six homozygote males were each mated with two wild type females and individually produced no pups after four months. In contrast, wild type males sired at least two litters each during that time, suggesting that the homozygous Boule males were sterile. Female Boule homozygotes showed no obvious defects and were fertile, producing an average of 8.3 pups per litter (8.3±1.4, n = 12) with heterozygote males, similar to wild type or heterozygote females (7.7±2.1; n = 7 for wildtype; 7.6±2.4; n = 18 for heterozygotes). Homozygote females continued to be fertile up to the oldest age tested (12 months). Thus, mutation of mouse Boule disrupts male reproduction but does not affect normal development, growth or female fertility, suggesting that mammalian Boule is required only for male reproduction, similar to fly boule but different from worm daz-1.
Similar physiological requirements of Drosophila and mouse Boule homologs suggest possible conservation of an ancient male gametogenic requirement. However, it is also possible that such similarity is a mere coincidence since out of hundreds of bilaterian species, only homologs from three species are functionally characterized, with one out of the three being functionally divergent. We reasoned that if mouse and Drosophila Boule function is conserved, then the specific reproductive defects of the loss-of-function mutations in both species should be more likely to be similar than if they had evolved independently by chance. We therefore determined whether mouse Boule and fly boule function within similar processes of male reproduction. In both flies and mice, the major reproductive organs are clustered together in the male reproductive tract. The tract consists of a pair of testes for sperm production; a pair of sperm storage/maturation organs (the epididymis in mice and seminal vesicle in flies); accessory glands for providing proteins, other nutrients and seminal fluid that accompany sperm migration and fertilization (prostate, seminal vesicles and coagulating glands in mice and accessory glands in flies); and a sperm transport duct used for sperm transportation and maturation (vas deferens and urethra in mice and ejaculatory duct in flies) (Figure 5B) [30]. While the major components of the male reproductive tract in mammals and insects appear to serve similar reproductive functions, it is not known if any components are evolutionarily related between vertebrates and invertebrates. In mouse Boule mutants, all the components of the male reproductive tract are present and intact (compare Figure 5B and 5C), similar to that of the fly boule mutant [39]. Compared with wild type mice, the male reproductive tracts of Boule homozygous mutant mice were morphologically indistinguishable except for the testes, which are smaller by weight [61]. Hence the sterility defect is the result of a defect in the testis, similar to the sterility defect associated with the Drosophila boule mutation [39], [61].
Further characterization of the reproductive defects revealed that mouse Boule mutant epididymides lacked mature sperm (Figure 5E), and instead contained degenerating cells that were not seen in the wild type (compare Figure 5D and 5F to Figure 5E and 5G). Therefore, the observed male sterility of the Boule homozygous mutant mice appears to be due to a complete absence of sperm in the epididymis.
Next, we examined the developmental impact of the mouse Boule mutation on sperm production. While the overall testicular structure was normal and all the somatic cell types were present in mouse Boule mutants, the effect of mouse Boule mutation on sperm development was dramatic, with a complete halt of spermatogenesis inside all seminiferous tubules of the testis. Both mature sperm and developing elongating spermatids were entirely absent from the lumen of individual seminiferous tubules (compare Figure 5H and 5J to Figure 5I and 5K). This indicates that the failure to produce sperm resulted from a major block in sperm production due to a global arrest of spermatogenesis prior to spermatid differentiation, similar to the spermatogenic defect seen in the testes of boule mutant flies [39], [61]. Such similar global, spermatogenic-specific impacts of mutations in orthologs in divergent phyla is surprising and unprecedented, given the vastly different organization of testicular structure, type of spermatogenesis and differences in the contribution of hormonal control in mammals and insects.
In Drosophila, spermatogenesis is cystic, where a single spermatogonial cell and its clonal descendants are encapsulated in a somatic cyst throughout sperm development. Though mouse spermatogenesis is acystic, we observed prominent ball-like structures containing degenerating cells in the mouse Boule mutant, resembling the degenerating cysts seen in the fly boule testis (Figure 5I and 5K, arrowheads). Although the cyst structure is not present in mammalian spermatogenesis, descendant cells from a single spermatogonial stem cell remain connected with each other through cytoplasmic bridges during mouse sperm development, similar to that in Drosophila spermatogenesis [62]–[65]. This phenomenon could lead to the merging of multiple interconnected arrested spermatogenic cells in Boule mutant testes, resulting in such giant “cysts” with multiple nuclei. Despite the distinct modes of spermatogenesis in mice and Drosophila, the mouse Boule and fly boule mutations caused a remarkably similar and specific global arrest of spermatogenesis. Though further characterization of the developmental and cellular defects in mouse Boule mutant testes is needed to determine the full extent of similarity in developmental and cellular defects between mouse and Drosophila Boule mutants [61], our data demonstrate a key physiological requirement of Boule in sperm development and conservation of its male reproduction function between two distant lineages of a protostome (Drosophila) and a deuterostome (mouse).
Evolution of reproductive traits and genes is of the utmost interest to our understanding of the central questions in evolutionary biology such as speciation. However, the relatively rapid divergence of sex-biased reproductive genes in comparison with somatic cell proteins or non sex-biased reproductive proteins during evolution has made it difficult to study the evolution of sex-specific reproductive systems across extended evolutionary distances. Even though some reproductive genes are conserved beyond a given phyla, they are often also involved in other developmental processes. Such broad functionality compounds studies of their reproductive evolution because the selective pressures driving their evolution may be due to critical somatic functions, and not a reproduction-related function. The human DAZ family of reproductive genes, with homologs in diverse species, many of which are specifically expressed in reproductive tissues, are ideal candidates for the study of reproduction-specific gene evolution [35], [38], [40], [41], [46], [53], [60]. In particular, Boule, the ancestral gene member, is reproductive specific in flies and worms.
We identified homologs of Boule in the major phyla of metazoans, reconstructed the evolutionary history of Boule, and began to determine its functional divergence. We found that Boule, unlike other reproductive proteins, has been maintained in all major phyla of bilaterian animals as well as in Cnidarians, but are absent in the most primitive animals (the placozoan Trichoplax), fungi and plants (Figure 1). We found that Dazl homologs are only present in vertebrates, supporting the hypothesis that Boule is the ancestral member of the DAZ family [35]. Dazl homologs were absent in representative species of non-vertebrate deuterostomes and cartilaginous fish (elephant shark), but were present in bony fish and tetrapod animals (Figure 1). This places the origin of Dazl after the divergence of bony fish from cartilaginous fish but before the arrival of tetrapod animals (Figure 6). On the other hand, the widespread presence of Boule in eumetazoan animals indicates that the ancient Boule gene was present as early as 600 million years ago in the Precambrian era, in the common ancestors of Bilaterians (often called Urbilateria) as well as eumetazoans (Figure 6 and Figure 7) [66], [67].
Interestingly, human BOULE has previously been shown to be able to function in Drosophila testes, and can even rescue meiotic defects of boule mutant flies, suggesting a conservation of a spermatogenesis-specific function [35], [44]. However, the C. elegans boule homolog daz-1 is required only in oogenesis [40], making it unclear whether such a transgenic replacement in the fly actually represents a legitimate functional conservation. Furthermore, both C. elegans and Drosophila are protostomes, so whether Boule is even required for reproduction, let alone restricted to spermatogenesis, in any deuterostome species was not known. Using mice as a representative deuterostome, we generated a Boule null allele to address this question (Figure 4). Boule is required only for male reproduction in mice (Figure 5), similar to insect boule, revealing not only a conserved function, but suggesting an ancient requirement of Boule in gametogenesis. Furthermore, the requirement of mouse Boule for male reproduction and its dispensability for female fertility suggests that low level expression of Boule in embryonic germ cells and adult ovaries is not essential for either the development of germ cells or the production of female gametes. Similarly, Drosophila boule, initially thought to be testis-specific, has also been found to be alternatively spliced and expressed in the ovary and even some somatic tissues at a low level, though loss-of-function similarly only causes male sterility [59], [68]. Interestingly, this result shows that Boule has a spermatogenesis-specific requirement conserved in at least two distant lineages of bilateral animals, making it a strong candidate for a conserved male gametogenesis factor between Drosophila and mammals.
Given that mouse Boule is required for sperm production like fly boule but different from C. elegans daz-1, we propose that Urbilaterian Boule had an ancestral function in male gametogenesis which was lost during the evolution of the nematode lineage (Figure 7). This is consistent with the higher sequence divergence of the C. elegans daz-1 RNA binding motif than most other bilaterian Boule homologs (Figure 2). While we can not rule out the possibility that the similar male gametogenic requirement in mice and Drosophila is a coincidence and both evolved independently, the striking similarity in the reproductive defects of loss-of-function mutants of Drosophila and mouse Boule homologs (male specific infertility, global arrest of spermatogenesis, absence of elongating spermatids and mature sperm) argue against such a possibility. Furthermore, the predominance of testis-biased expression of Boule homologs among distinct bilaterian species (Figure 3), supports a model of an ancient male gametogenic function (Figure 7).
It is important to note, however, that this model does not exclude the possibility of an additional ancestral ovarian function of Urbilaterian Boule. Since ovary expression of Boule is also prevalent among diverse animals, and C. elegans daz-1 is required in females, the ancestral Boule gene may have also played a role in oogenesis, which may have been subsequently lost in specific lineages. Our data does not rule out this possibility, and such an ancestral oogenesis function of Boule could be in addition to our proposed ancient spermatogenesis function. Further functional analysis in other lineages, including medaka where strong ovarian Boule expression has been observed, could help determine the more likely scenario [53]. Additionally, characterization of Boule homolog(s) in the sea anemone, an outgroup to the bilaterian lineage, could provide further insights into the ancestral roles of Boule.
Whether or not an ovarian function of Boule is also conserved, our discovery that mammalian Boule is required only for sperm development like its fly counterpart is the first such demonstration of a conserved spermatogenesis-specific function in both lineages. While spermatogenesis occurs in the testes of different animal lineages, it is not known if either spermatogenesis or the testis itself is evolutionarily related between vertebrates and invertebrates. The fly testis, which is a single tube with a linear progression of spermatogenesis, appears different from the mouse testis, which is composed of many seminiferous tubules with a concentric progression of spermatogenesis from the periphery of the seminiferous tubules towards the lumen. However, if we focus on a single cycle of spermatogenesis within a segment of a mouse seminiferous tubule and compare it with a fly testis tubule, we see similar spermatogenic cell types present inside the fly testis tubule and the mouse seminiferous tubule segment [69]. Spermatogenesis in both species starts with spermatogonial stem cells located in a specific position of the tubule, attached to the apical end in fly and to basement membrane in mice, which move and progress into later stages of cell types in one direction, towards the basal end of the testis tubule in fly and towards the lumen in mice. All the major stages of sperm development appear to be present in both species and arranged in a similar spatial and temporal pattern. If both developmental processes evolved from an ancient primitive spermatogenesis prototype, one would predict the presence of at least some common male gametogenesis-specific regulators in both lineages. Yet no such common male gametogenesis factor has been demonstrated to be required exclusively for sperm production in both lineages. The lack of a universal male reproductive factor among all animal lineages, while consistent with rapid evolution of male reproductive genes, is in contrast to the prevalence of sexual reproduction and in particular to the similarity in male gametogenesis among metazoan animals [34], [69]. This paradox led to the question of whether such similarity in the reproductive traits arose from convergent evolution or from conservation of an ancient prototype in the common ancestor.
Furthermore, male reproductive traits and genes undergo rapid adaptive evolution in diverse lineages such as Drosophila, fish, rodents and primates [5], [9]–[12], [16], [17]. Male-biased genes exhibit a higher divergence of expression among closely related species than female-biased genes or genes expressed in both sexes [7], [13]. Additionally, testis-biased genes have the highest rate of extinction and species-specific de novo gene formation during evolution [13], [70], [71]. For example, the most widespread testis-specific proteins among both vertebrates and invertebrates appear to be sperm nuclear basic proteins (SNBP). Many organisms replace histones with a set of small basic structural proteins (SNBP) or protamines to establish a highly compact sperm chromatin structure [72], [73]. Although all metazoan SNBP homologs share their common ancestry with somatic histone H1 protein, the testis-specific SNBPs in different lineages have undergone extensive lineage-specific loss and dynamic evolution, including adaptive evolution [5], [73]. Furthermore it remains unclear if vertebrate and invertebrate protamine homologs are functionally conserved. Loss of one copy of either mouse Protamine-1 or Protamine-2 leads to male sterility, but in contrast, fly sperm carrying a deletion of both protamine-like homologs appears to be functional [74], [75]. Sexual selection has been proposed to be the major force driving this fast divergence of male reproductive traits, gene sequences, and their expression patterns [9], [10], [12]. Given that sexual reproduction is widespread among animals and sperm production appears to be present in all major phyla of metazoan animals, it raised a question whether any male-biased reproductive gene could be exempt from such selective pressure and remain conserved through extended evolutionary distances.
However, Boule homologs have been maintained throughout all major lineages of animals from a common eumetazoan ancestral gene and are required only for sperm development in both Drosophila and mice. We have shown that Boule proteins have resisted sexual selective pressure, and instead evolved under purifying selection. Though ancestral Boule may have also functioned in oogenesis, our findings that bilaterian Boule homologs tend toward male-biased expression, taken together with the similar spermatogenesis arrest phenotypes in both Drosophila and mouse mutants, supports the model of a common origin of bilaterian spermatogenesis.
While it remains to be seen if Boule homologs are restricted only to spermatogenesis or also function in the ovary, we have shown a clear case of conservation of a reproduction-specific gene across Bilateria. We found that among a broad representation of bilaterian animals, Boule expression was restricted to the gonads (Figure 3), indicating that it has remained reproduction-specific throughout evolution. In addition, DNA sequence analysis of multiple Drosophila and mammalian Boule homologs revealed that, unlike other reproductive proteins [11], [16], [17], Boule evolution has been driven not by positive selection, but by purifying selection. This establishes an unambiguous case of a reproduction-specific gene being driven predominantly by purifying selection, in two distinct animal lineages, suggesting a strong functional constraint. Interestingly, our in-depth analysis of the developmental defects in Boule null mice revealed a novel requirement in spermatid differentiation [61]. Such a postmeiotic function for boule is also likely present in Drosophila, though its requirement for spermatid differentiation would not have been revealed in the boule mutant flies due to an earlier block at meiosis [39], [61]. The previously established function of Boule in meiotic progression in both Drosophila and nematodes [39], [40] may also be conserved in mice, despite the lack of a similar meiotic defect in Boule null mice [61]. We proposed that Dazl and Boule may redundantly regulate meiosis, and that Dazl may compensate for Boule loss during meiosis in mice [61]. Yet despite this possibility of a partial redundancy of function with Dazl, mouse Boule has been maintained under purifying selection, further indicating that the presence of other DAZ family genes has had little impact on the functional constraint of Boule. While meiosis is fundamental to sexual reproduction and key components of meiotic machinery for chromosomal synapses and recombination are conserved from yeast to mammals [2], [76], the absence of Boule homologs in fungi together with the requirement of Boule homologs in only one sex of animals suggest that conservation of Boule is unlikely due to the same functional constraint that keeps components of meiotic machinery conserved. Another main functional constraint on metazoan reproduction appears to be associated with germ cell specification and maintenance. Mutations disrupting those conserved germ cell components, such as Vasa or Piwi, often result in a failure to form germ cells or a loss of germ cells before meiotic stages. Furthermore, the resulting infertility sometimes affects both males and females of the same species [21]–[23], [77]–[79]. These phenotypes differ from the sex-biased infertility of Boule mutations in all species examined, and the gametogenesis defects in Boule mutants are much less variable than those from either Vasa or Piwi mutants across species [21]–[23], [61], [78], [79]. Further characterization of the subcellular expression and molecular function of Boule will help to discern the relationship between Boule and these other highly conserved germ cell proteins.
We've shown the widespread presence of Boule homologs throughout bilaterian animals and the functional conservation of a reproductive-exclusive requirement among Drosophila, worm and mouse. This has revealed an ancient reproductive requirement in the Urbilaterian, the common ancestor of all bilaterian animals and highlights a fundamental reproductive function associated with Boule protein conserved over six hundred million years of evolution. With the identification of Boule and possibly more reproductive genes conserved across such large evolutionary distances, we can begin to compare the impact of sexual selection on the molecular evolution of the same components of reproductive traits in different animal lineages at both the microevolution and macroevolution levels.
For known Boule homologs, DNA sequences from various species were retrieved from the literature and the Genbank database. For species where the presence of Boule or Dazl homologs was unknown, we first searched the EST and cDNA database in Genbank using consensus RRM sequences of either Boule or Dazl using Tblastn and positively identified the homologs using our established criteria. The homolog sequences were further confirmed by the presence of Boule/Dazl homologs with high sequence similarity in other species within the same taxon. In the absence of EST or cDNA information, we then focused on a representative species from the same phylum whose genome had been completely sequenced. The specific genome databases were searched using the consensus Boule RRM sequences and Tblastn, and the top hits were analyzed to determine if they were Boule homologs based on criteria described above. The identified homologs were verified by BLASTing against the human protein database, which should identify human BOULE as the top hit sequence with highest similarity. New Boule and Dazl homologs we have identified as well as known homologs from previous publications are summarized in Table S3. Sequence alignment of RRMs and entire proteins was performed using ClustalW2 and ClustalX programs [80]. The parameters for alignment were protein Gap open penalty = 10, protein extension penalty = 0.2, and other parameters at default settings. Phylogenetic analysis was done using Mega 4.0 [81]. Ka and Ks were calculated as described [44].
RNA was extracted by Trizol from tissues and reverse transcribed for amplification of Boule cDNA. For tissues collected in RNA later (Applied Biosystems/Ambion, Austin TX), samples were stored at 4°C overnight, solution was removed, and the tissues were stored at −80°C for later RNA extraction. A minimum of two pairs of primers spanning the RRM region and other regions were used to confirm the expression of the Boule gene (Table S4). All the Boule amplicons were confirmed by sequencing. Dmrt1 (Doublesex and mab-3 related transcription factor 1) was a testis-specific positive control in chicken [82] and Bnd (Bindin) was a testis-specific positive control in the sea urchin [83]. We used multiple sets of primers covering exons 2, 3, 4, 5 and 6, and determined that the main transcript is testis-specific (Table S4).
Mice (Mus musculus) were housed and bred in a barrier facility according to the guidelines approved by the ACUC committee at Northwestern University. Boule mutant mice were created in the mixed background of C57B6 and 129svj. Ripe purple sea urchins (Strongylocentrotus purpuratus) in spawning season (May, 2009) were collected from Pacific Ocean off Carslad, California (M-REP, Carslad, California) and shipped overnight on ice to Chicago. The sex of sea urchins was determined by the presence of eggs (often a milky spill on the outside of female urchins upon arrival) and by the presence of distinct gametes in the gonad tissue biopsies. Gonads, guts and ampullae from at least three male and three female purple sea urchins were collected and either stored in Trizol for immediate RNA extraction or snap-frozen in liquid nitrogen and stored at −80°C. Chicken gonadal and other tissues were collected from euthanized White leghorn chickens at the completion of a research project approved by UIUC animal committee at the University of Illinois at Urbana-Champaign Veterinary School. Tissues from two four-year old roosters and two three-year old hens were snap-frozen in liquid nitrogen or stored directly in RNA later.
Tissue histology was performed as described previously [35]. Testes were fixed in Bouins' solution overnight and sectioned at 5-µm thickness for hematoxylin/eosin staining. Bright field images were captured using a Leica DM 5000B compound microscope with a DFC320 camera and the Leica image capture suite software.
We replaced exon 3 with lacZ-neo using the NZTK2 vector (Richard Palmiter, University of Washington, Seattle, WA). We designed primers with built-in SalI sites to amplify a 2-kb left arm next to exon 3, and primers with built-in XhoI and NotI sites to amplify a 5.9-kb right arm from 129svj mouse genomic DNA. High fidelity platinum PCR kits (Invitrogen) were used to amplify the fragment with minimal PCR error. The amplified fragments were cloned into Topo vectors and later released with appropriate enzymes for subcloning into the NZTK2 vector. The clones with correct orientation of left and right arm insertions were chosen for sequencing. Sequencing of the genomic arms in the selected clones indicated that both arms had greater than 99% sequence identity with the genomic sequence. Gene targeting was performed on 200 ES cell clones (129svj E14 feeder cell-less ES cells) and four positive ES clones (1D5, 1H5, 2A4 and 2D7) were identified by the presence of both the 2-kb and 6-kb arms using primers outside each arm and on the vector. The 1D5 clone was used to inject blastocysts at the Northwestern University Transgenic Core Facility. Four chimerical mice produced lacZ-positive progeny and mice from two independent founders were used to generate mutant mice for the analysis. The phenotypes were identical among the mutant mice from two independent lines and we did not distinguish our analyses between the two lines.
|
10.1371/journal.pgen.1000372 | Loss of Myotubularin Function Results in T-Tubule Disorganization in Zebrafish and Human Myotubular Myopathy | Myotubularin is a lipid phosphatase implicated in endosomal trafficking in vitro, but with an unknown function in vivo. Mutations in myotubularin cause myotubular myopathy, a devastating congenital myopathy with unclear pathogenesis and no current therapies. Myotubular myopathy was the first described of a growing list of conditions caused by mutations in proteins implicated in membrane trafficking. To advance the understanding of myotubularin function and disease pathogenesis, we have created a zebrafish model of myotubular myopathy using morpholino antisense technology. Zebrafish with reduced levels of myotubularin have significantly impaired motor function and obvious histopathologic changes in their muscle. These changes include abnormally shaped and positioned nuclei and myofiber hypotrophy. These findings are consistent with those observed in the human disease. We demonstrate for the first time that myotubularin functions to regulate PI3P levels in a vertebrate in vivo, and that homologous myotubularin-related proteins can functionally compensate for the loss of myotubularin. Finally, we identify abnormalities in the tubulo-reticular network in muscle from myotubularin zebrafish morphants and correlate these changes with abnormalities in T-tubule organization in biopsies from patients with myotubular myopathy. In all, we have generated a new model of myotubular myopathy and employed this model to uncover a novel function for myotubularin and a new pathomechanism for the human disease that may explain the weakness associated with the condition (defective excitation–contraction coupling). In addition, our findings of tubuloreticular abnormalities and defective excitation-contraction coupling mechanistically link myotubular myopathy with several other inherited muscle diseases, most notably those due to ryanodine receptor mutations. Based on our findings, we speculate that congenital myopathies, usually considered entities with similar clinical features but very disparate pathomechanisms, may at their root be disorders of calcium homeostasis.
| Congenital myopathies are inherited muscle conditions typically presenting in early childhood. They are individually rare, but as a group are likely as common as conditions such as muscular dystrophy. The zebrafish is an emerging experimental system for the study of myopathies. We have utilized the zebrafish to develop a model of myotubular myopathy, one of the most severe childhood muscle diseases and a condition whose pathogenesis is poorly understood. We have generated fish that have the characteristic behavioral and histological features of human myotubular myopathy. Using this model, we then made novel insights into the pathogenesis of myotubular myopathy, including the identification of abnormalities in the muscle tubulo-reticular system. We subsequently identified similar changes in muscle from patients with myotubular myopathy, corroborating the importance of our zebrafish findings. Because a functional tubulo-reticular complex is required for normal muscle contraction, we speculate that the weakness observed in myotubular myopathy is caused by breakdown of this network. In all, our study is the first to identify a potential pathomechanism to explain the clinical features of myotubular myopathy. Furthermore, by revealing abnormalities in the tubulo-reticular system, we provide a novel link between myotubular myopathy and several other congenital myopathies.
| Myotubular myopathy is a severe, X-linked congenital myopathy with onset in infancy [1]. It is characterized by profound neonatal hypotonia and skeletal muscle weakness. It is associated with substantial mortality, with approximately half of all affected boys dying in the first year of life [2]. Surviving children have significant morbidity associated with respiratory compromise and difficulties with ambulation. Currently there are no treatments or disease modifying therapies available for this condition.
The condition is defined by characteristic changes observed on muscle biopsy [1]. Biopsies show muscle fiber hypotrophy and an abundance of fibers with large, centralized nuclei of unusual appearance. These nuclei are distinct in appearance from those observed in degenerative conditions like Duchenne muscular dystrophy, and are the defining pathologic features of a group of congenital myopathies called centronuclear myopathies [3].
Myotubular myopathy is caused by mutations in the myotubularin gene [4]. Over 200 mutations have been reported in the myotubularin gene, the majority of which result in loss of functional gene expression [1]. Myotubularin is the only gene associated with myotubular myopathy. It is the canonical member of a large family of homologous proteins called the myotubularin related proteins (MTMRs) [5]. Of interest is the fact that several MTMRs are mutated in human neurologic diseases, including mutation of MTMR14 in an autosomal form of centronuclear myopathy [6].
Myotubularin was originally characterized as a protein tyrosine phosphatase, but was subsequently found instead to function primarily as a lipid phosphatase [7],[8]. It acts specifically to remove phosphates from the 3-position of phosphoinositide rings. As demonstrated in cell free biochemical assays [7],[8] and with forced exogenous expression [9],[10], myotubularin converts phosphoinositide-3-phosphate (PI3P) to phosphoinositide phosphate (PIP) and phosphoinositide-3,5-bisphosphate (PI3,5P2) to phosphoinositide-5-phosphate (PI5P). Most recently, Cao and colleagues have demonstrated using RNAi in A431 cells that knockdown of myotubularin results in a 60–120% increase in PI3P levels, thus substantiating the requirement for myotubularin in the regulation of endogenous PI3P [11]. Increased PI3P levels have also been observed in yeast lacking the myotubularin homolog ymr1 [8],[12]. As yet, however, this activity has not been directly examined in whole vertebrates or in specific organ systems, including muscle. The functional importance of myotubularin's phosphatase activity is assumed from the fact that missense mutations that alter critical amino acids in the phosphatase domain without affecting protein stability result in myotubular myopathy [1].
Phosphoinositides are implicated in myriad cellular functions, chief among them the regulation of membrane traffic and vesicle/organelle movement [13]. Because it acts to modify certain PI residues, myotubularin is assumed to function as a regulator of membrane traffic and in particular the movements of vesicles between endosomal compartments [14],[15]. Overexpression of myotubularin in cell culture delays traffic out of the endosomal compartment and causes vacuole accumulation. However, as with myotubularin phosphatase activity, a role for myotubularin in the regulation of membrane traffic in vivo and specifically in skeletal muscle has yet to be determined. In addition, unlike with other myopathies due to altered membrane traffic (examples include Danon Disease due to LAMP2 mutation [16]), myotubular myopathy is not characterized by the pathologic accumulation of vesicles.
Many critical questions remain unanswered concerning myotubularin function and myotubular myopathy pathogenesis. These include whether myotubularin truly functions as a lipid phosphatase and regulator of membrane traffic in vivo. Furthermore, the relationship between the proposed functions of myotubularin and disease pathogenesis is not clear. The same is true with the association between the unique histologic appearance of the muscle in myotubular myopathy patient biopsies and the etiology of muscle weakness and hypotonia. The lack of knowledge concerning these fundamental issues is a significant barrier in the development of therapeutic strategies for the disease.
A murine model of myotubular myopathy exists, generated by targeted mutagenesis [17]. It recapitulates the clinical and histopathologic features of the disease, thus confirming the association between myotubularin and myotubular myopathy. However, due in part to technical limitations with the murine system, it does not address many of the fundamental questions mentioned above. To begin answering these questions, and to develop a model system amenable to rapid testing of therapeutic strategies, we report here the development of a zebrafish model of myotubular myopathy. Using antisense morpholino technology, we generated zebrafish embryos with reduced myotubularin protein expression. These embryos have severely impaired motor function, muscle fiber atrophy and the presence of large, abnormally located nuclei. These findings are reminiscent of those seen in myotubular myopathy. We also demonstrate that loss of myotubularin causes increased PI3P levels in muscle, thus confirming for the first time that myotubularin functions as a lipid phosphatase in a vertebrate model system. Using RNA-mediated rescue experiments, we show that the homologous myotubularin-related genes MTMR1 and MTMR2 are able to functionally compensate for the loss of myotubularin. Lastly, and most significantly, we identify alterations in the T-tubule and sarcoplasmic reticular networks in morphant zebrafish muscle. We confirm that similar disorganization of the tubulo-reticular network is present in biopsy samples from patients with myotubular myopathy. In all, we have successfully created a zebrafish model of myotubular myopathy, and have used this model to both answer fundamental questions concerning myotubularin function and to uncover a novel mechanism to explain the pathogenesis of the disorder.
To study the function of myotubularin (MTM1) in zebrafish, we employed antisense morpholinos to achieve functional gene knockdown. We first identified the zebrafish homolog of MTM1 using the Ensembl genome browser (ENSDARG00000037560). By bioinformatics and RT-PCR from zebrafish embryonic RNA, we found that MTM1 and 12 of 14 of the MTM1-related gene products (MTMRs) are expressed in the developing fish (Figure S1). We then designed morpholinos to the translation start site (ATG MO), to the splice donor site of exon 1 (Ex1 MO), and to the splice acceptor site of exon 3 (Ex3 MO). Both splice morphants were predicted to result in the loss of an exon and the introduction of a premature stop codon. These morpholinos were independently injected into 1–4 cell stage embryos and then embryos were phenotypically analyzed at 24, 48, and 72 hours post fertilization (hpf). A control morpholino (CTL MO) designed to a random sequence of nucleotides not found in the zebrafish genome was used to control for injection related non-specific effects [18].
The efficacy of the ATG morpholino to interfere with translation was verified by the demonstration of reduced myotubularin protein levels by immunofluorescence and western blot analysis of samples from ATG MO injected embryos (Figure S2A, B). The ability of the splice morphants to alter myotubularin RNA processing and stability was confirmed by RT PCR analysis using primers to flanking exons (Figure S2C). Of note, all 3 morpholinos yielded indistinguishable phenotypes. The ATG morpholino was used for analysis and quantitation in all subsequent experimentation, with all phenotypic observations additionally verified using the two splice morpholinos.
Zebrafish embryos undergo rapid skeletal muscle development, and multinucleated myotubes are present and easily recognizable by 24 hours post fertilization. We thus began our analysis at this time point. Live microscopic analysis of myotubularin morphant embryos revealed a subtle but reproducible abnormality in body shape. Specifically, knockdown embryos exhibited a dorsal curvature (**) through the back and tail instead of the normal flat or C-shaped dorsum (Figure 1A). A similar morphologic abnormality has been observed in other zebrafish models of congenital myopathies [19],[20].
Myotubularin morphant zebrafish began exhibiting more distinct morphologic abnormalities starting at 48 hpf, with the most obvious changes present in embryos at 3 days post fertilization (Figure 1B). The most consistent finding was thinning of the muscle compartment (bracket, Figure 1B). Morphant embryos also frequently had bent and/or foreshortened tails, a feature commonly associated with abnormalities in muscle structure or function (arrow, Figure 1B). Of note, the most severely affected embryos (ex: bottom embryo, Figure 1B) also exhibited changes consistent with an overall delay in embryonic development (small heads, abnormally shaped yolk balls, and reduced body extension).
In zebrafish, the first recognizable muscle dependent motor function, detected between 17 and 26 hpf, is spontaneous embryo coiling [21]. On average, control injected embryos had 10.2 (+/−0.4) spontaneous muscle contractions per 15 second period (Supplemental Video 1). Conversely, embryos injected with myotubularin morpholinos had only 5.2 (+/−0.5) contractions in the same period (Figure 2A and Supplemental Video 2). This abnormality was highly reproducible (P<0.0001), and marked the earliest observed functional abnormality in zebrafish with reduced myotubularin levels.
In addition to a decrease in spontaneous coiling frequency, myotubularin morphants also displayed defective motor behaviors later in development. Normally bouts of muscle activity contribute to the hatching of larvae from their protective outer chorion between 48 and 60 hpf. Typically approximately 90% (87.2%+/−2.3%) of control injected embryos by 60 hpf had hatched from their chorions (Figure 2B). In contrast, only 35.3% (+/−3.3%) of age-matched myotubularin morpholino-injected embryos were found to have hatched (Figure 2B), consistent with a continued decrease in muscle activity. In the most severe morphants, delayed embryonic development also likely contributed to the reduction in chorion hatching.
Once hatched, the myotubularin morphant larvae also displayed profound abnormalities in touch-evoked escape behaviors. Typically, 72 hpf larvae respond to tactile stimuli with a rapid and vigorous escape contraction, followed by swimming, which often resulted in larvae swimming out of the field of view (Figure 2C; Supplemental Video 3). In contrast, myotubularin morphants displayed weak escape contractions, followed by diminished swimming that often failed to propel the larvae out of the field of view (Figure 2C; Supplemental Video 4). The delayed chorion hatching, diminished touch-evoked escape behaviors, and morphologic changes were highly indicative of decreased muscle function.
Severe muscle pathology, observed at both the light and electron microscopic levels, underlied the functional defects described above. We focused our analysis on muscle from 72 hpf embryos, as the muscle structure at this age is mature and greatly resembles that of human muscle. Light microscopic analysis of hematoxylin/eosin stained myotubularin morphant muscle revealed thin myofibers with abnormally located nuclei (**, Figure 3B). Analysis of semi-thin sections more dramatically illustrated these abnormal nuclei, which were mislocalized, large and filled with nucleoli of unusual appearance (Figure 3C). These findings are highly reminiscent of the nuclear abnormalities observed in human myotubular myopathy, shown in longitudinal section in Figure 3A.
We further characterized the perinuclear compartment using electron microscopy (Figure 4). Nuclei from myotubularin morphants were again found to be unusual in appearance (Figure 4B). The nuclei were surrounded by enlarged areas of disorganized cytoplasm which had a relatively paucity of normally appearing organelles. Higher magnification of the perinuclear compartment underscored the perinuclear changes, revealing abnormal mitochondria, areas lacking any organellar structure, and disorganized tubule-like structures (Figure 4C) In addition, some fibers contained large, bizarre, membranous structures of unclear origin (Figure 4D). This perinuclear disorganization is commonly observed in human myotubular myopathy muscle biopsies, and similar membranous structures have also been reported [22]. Of note is that we did not observe obvious vacuoles in the perinuclear area of any myofibers examined, which is contrary to what might be expected for a defect in endosomal trafficking.
The fact that myotubularin morphants had thin appearing muscle compartments by live image analysis (Figure 2) suggested that the muscle fibers may be hypotrophic as compared to controls. To examine this, we isolated myofibers from 72 hpf control and myotubularin morpholino injected embryos. Myofiber size was determined by calculating the area of myofibers stained by immunofluorescence with a myosin heavy chain (MHC) antibody. Myofibers from myotubularin morphants were significantly smaller than those from controls, measuring only 50% of control area (Figure 5A, B). The reduced size was not due to loss of myofiber structural integrity, as evidenced by the normal appearance of sarcomeric structures with MHC antibody labeling. Myofiber hypotrophy is an abnormality that is commonly observed in the muscle from myotubular myopathy patients [2].
One of the central questions related to myotubularin function is whether it has lipid phosphatase activity in vivo. To address this, we measured levels of PI3P, the primary lipid upon which myotubularin acts in vitro, in morpholino-injected embryos. For whole embryo measurements, we extracted total lipids and then used a lipid-protein-antibody overlay technique. When normalized to PI4P levels, the amount of PI3P detected in lipid preps from myotubularin morphants was not significantly different from the level in controls (Figure S3). The fact that overall PI3P levels were not changed was unsurprising considering that 7 other MTMRs with PI3P phosphatase activity are present in the fish embryo (see Figure S1).
Given that myotubularin is specifically required for muscle function, we next wanted to measure PI3P levels in muscle only. To accomplish this, we performed quantitative immunofluorescence on isolated myofibers using a PI3P antibody. Myotubularin morphant myofibers had readily observable increases in PI3P antibody staining, in particular in the perinuclear area (Figure 6A). We quantitated the pixel intensity of the perinuclear PI3P staining, and found that myotubularin morphants had levels 1.6 times higher than observed in controls (Figure 6B). This was consistent with a loss of myotubularin's phosphatase activity in the muscle, and provided evidence that myotubularin functions to regulate PI3P levels in muscle in vivo.
A potential explanation for the fact that PI3P levels are normal in the whole embryo but increased in muscle is that myotubularin is the sole or primary PI3P phosphatase in muscle while other MTMRs are expressed in other tissues. This question has been examined in murine myocytes by RT-PCR, and myotubularin was found to be the predominant phosphatase expressed in differentiated fibers [23]. We examined this question in the developing zebrafish using whole mount RNA in situ hybridization. We focused on the expression of myotubularin and its two most closely related homologs, MTMR1 and MTMR2. We found that between 24 hpf and 72 hpf, only myotubularin was expressed in muscle (Figure 7A and data not shown), supporting the idea that it is the primary PI3P phosphatase in that tissue.
We thus hypothesized that myotubularin knockdown results specifically in abnormalities in muscle because other functionally similar MTMRs are not expressed in muscle. To test this, we performed a series of gene rescue experiments (Figure 7B). We injected embryos with myotubularin morpholino and RNA from either myotubularin, MTMR1, or MTMR2 and measured the ability of embryos to hatch from their chorions by 60 hpf. In these experiments, the morpholino and the RNA are expressed ubiquitously. As expected, injection of morpholino alone caused a significant reduction in hatching (35% hatched; see also Figure 2B), while co-injection with full-length zebrafish myotubularin RNA resulted in significant amelioration of this hatching defect (71% hatched). Interestingly, co-injection of morpholino with either MTMR1 or MTMR2 RNA also restored the ability to hatch from the chorion. MTMR1 rescued hatching nearly to control levels (82%hatched; Figure 7B), while MTMR2 resulted in more modest improvement (55% hatched; Figure 7B). Therefore, these functionally similar MTMRs can compensate for the lack of myotubularin function in skeletal muscle.
A recent study on mouse myotubularin by Buj-Bello and colleagues reported localization of the protein to the T-tubule/sarcoplasmic reticulum junction [24]. We examined myotubularin subcellular localization in zebrafish myofibers, and determined by immunofluorescent analysis that the protein was expressed in a distinctive linear pattern that overlaps with that of the dihydropyridine receptor (DHPR), a marker for T-tubules (Figure 8). This pattern is thus consistent with localization to T-tubules. As expected, this staining was essentially undetectable in myofibers derived from myotubularin morphants (Figure S2).
Based on this localization, we were interested in the effect of myotubularin knockdown on T-tubule organization. We performed ultrastructural analysis using electron microscopy (Figure 9). Muscle from control morpholino injected embryos exhibited the normal pattern of T-tubules and sarcoplasmic reticulum (SR), with the SR coursing tightly through the sarcomeres and the T-tubules forming triads at regular periods. Conversely, muscle from myotubularin morpholino injected embryos had grossly aberrant SR and T-tubule networks (Figure 9). The SR networks were irregular, disorganized, and often randomly interspersed throughout the sarcomere. The T-tubule triads showed a range of abnormalities, from mild changes in electron density of the triad (arrow, upper right panel), to severe dilation of the triad structure (arrows, lower right panel), to fibers with essentially unrecognizable SR/triad areas (arrow, lower left panel).
We next determined if these ultrastructural changes corresponded to alterations in T-tubule function. We focused on excitation-contraction coupling, a process that requires intact T-tubules. We first verified that nervous system output to muscle was normal by assaying touch-evoked fictive swimming. To examine this, whole-cell voltage recordings were made from myofibers in vivo. In both control and myotubularin morphants, tactile stimulus resulted in rhythmic membrane depolarization in skeletal muscle (Figure S4). These data are consistent with intact output from the nervous system and through the neuromuscular junction [25],[26].
We then proceeded to study excitation-contraction coupling (Figure 10). This was accomplished by measuring the ability of myofibers to contract when exposed to depolarizing stimuli of progressively higher frequencies [27]. Employing this technique, we found that control myofibers consistently contracted at all stimuli up to 30 Hz, with the average maximal frequency equaling 27.0 Hz. Conversely, myofibers from myotubularin morphants exhibited increasing abnormalities above 10 Hz, with no myofibers able to contract to stimuli at 25 Hz and the average maximal frequency equaling only 11.5 Hz. This result is consistent with a defect in excitation-contraction coupling, and provides functional evidence to support the morphologic abnormalities observed in the T-tubules.
We were interested to correlate our findings with muscle from myotubular myopathy patient biopsies. T-tubule abnormalities have not been specifically mentioned in previous pathologic analyses from myotubular myopathy patients. We examined T-tubule organization in human biopsy samples using immunohistochemistry and antibodies to DHPRa1, a T-tubule marker, and to RYR1, a marker of the adjacent sarcoplasmic reticulum. A similar technical approach was successfully utilized by Laporte and colleagues to examine T-tubule organization in centronuclear myopathy patients with BIN1 mutations [28]. As a staining control, we used muscle from an unaffected, age matched control sample. DHPRa1 and RYR1 staining in the control muscle were found in the expected pattern along the membrane and throughout the cytoplasm (Figure 11A and Figure 11B, respectively). Conversely, samples from three patients revealed clear abnormalities in both DHPR and RYR1 staining patterns. T-tubules were found concentrated around the abnormally located central nuclei, or in irregular densities in the centers of several fibers. Importantly, other plasma membrane components were not found in this distribution (Figure S5), indicating that this disorganization is relatively specific for T-tubules.
We lastly examined electron micrographs obtained from patient muscle biopsies (Figure 12). Age matched control muscle showed the typical tight triad structure with well-organized adjacent sarcoplasmic reticulum. In contrast, micrographs from 3 myotubular myopathy patients showed various degrees of dilatation and disorganization of the T-tubules and adjacent sarcoplasmic reticulum. In conjunction with the immunostaining, these data confirm that T-tubule abnormalities are present in both our zebrafish model and in patients with myotubular myopathy.
We used antisense morpholinos to investigate the effect of myotubularin knockdown on zebrafish development. Our data from these studies illuminate several novel insights into myotubularin function/dysfunction. The first is that knockdown of zebrafish myotubularin recapitulates the features of myotubular myopathy, and thus demonstrates that zebrafish are an excellent model for studying the disease. The second is that closely related MTMRs can functionally compensate for the loss of myotubularin, suggesting that homolog upregulation is a viable therapeutic strategy in myotubular myopathy. The third is that T-tubule abnormalities are present in both zebrafish and human patients lacking myotubularin. As discussed below, T-tubule abnormalities are a unifying pathologic feature shared now by several congenital myopathies.
Myotubular myopathy is characterized clinically by the early onset of weakness and hypotonia, and pathologically by Type I fiber hypotrophy and the presence of centralized nuclei with abnormal appearance surrounded by areas of sarcoplasmic disorganization [1]. Zebrafish with reduced levels of myotubularin share all of these essential disease features. Embryos have defects in the earliest muscle dependent functional processes, including diminished spontaneous contractions and an inability to hatch from their chorions. The histopathology of myotubularin morphant fish closely mirrors the appearance of human myotubular myopathy muscle. Fibers are small (50% of control size) and have large, unusual and mislocalized nuclei surrounded by areas of sarcoplasmic disorganization. The perinuclear area also often contains accumulation of abnormal membranous structures; such structures have been reported in human ultrastructural analyses [22].
The myotubularin morphant zebrafish described here are now the second model system that recapitulates the “clinical” and pathologic features of myotubular myopathy by knocking down myotubularin levels during development. The other model is a mouse gene knockout generated by Buj-Bello, Laporte, Mandel and colleagues [17]. One interesting difference between our model and the knockout mice is the timing of the muscle phenotype. Our phenotype is present at a very early time point (essentially when primary myogenesis is completed), whereas the knockout mice have a period of normal development followed by precipitous degeneration. It is not clear which more accurately reflects the human disease, for while patients often have symptoms at birth, the ability to measure/detect in utero abnormalities in muscle function is difficult [1]. The difference between the two models may be reflective of the rapid and compacted development of the zebrafish. Conversely, it may be due to the fact that muscle maturation in the mouse continues for the first several postnatal weeks. Thus the difference may reflect the specifics of muscle development in the two organisms instead of intrinsic differences in myotubularin function in the species.
Our zebrafish model joins a growing list of myopathies and dystrophies that are successfully modeled in zebrafish [19], [20], [25], [29]–[32]. Given the large number of offspring that can be studied and the highly reproducible nature of our morphant phenotype, the myotubular myopathy zebrafish should provide an excellent springboard for high throughput testing of small molecule therapeutics.
One of the fundamental questions regarding myotubularin function was whether it behaved as a lipid phosphatase in vivo. We were able to address this question using our zebrafish model. Using quantitative immunohistochemistry, we demonstrate that PI3P, the principal substrate for myotubularin phosphatase activity, accumulates in myofibers from myotubularin morphant embryos. This is the predicted result from loss of myotubularin expression if it acts as a 3-position phosphatase. Significantly, these data are very consistent with the previously reported changes in PI3P levels found when myotubularin protein levels are reduced in cell culture or in yeast. We observed a 1.6 fold increase in PI3P in skeletal muscle, while Cao et al detected a 1.6 to 2 fold increase using RNAi in A431 cells and Dixon and colleagues found a 2 fold increase in ymr1 null yeast. Of note, our results represent one of the first assessments of potential phosphatase activity for any myotubularin family member in vivo and the first specifically in muscle.
Including myotubularin, 15 MTMRs are present in the vertebrate genome. All are expressed in zebrafish, mouse and man. Eight of the 15 have apparently identical phosphatase activity, with the remaining 7 are considered “phosphatase-dead” MTMRs [33]. Because myotubularin mutations result in severe muscle disease, it seems clear that none of the phosphatase active MTMRs compensate in myotubular myopathy patients [14]. It was not known whether this is due to unique non-phosphatase properties of myotubularin, or rather due to expression differences between MTMRs. Our data convincingly support the later conclusion. We show that MTMR1 and MTMR2, the MTMRs with the highest homology to myotubularin, are not expressed in zebrafish muscle. Furthermore, exogenous ubiquitous expression of either gene rescued the myotubularin morpholino phenotype. Importantly, expression of these MTMRs in control fish did not result in obvious phenotypic abnormalities. This implies that increasing the expression of either MTMR1 or MTMR2 in patient muscle is a viable potential treatment strategy for myotubular myopathy.
Perhaps the most significant finding from our study is that decreasing myotubularin expression or function results in both structural and functional abnormalities in the T-tubule network. This finding is significant for several reasons. The first is that it provides the first viable explanation for why patients (and mice and zebrafish) have significant weakness. T-tubules are critical for several aspects of muscle contractions and force generation, in particular for excitation-contraction coupling [34]. Impairment of this network should clearly lead to diminished force production and muscle weakness. We demonstrate this functionally in the zebrafish, as embryos with decreased myotubularin have excitation-contraction coupling abnormalities.
A second significance to these data is that they provide a plausible hypothesis for myotubularin function in myofibers. T-tubules biogenesis and maintenance is dependent on the continuous recycling of its membranous contents [15]. Membrane recycling is in turn dependent on tight regulation of phosphoinositides. Therefore, one possible explanation for our results is that myotubularin functions to regulate the recycling of T-tubule membrane components via its ability to participate in the regulation of phosphoinositide levels.
An association between T-tubule homeostasis and myotubularin is especially attractive given the potential functional similarities between T-tubule recycling and endosomal dynamics. Previous studies have shown that both endosomes and T-tubules share similar structural and regulatory components. Most notably, BIN1/amphiphysin2 and dynamin-2 are critical regulators of membrane trafficking at the endosome [35],[36], and both are expressed at the T-tubule [28]. In addition, BIN1 is required for T-tubule biogenesis in cultured myocytes and for T-tubule organization in Drosophila [37],[38]. As discussed below, mutations in both BIN1 and dynamin-2 result in centronuclear myopathy [1], a myopathy with similar pathologic features to myotubular myopathy. Such overlapping roles are also seen with caveoli, which are critical for both T-tubule formation/maintenance and for endocytosis [39],[40]. Thus, given the many observations functionally linking T-tubule dynamics and the regulation of endosomes, it seems very likely that myotubularin's primary function in muscle is controlling T-tubule dynamics in a fashion analogous to that described for its regulation of endosomal trafficking in vitro [10],[11].
A final importance relates to other congenital myopathies. Traditionally, congenital myopathies are considered a group of independent conditions, distinguished by their histopathologic features on muscle biopsy. However, they are in general similar in clinical presentation, characterized by neonatal hypotonia and non-progressive weakness [41]. The discovery of T-tubule abnormalities in myotubular myopathy now pathogenically links the three most prevalent groups of congenital myopathies. Core myopathies are caused by mutations in the ryanodine receptor (RYR1) [42], the calcium channel located at the T-tubule/sarcoplasmic reticulum border that is required for excitation-contraction coupling [25], and by mutations in Selenoprotein-N [43], a modifier of RYR1 [20]. Most nemaline myopathies are caused by mutations in the components of the thin filaments, proteins which function downstream of RYR1 and calcium release to initiate contraction [44]. Along with the centronuclear myopathies due to BIN1 (where T-tubule abnormalities have already been documented) and dynamin-2 mutation [1], myotubular myopathy likely is an “upstream” defect, resulting in abnormalities in the underlying T-tubule and sarcoplasmic reticular structure upon which RYR1 function is dependent.
In light of this commonality between congenital myopathies, the next important step is to see if modifiers of excitation-contraction coupling and T-tubule function can ameliorate the muscle weakness in the relevant disease models. We are currently at work developing and testing such agents in our zebrafish model of myotubular myopathy.
We have developed a new vertebrate model of myotubular myopathy, which has allowed us to answer fundamental questions regarding myotubularin function, and to make novel insights into the pathogenesis of the human disease. In the future, this model may provide a valuable platform for developing and testing novel therapeutics based on our new insights.
Morpholinos were designed to the putative ATG, to the exon 1 splice donor site, and to the exon 3 splice acceptor site of the zebrafish myotubularin gene (GeneTools, LTC). The control morpholino (GeneTools) was designed to random sequence with no homology by BLAST analysis in the zebrafish genome. The morpholino sequences are as follows:
1.5 nL of morpholino (0.08 mM) was injected into the yolk of 1–4 cell stage zebrafish embryos as described previously [18]. Embryos were subsequently grown in egg water and then analyzed at various time points. Western blot and RT-PCR analysis for determining morpholino efficacy were described previously [18].
Embryos were examined by live image analysis using a Leica stereomicroscope and camera. Both photomicrographs and videos were obtained using this system.
To measure spontaneous coiling, embryos were manually dechorionated at 24 hpf and recorded for 15 seconds. Records were obtained approximately 5 minutes after dechorionation.
Touch-evoked motor behaviors were elicited by touching the head, yolk sac or tail with a pair of No. 5 forceps. Motor behaviors were recorded by video microscopy (∼20×) using a Panasonic CCD camera (wv-BP330) mounted on a Leica dissection microscope. Videos captured (30 Hz) on a Macintosh G4 computer with a Scion LG-3 video card using Scion Image software (www.scioncorp.com) were processed with ImageJ.
For hematoxylin/eosin sections, 72 hpf embryos were fixed overnight at 4°C in 4% paraformaldehyde, washed in PBS, dehydrated in alcohols and xylenes, and embedded in paraffin. Microtome sections were cut at 2 mm. H/E was done per standard protocol. For semi thin sections and electron microscopic analysis, 72 hpf embryos were fixed overnight at 4°C in Karnovsky's fixative and then processed for sectioning by the Microscopy and Imaging Laboratory (MIL) core facility. Semi-thin sections were stained with Toludine blue. Light microscopy was performed using an Olympus BX-51 microscope and images captured with an Olympus DP-70 digital camera. Electron microscopy was performed using a Phillips CM-100 transmission electron microscope.
Mixed cell cultures from 72 hpf embryos were obtained as follows. Embryos were euthanised with tricaine and the dissociated in 10 mM collagenase type I (Sigma) for 60–90 min at room temperature. Embryos were triturated approximately every 30 min. Dissociated preps were pelleted by centrifugation (5 min at 3000 rpm), resuspended in CO2 independent media (Invitrogen), passed through a 70 mm filter (Falcon), and plated onto chamber slides (Falcon) precoated with poly-L-Lysine (Sigma). Culture media was changed after one hour, after which cells were fixed for 15 min in 4% paraformaldehyde.
Fixed cells were blocked in 10%NGS/0.3% Triton, incubated in primary antibody overnight at 4°C, washed in PBS, incubated in secondary antibody, washed in PBS, then mounted with ProLong Gold plus DAPI (Invitrogen). For PI3P antibody staining, cells were processed according to manufacturers protocol (Echelon Biosciences). The following primary antibodies and dilutions were used: mouse anti-myosin heavy chain (MF20; 1∶20; Developmental Hybridoma Bank); mouse anti-PI3P (1∶100; Echelon); rabbit anti-myotubularin (1∶50; Stratagene); and mouse anti-DHPR1a (1∶200; Abcam). Alexafluor conjugated secondary antibodies were used at 1∶250 (Invitrogen). Images were obtained by confocal microscopy as described previously [18].
Myofiber area was measured from photomicrographs using Metamorph software. Myofibers were outlined using the freehand tool and analyzed for total two-dimensional area.
PI3P antibody staining was performed as described above. Samples were analyzed on an Olympus IX-71 inverted confocal microscope and images captured using the FluoView v4.3 software. Fluorescent images were processed for quantitation using Metamorph (Sunnyvale, CA). Identical regions (immediate perinuclear area) were selected from each fiber using the rectangle tool set to a constant area. Boxed areas were then analyzed for pixel intensity. 15 myofibers from control and myotubularin morphant myofibers were compared for each region (5 per single myofiber prep×3 independent preps). Statistical significance was determined using a Students one-tailed t-test (Prism software) [45].
The lipid overlay assay was performed per manufacturer protocol for PI3P and PI4P on lipids extracted from 72 hpf embryos (Echelon Biosciences) [11].
In situ hybridization was performed as described previously [18]. Probes were made by in vitro transcription from zebrafish cDNA plasmids (all clones obtained from Open Biosystems).
RNA for morpholino rescue was prepared by in vitro transcription using the mMessage mMachine kit (Ambion). RNA was co-injected with morpholino at a concentration of 100 ng/ml. Rescue was determined by measuring the percentage of embryos hatched from their chorions at 60 hpf.
For in vivo electrophysiological measurements [46], larvae (72–80 hpf) were pinned in a Sylgard-coated petri dish (Dow Corning, Midland, MI) containing extracellular recording solution with curare [in mM:134 NaCl, 2.9 KCl, 2.1 CaCl2, 1.2 MgCl2, 10 glucose, 10 HEPES, pH 7.8 and 3 µM d-tubocurarine]. Larvae were manually skinned on one side, exposing muscle tissue. Electrodes were pulled from borosilicate glass and filled with internal recording solution [in mM: 116 K-gluconate, 16 KCl, 2 MgCl2, 10 HEPES, 10 EGTA, at pH 7.2 with 0.1% Sulforhodamine B for muscle cell type identification]. Whole-cell recordings were performed on individual adaxial myocytes using an Axopatch 200B amplifier (Axon Instruments, Union City, CA), low pass filtered at 1 kHz and sampled at 2–10 kHz. For each patch-clamped myocyte, steps of depolarizing current (3–6 nA) were injected to induce contraction. Current pulses were first delivered at a frequency of 1 Hz for 10 s. Frequency was increased by 5 Hz intervals until the myocyte reached tetanus. Contractions were recorded by video imaging and data acquired using a Digidata 1322A interface controlled by pClamp 8 software (Axon Instruments). Data analysis was performed using Clampfit 10.
Touch-evoked fictive swimming was elicited with a fire-polished recording electrode (∼50 µm) controlled by a Burleigh PCS-1000 piezoelectric manipulator and PCS-250 patch clamp driver (EXFO Life Sciences) as described previously [47] until fictive swimming was evoked.
Cryosections from human muscle biopsies were incubated overnight at 4°C in primary antibody (DHPRa1, 1∶200; RYR1, 1∶100; α-dystroglycan, 1∶50), washed in TBS, and then processed using the kit (Novacastra). Photomicrographs were obtained on an Olympus XL.
Western blot analysis was performed as previously described [48]. Rabbit anti-myotubularin (Stratagene) was used at 1∶1000 and anti-rabbit secondary (Santa Cruz Biotech) at 1∶2000. Goat anti-actin (Santa Cruz Biotech) was used at 1∶1000 and anti-goat secondary (Santa Cruz Biotech) at 1∶200. Luminenscent detection was performed using the Lumiglo reagent (Cell Signalling).
All animals were handled in strict accordance with good animal practice as defined by national and local animal welfare bodies, and all animal work was approved by the appropriate committee (UCUCA #09835).
|
10.1371/journal.pbio.1001564 | Control of Translation and miRNA-Dependent Repression by a Novel Poly(A) Binding Protein, hnRNP-Q | Translation control often operates via remodeling of messenger ribonucleoprotein particles. The poly(A) binding protein (PABP) simultaneously interacts with the 3′ poly(A) tail of the mRNA and the eukaryotic translation initiation factor 4G (eIF4G) to stimulate translation. PABP also promotes miRNA-dependent deadenylation and translational repression of target mRNAs. We demonstrate that isoform 2 of the mouse heterogeneous nuclear protein Q (hnRNP-Q2/SYNCRIP) binds poly(A) by default when PABP binding is inhibited. In addition, hnRNP-Q2 competes with PABP for binding to poly(A) in vitro. Depleting hnRNP-Q2 from translation extracts stimulates cap-dependent and IRES-mediated translation that is dependent on the PABP/poly(A) complex. Adding recombinant hnRNP-Q2 to the extracts inhibited translation in a poly(A) tail-dependent manner. The displacement of PABP from the poly(A) tail by hnRNP-Q2 impaired the association of eIF4E with the 5′ m7G cap structure of mRNA, resulting in the inhibition of 48S and 80S ribosome initiation complex formation. In mouse fibroblasts, silencing of hnRNP-Q2 stimulated translation. In addition, hnRNP-Q2 impeded let-7a miRNA-mediated deadenylation and repression of target mRNAs, which require PABP. Thus, by competing with PABP, hnRNP-Q2 plays important roles in the regulation of global translation and miRNA-mediated repression of specific mRNAs.
| The regulation of mRNA translation and stability is of paramount importance for almost every cellular function. In eukaryotes, the poly(A) binding protein (PABP) is a central regulator of both global and mRNA-specific translation. PABP simultaneously interacts with the 3′ poly(A) tail of the mRNA and the eukaryotic translation initiation factor 4G (eIF4G). These interactions circularize the mRNA and stimulate translation. PABP also regulates specific mRNAs by promoting miRNA-dependent deadenylation and translational repression. A key step in understanding PABP's functions is to identify factors that affect its association with the poly(A) tail. Here we show that the cytoplasmic isoform of the mouse heterogeneous nuclear ribonucleoprotein Q (hnRNP-Q2/SYNCRIP), which exhibits binding preference to poly(A), interacts with the poly(A) tail by default when PABP binding is inhibited. In addition, hnRNP-Q2 competes with PABP for binding to the poly(A) tail. Depleting hnRNP-Q2 stimulates translation in cell-free extracts and in cultured cells, in agreement with its function as translational repressor. In addition, hnRNP-Q2 impeded miRNA-mediated deadenylation and repression of target mRNAs, which requires PABP. Thus, competition from hnRNP-Q2 provides a novel mechanism by which multiple functions of PABP are regulated. This regulation could play important roles in various biological processes, such as development, viral infection, and human disease.
| Proteins that form dynamic multiprotein complexes with eukaryotic mRNAs play important roles in the control of gene expression [1]. The composition and architecture of messenger ribonucleoprotein particles (mRNPs) largely determines their distribution between different subcellular structures (i.e., polysomes, stress granules, or processing bodies), and ultimately the rates of mRNA translation and degradation [2]. Recent analysis revealed an unexpectedly broad repertoire of mammalian mRNA-binding proteins with largely unknown functions [3]. During translation initiation, which is the rate-limiting and most regulated step of protein synthesis, the 80S ribosome is recruited to the mRNA and positioned at the initiation codon [4]. This process is facilitated by the binding of eukaryotic initiation factor 4E (eIF4E) to the m7G cap structure at the 5′ end of the mRNA and poly(A) binding protein (PABP) to the 3′ poly(A) tail. The stimulatory effects of the cap structure and the poly(A) tail on translation are synergistic. eIF4E is a subunit of the eIF4F complex, which also includes eIF4A, an RNA-dependent ATPase/RNA helicase, and eIF4G, a high-molecular-weight scaffolding protein [5]. eIF4G interacts with PABP [6]–[8] and eIF3, which bridges between the eIF4G and 40S ribosomal subunit [9]. These interactions circularize the mRNA [10] and enhance translation (reviewed in [11]–[13]).
PABP is a highly evolutionarily conserved protein, which was first described, to our knowledge, four decades ago [14]. It contains four RNA recognition motifs (RRMs) and a proline-rich C-terminal region, which is involved in protein–protein interactions. PABP binds poly(A) in a cooperative manner and a periodicity of ∼27 nucleotides [15]. The 3′ end-associated PABP is a critical determinant of translational activity of an mRNA, which acts in cis [16]. Depletion of PABP from translation extracts decreases the binding of eIF4E to the cap structure and dramatically inhibits 48S and 80S ribosome initiation complex formation [17]. PABP might also play roles at the late step of initiation by promoting ribosomal subunit joining [18], and during termination and ribosome recycling by forming a complex with eRF3 [19]. These results underscore the importance of PABP for global translation. In a different role, PABP enhances the association of the microRNA-induced silencing complex (miRISC) with specific mRNAs to augment miRNA-mediated translation repression [20]. Finally, PABP regulates mRNA deadenylation, which is the first and generally rate-limiting step of mRNA degradation [21]. We and others recently showed that miRISC, which includes the Argonaute (AGO) and GW182 proteins [22]–[24], binds PABP (via GW182) and recruits CNOT7/CAF1 deadenylase to promote poly(A) tail shortening [25],[26]. These and other studies implicate PABP in mRNA-specific regulation of protein synthesis [27]. Intriguingly, PABP is a subject of posttranslational modifications, whose functional significance remains to be established [28].
Two PABP-interacting proteins, Paip1 and Paip2, modulate PABP activity in translation. Paip1 is a positive regulator of translation [29],[30]. In contrast, Paip2 inhibits translation by displacing PABP from the poly(A) tail and eIF4G [31],[32]. The dissociation of PABP from the poly(A) tail would be expected to remodel the mRNP. Here, we report the interaction of mouse heterogeneous nuclear ribonucleoprotein Q isoform 2 (hereafter referred to as hnRNP-Q2) with the poly(A) tail. We also show that hnRNP-Q2 competes with PABP for poly(A) binding to inhibit global protein synthesis both in vitro and in vivo and attenuate miRNA-mediated repression of mRNAs. These findings implicate hnRNP-Q2 in the control of the multifunctional activities of PABP.
Paip2 dramatically decreases the affinity of PABP for poly(A) in model systems containing poly(A) and pure recombinant proteins [31]. To examine the effect of Paip2 on PABP-poly(A) complex formation under more physiological conditions—that is, in the context of cell extracts—we used UV-induced crosslinking, which is a reliable technique to detect specific RNA–protein interactions [33]. For this assay, rabbit globin mRNA was 3′ end extended using [α-32P]ATP and yeast poly(A) polymerase. The mRNA was incubated with micrococcal nuclease-treated rabbit reticulocyte lysate (RRL), HeLa, or Krebs cell-free S10 extract, UV irradiated, and digested with a mixture of RNases. Proteins bound to the poly(A) tail were analyzed by SDS-PAGE and autoradiography. In all extracts, almost exclusive crosslinking of a ∼70 kDa polypeptide (p70) to poly(A) was observed (Figure 1A, B, C; lane 1). Adding Paip2 prevented the crosslinking of p70, indicating the dissociation of the p70-poly(A) complex (Figure 1A, B, C; lane 2). These results strongly suggest that p70 is PABP. Indeed, p70 was not detected in extracts that were depleted of PABP using GST-Paip2 affinity matrix (Figure 1A, B, C; lane 3) [34]. Surprisingly, we observed binding of novel proteins (p68 and p58/59) to the poly(A) tail following the sequestration of PABP by Paip2 or PABP depletion (Figure 1A, B, C; lanes 2, 3). Supplementing PABP-depleted extracts with recombinant PABP abolished the binding of these proteins to poly(A) (Figure 1A, B, C; lane 4). Interestingly, p68 appeared as a single prominent band in RRL that was supplemented with Paip2 or depleted of PABP (Figure 1A, lanes 2, 3). p68 was also detectable in control RRL, albeit as a faint band (Figure 1A, lane 1). Consequently, we wished to identify this protein.
To determine the identity of p68 in HeLa cells, a HeLa cytoplasmic extract was depleted of PABP and subjected to chromatography on poly(A)-Sepharose. After washing the beads with buffer containing 0.2 M KCl, poly(A) interacting proteins were sequentially eluted with buffers containing 1 M KCl and 2 M LiCl and resolved by SDS-PAGE (Figure S1A). In contrast to UV crosslinking, this assay captured many proteins, most likely because a great deal of them bind to the sugar-phosphate backbone of the RNA or other RNA-binding proteins. In agreement with the role of ionic RNA–protein interactions, the pattern of poly(A) binding proteins in a more stringent (2 M LiCl) wash was less complex than in 1 M KCl wash. Two candidates for p68 (polypeptides of ∼69 kDa and ∼70 kDa; bands 1 and 2, respectively) were excised from the gel, digested with trypsin, and analyzed by mass spectrometry. One protein in band 1 was identified as hnRNP-Q (UniProtKB accession number O60506), with 36.3% sequence coverage by 22 unique peptides (Table S1). Band 2 was identified as heat shock cognate 71 kDa protein (HSP7C, UniProtKB accession number P11142) with 50.3% sequence coverage by 24 unique peptides (unpublished data). To determine whether any of the two proteins is the poly(A)-crosslinkable p68, we performed immunoprecipitation of p68 from RRL following UV-induced crosslinking. A monoclonal antibody against hnRNP-Q (18E4) efficiently precipitated 32P-labeled p68 (Figure S1B). In contrast, two antibodies against hsp70 as well as an antibody against PABP failed to do so. The p68 protein of Krebs extract was also precipitated with the anti-hnRNP-Q antibody (unpublished data). These results demonstrate that p68 is identical to hnRNP-Q.
hnRNP-Q, also termed as NS1-Associated Protein-1 (NSAP1) [35], is an abundant and ubiquitously expressed protein [36] that has been assigned functions in pre-mRNA splicing and mRNA metabolism [37]–[40] as well as a role in IRES-mediated translation [41]–[45]. hnRNP-Q is highly homologous to hnRNP-R and contains an N-terminal acidic domain, three RRMs, and an RGG-rich C-terminal region, which may be involved in RNA binding and protein–protein interactions [37]. Multiple hnRNP-Q isoforms (seven in humans and two in mouse) are derived from alternative splicing of a single gene [37]. Posttranslational modifications of hnRNP-Q, which include phosphorylation and methylation, may determine its subcellular localization and RNA-binding properties [46],[47]. In mouse, the small (562 amino acid long) splicing variant of hnRNP-Q, referred to as SYNaptotagmin-binding Cytoplasmic RNA-Interacting Protein (SYNCRIP) or hnRNP-Q isoform 2 (hnRNP-Q2; accession number NP_062770.1), is mostly cytoplasmic, while the longer hnRNP-Q isoform 1 is in the nucleus [40],[46],[48]. The human ortholog of mouse hnRNP-Q2 is hnRNP-Q isoform 6 (hnRNP-Q6; accession number NP_001153149.1), whose sequence is identical to that of hnRNP-Q2 except for alanine instead of serine in position 357. Notably, the cytoplasmic isoforms of hnRNP-Q contain one, instead of two, nuclear localization signal [37]. We found that RRL contains a single isoform of hnRNP-Q that co-migrates with the smallest isoform of hnRNP-Q of Krebs or HeLa cells (Figure S1C). To determine whether the cytoplasmic hnRNP-Q isoform is associated with actively translated polysomal mRNAs or inactive mRNPs, a HeLa cytoplasmic extract was centrifuged through a sucrose density gradient. Proteins from each fraction were analyzed by Western blotting using antibodies against hnRNP-Q and PABP. Significantly, despite sharing high binding affinity for poly(A), hnRNP-Q and PABP differed with respect to their subcellular distribution. HnRNP-Q was present in the unbound protein/free mRNP fractions (along with eIF4A, eIF4E, and Paip2), as well as in the 40S ribosomal subunit fraction, but not in polysome fractions (Figure S2). In contrast, PABP and the mRNA packaging protein YB-1 [49] associated with polysomes in addition to their presence in the upper gradient fractions. These results indicate the association of hnRNP-Q with untranslated mRNPs, and are consistent with other studies on the subcellular distribution of NSAP1/hnRNP-Q, PABP, and YB-1 [42],[45],[50],[51].
Interestingly, it has been reported that SYNCRIP/hnRNP-Q2 exhibits preferential binding to poly(A) [48]. Consequently, we wished to investigate the poly(A) binding specificity of hnRNP-Q2 by performing RNA competition experiments in RRL. The addition of poly(A) (5 µg/ml) to a PABP-depleted RRL completely inhibited UV crosslinking of hnRNP-Q2 to the poly(A) tail (Figure 2A). Poly(G), poly(U), and 18S ribosomal RNA (rRNA) had no effect on the crosslinking of hnRNP-Q2, while poly(C) was slightly inhibitory. Poly(A) also specifically inhibited crosslinking of PABP to the poly(A) tail, as assayed in control (mock-depleted) RRL. Thus within RRL, hnRNP-Q2 exhibits preference for poly(A).
To quantitatively characterize the hnRNP-Q2/poly(A) interaction, a bacterially expressed His-tagged mouse hnRNP-Q2 was affinity purified by Ni2+-NTA agarose chromatography. SDS-PAGE and UV spectrum analyses revealed that the hnRNP-Q2 preparation is largely free of contaminating proteins and nucleic acids (Figure 2B and unpublished data). The recombinant hnRNP-Q2 was used in an electrophoretic mobility shift assay (EMSA), in which a constant small amount of 5′ 32P-labeled oligo(A30) was titrated with increasing amounts of hnRNP-Q2. Following incubation, RNA/protein complexes were separated from free oligo(A30) by native gel electrophoresis and quantified using a phosphorimager (Figure 2C). The apparent Kd for hnRNP-Q2 (which equals to the protein concentration at which 50% of the probe is shifted into a complex) was estimated to be ∼30 nM. Interestingly, the amount of the shifted probe gradually increased with the increased amounts of hnRNP-Q2, indicating that stable complex formation requires cooperative binding of hnRNP-Q2. Thus, hnRNP-Q2 binds avidly to oligo(A), albeit less strongly than PABP, for which a Kd of 4–7 nM has been reported [52],[53].
To determine whether hnRNP-Q2 competes with PABP for poly(A) binding, it was depleted from RRL using the anti-hnRNP-Q antibody. Western blotting revealed efficient (∼90%) depletion of hnRNP-Q2 but not β-actin, which served as a loading control (Figure 3A). In the absence of endogenous hnRNP-Q2, efficient crosslinking of PABP to the globin mRNA poly(A) tail occurred (Figure 3B). Adding increasing amounts of hnRNP-Q2 gradually diminished the crosslinking of PABP. In a reciprocal experiment, PABP-depleted RRL [54] was supplemented with increasing concentrations of recombinant PABP. In the absence of PABP, hnRNP-Q2 was bound by the poly(A) tail by default, appearing as a prominent band (Figure 3C). The hnRNP-Q2 band gradually faded away, while the PABP band intensified, when increasing concentrations of PABP were added to the reaction mixture. These results clearly demonstrate that PABP and hnRNP-Q2 compete with each other for poly(A) binding.
Having shown that hnRNP-Q2 and PABP compete for binding to the poly(A) tail, we predicted that hnRNP-Q2 would counteract PABP activity in translation. To investigate this, endogenous hnRNP-Q2 was immunodepleted from Krebs extract (∼90% depletion; Figure 4A). Depletion of hnRNP-Q2 stimulated the translation of capped and polyadenylated (A98) luciferase mRNA [designated as Cap-Luc-p(A)98 mRNA], by ∼3.5-fold (Figure 4B). The stimulatory effect of hnRNP-Q2 depletion on translation was not due to co-depletion of YB-1 (Figure 4A), an mRNA packaging protein and a general repressor of translation [49],[54],[55]. Adding back hnRNP-Q2 to the depleted extract decreased translation, and this inhibition was hnRNP-Q2 dose-dependent. It is noteworthy that the amounts of hnRNP-Q2 added in this and other assays were in the range of the concentrations normally found in Krebs extracts (∼30 µg/ml; Figure S3). To rule out the possibility that hnRNP-Q2 inhibits protein synthesis by destabilizing mRNA, the decay of 32P-labeled Cap-Luc-p(A)98 mRNA was monitored in translation extracts either containing or lacking hnRNP-Q2. Cap-Luc-p(A)98 mRNA remained intact in control and hnRNP-Q2-depleted translation extracts over a 2 h incubation period (Figure 4C). Furthermore, adding hnRNP-Q2 (30 µg/ml) to the depleted extract had no effect on the stability of Cap-Luc-p(A)98 mRNA.
We next examined the effect of hnRNP-Q2 on cap-independent translation driven by different viral internal ribosome binding sites (IRESs). The hnRNP-Q2-depleted extract was ∼2.2-fold more active than mock-depleted extract in supporting translation from the poliovirus (PV) IRES (Figure 4D). However, the translation from the hepatitis C virus (HCV) IRES (which is PABP and eIF4G-independent, in contrast to PV IRES) was not significantly augmented by hnRNP-Q depletion (Figure 4E). Consistent with these results, in hnRNP-Q2-depleted extract, PV IRES exhibited greater susceptibility to inhibition by recombinant hnRNP-Q2, as compared to HCV IRES (Figure 4, compare panels D and E). Thus, competition from hnRNP-Q2 does not significantly affect the function of ribosomes and translational factors other than the PABP/eIF4G complex.
To elucidate whether competition from hnRNP-Q2 targets the initiation step of translation, we examined ribosome binding using commercial nuclease-treated RRL. Although cap- and poly(A) tail dependence of RRL is decreased after nuclease treatment [56],[57], a significant dependence on these structures for translation is observed at low levels of input mRNA and elevated potassium ion concentrations (Figure S4) [54],[58]–[60]. For instance, at 60 mM additional KCl concentration, capping and polyadenylation enhance the translation of Luc mRNA (0.5 µg/ml) by 12.5- and 3.3-fold, respectively (Figure S4B, D). Therefore, the assays below were carried out using KCl (60 mM)-supplemented RRL and low (<0.5 µg/ml) mRNA concentrations. To investigate the formation of 80S ribosome initiation complex, RRL was incubated with radiolabeled globin mRNA in the presence of cycloheximide. The 80S complex was resolved from the unbound mRNA by sucrose gradient centrifugation. The addition of hnRNP-Q2 (20 µg/ml) to control or hnRNP-Q2-depleted RRL inhibited 80S initiation complex formation by 2.3-3-fold (Figures 5A, B). A similar reduction of 80S ribosome recruitment in the presence of hnRNP-Q2 was observed in normal or hnRNP-Q2-depleted Krebs extracts (Figure S5). To determine whether hnRNP-Q2 also targets 48S pre-initiation complex formation, 60S ribosomal subunit joining was inhibited by GMPPNP, a nonhydrolysable GTP analog [17]. In GMPPNP supplemented RRL, the labeled mRNA redistributed from the 80S fractions of the gradient to the 48S fractions, thereby validating the assay (Figure 5C). Importantly, adding hnRNP-Q2 (24 µg/ml) to hnRNP-Q2-depleted RRL inhibited 48S initiation complex formation by ∼5-fold with a profound shift of mRNA to the RNP fractions (Figure 5D). To determine whether hnRNP-Q2 inhibits translation prior to 48S complex formation, we examined the interaction of eIF4E with the cap-structure in mock- and hnRNP-Q2-depleted RRL by chemically crosslinking the lysates with polyadenylated Luc mRNA 32P-labeled at the 5′ cap-structure. We earlier showed that this assay provides a highly reliable measure of eIF4F activity [54]. In hnRNP-Q2-depleted RRL, eIF4E crosslinking was enhanced ∼1.5-fold relative to mock-depleted RRL (Figure 5E). Adding increasing concentrations of hnRNP-Q2 decreased crosslinking in a dose-dependent manner (to 25% of control). Thus, hnRNP-Q2 impairs the interaction of eIF4E with the cap-structure. Since PABP stimulates eIF4E-cap interaction [17],[61], it is most probable that hnRNP-Q2 acts by inhibiting this function of PABP. To gain evidence that hnRNP-Q2 targets eIF4-group factors in Krebs extract, purified eIF4F, eIF4A, eIF4E, and eIF4B were added to this system either lacking or containing hnRNP-Q2. These factors stimulated the translation of Cap-Luc-p(A)98 mRNA, consistent with their presence in limiting amounts in Krebs extracts (Figure S6 and [62]). In agreement with the partial repression of eIF4F activity by hnRNP-Q2, exogenous eIF4F relieved hnRNP-Q2-mediated translation inhibition (from 5.5- to 1.4-fold). eIF4A, eIF4E, and eIF4B also relieved the inhibition of translation by hnRNP-Q2, albeit less efficiently than eIF4F.
The length of the poly(A) tail determines the number of PABP molecules bound to an mRNA, thereby indirectly controlling PABP-dependent translation. To study whether hnRNP-Q2-mediated inhibition of translation is dependent on the length of the poly(A) tail, we compared the effect of hnRNP-Q2 on the translation of Luc mRNAs either without (A0) or with a poly(A) tail of increasing length (A15, A30, A45, A90, and A250). For these studies we used Krebs extract that was not nuclease-treated. In several studies, the omission of nuclease treatment during the preparation of extracts has been proven ideal for mimicking cap-poly(A) synergy and other translational control mechanisms operating in vivo [56],[63]–[65]. As shown in Figure 6A, the untreated extract was strikingly poly(A) tail dependent, exhibiting up to 20-fold stimulation of translation by poly(A) tailing. Importantly, adding hnRNP-Q2 to the extract inhibited the translation of mRNAs with long (90–250 nt) poly(A) tails more strongly (3.2–3.6-fold) than the translation of the mRNA with short (15–30 nt) poly(A) tails (1.5–1.8-fold), while having a marginal effect on the translation of the poly(A-) mRNA (1.2-fold inhibition). To assure that the displacement of PABP from the poly(A) tail is required for the hnRNP-Q2-mediated translational inhibition, PABP was sequestered into the PABP–Paip2 complex. In the presence of Paip2, low-efficient PABP-independent translation was virtually insensitive to inhibition by hnRNP-Q2 (Figure 6B). Inactivation of PABP also abolished the response of translation to poly(A) length. Thus, the translational inhibition by hnRNP-Q2 in the nuclease untreated extract is both PABP and poly(A) tail dependent. To determine to what extent endogenous hnRNP-Q2 inhibits the translation of polyadenylated mRNAs, we attempted to deplete the untreated extract of hnRNP-Q2. However, we failed to achieve substantial immunodepletion of hnRNP-Q2 using the 18E4 antibody (unpublished data). It is possible that hnRNP-Q2 cannot interact with this antibody when bound to mRNA that is in untreated extract. Since the assays above employed the extract that was not nuclease treated, it was of interest to test the effect of hnRNP-Q2 on the translation of endogenous mRNAs (Figure 6C). HnRNP-Q2 reduced 35S-methionine incorporation in the untreated extract in a dose-dependent manner. However, this inhibition was relatively small (up to 1.4-fold), as compared to that of the exogenous mRNA translation (Figure 6A). It is likely that re-initiation of translation in the untreated extract is less efficient than in intact cells; hence, 35S-methionine incorporation primarily reflects the rate of polypeptide chain elongation on the preformed polysomes. Consistent with this notion, inhibiting re-initiation of translation with hippuristanol [66] only modestly (1.8-fold) reduced 35S-methionine incorporation (Figure 6C).
Next, we investigated whether hnRNP-Q2 inhibits protein synthesis in vivo by reducing the amount of hnRNP-Q2 in the mouse fibroblast-like cell line L929 using shRNA. One shRNA against hnRNP-Q (shRNA1) caused significant silencing of hnRNP-Q2 (∼90%; Figure 7A, B). Another shRNA (shRNA2) was less effective in hnRNP-Q2 silencing (∼75%). No changes in the levels of PABP, eIF4GI, eIF4A, and eIF4E were found. Overall translation rate was measured by [35S]methionine/cysteine incorporation into newly synthesized proteins. shRNA1 expressing cells showed a ∼2-fold increase in incorporation as compared to cells expressing nontargeting control shRNA (Figure 7C). hnRNP-Q knockdown by shRNA2 evoked less potent stimulation of translation, as compared to shRNA1 (∼1.4-fold). SDS-PAGE analysis of newly synthesized proteins indicated that hnRNP-Q2 inhibits global protein synthesis (Figure 7D).
MiRNAs, in addition to inhibiting translation, mediate deadenylation and decay of target mRNAs [67]. PABP facilitates miRNA-dependent deadenylation through its interaction with the GW182-CAF1/CCR4 deadenylase complex [25],[68]. We wished to determine whether hnRNP-Q2 antagonizes this function of PABP in a Krebs extract, which faithfully recapitulates PABP-dependent miRNA-mediated deadenylation [25]. An RNA bearing six let-7a targets sites and a 98 nucleotide long poly(A) sequence (6xB-3′UTR RNA), labeled uniformly with 32P UTP [25], was extensively deadenylated by Krebs extracts (Figures S7 and 8). The completely deadenylated (A0) RNA is likely unstable as it appeared as a less prominent band as compared to input (A98) RNA. The formation of the A0 RNA species was dependent on let-7a miRNA as it was blocked by the addition of anti-let-7a 2′-O-methylated oligonucleotide (2′-O-Me) and also not observed with a reporter bearing mutations in nucleotides complementary to the let-7a “seed” sequence (6xBMut-3′UTR RNA) (Figure S7). To investigate how deadenylation is affected by competition between PABP and hnRNP-Q2, hnRNP-Q2 was added to untreated or nuclease-treated Krebs extracts and the kinetics of deadenylation of 6xB-3′UTR RNA was followed. As expected, exogenous hnRNP-Q2 inhibited the conversion of the full-length A98 RNA into A0 RNA (Figure 8A, B). The effect of hnRNP-Q2 on deadenylation was somewhat stronger in the untreated extract, most likely because a fraction of PABP is withdrawn from competition as being sequestered by endogenous mRNAs. To better assess the effects of PABP and hnRNP-Q2 on deadenylation, we made use of extracts that were depleted of these proteins. In hnRNP-Q2-depleted extract, the full-length A98 RNA disappeared and the A0 RNA was formed within a 2-h incubation period (Figure 8C). Consistent with the importance of PABP for poly(A) tail shortening [25], the addition of Paip2 almost abrogated deadenylation. Importantly, adding hnRNP-Q2 markedly impaired deadenylation, as ∼30% of RNA retained the full-length poly(A) tail after 2 h of incubation. This demonstrates that competition from hnRNP-Q2 inhibits miRNA-mediated deadenylation. To determine whether hnRNP-Q2 also interferes with the function of exogenous PABP in miRNA-mediated deadenylation, the assay was carried out in Krebs extract devoid of both hnRNP-Q2 and PABP. In this extract, a vast proportion of the input RNA retained the poly(A) tail during the time course of reaction (Figure 8D). However, almost all the RNA became deadenylated within 2 h after addition of recombinant PABP. Importantly, adding back hnRNP-Q2 to the PABP-supplemented extract markedly decreased the rate of deadenylation. Under these conditions a significant fraction of RNA (∼25%) remained intact even after 3 h of incubation. In addition, the feeble deadenylation of the RNA in PABP and hnRNP-Q2 double-depleted extract (which could be due to incomplete depletion of PABP) was prevented by the addition of recombinant hnRNP-Q2. Taken together, these results demonstrate that hnRNP-Q2 stabilizes mRNAs by antagonizing PABP activity in miRNA-mediated deadenylation.
Next, we examined whether hnRNP-Q2 reduces miRNA-induced repression in vivo. Control and hnRNP-Q2 knockdown L929 cells were transfected with Renilla luciferase (RL) reporters, with or without six let-7a miRNA target sites (6xB) [69]. A firefly reporter (FL) was co-transfected to normalize for transfection efficiency. In control cells, the expression of RL-6xB was ∼4-fold lower than RL (Figure 9A, B). Importantly, hnRNP-Q2 knockdown significantly augmented this inhibition (from 4-fold to 8.1-fold). Co-transfection of anti-let-7a 2′-O-Me oligonucleotide, but not control anti-miR-122a oligonucleotide, dramatically reduced the inhibition of RL-6xB expression, consistent with the role of let-7a miRNA in silencing of the RL-6xB reporter (Figure 9A). We determined the amount of RL-6xB mRNA to be ∼2.6-fold and ∼4-fold lower than RL mRNA in control and hnRNP-Q knockdown cells, respectively (Figure 9C, top; compare lane 2 with 1 and lane 8 with 7; Figure 9D). This difference in the relative RL-6xB levels can partially explain the augmented reduction of expression of RL-6xB reporter after hnRNP-Q2 depletion (Figure 9B). Attesting to the dependence of RL-6xB mRNA decay on let-7a miRNA, the levels of RL-6xB mRNA were rescued by co-transfection of anti-let-7a, but not anti-miR-122a, 2′-O-Me oligonucleotide (Figure 9C).
In eukaryotic cells, the association of PABP with the poly(A) tail stimulates global translation [12],[17], but also promotes miRNA-dependent deadenylation and repression of target mRNAs [20],[25],[70]. Paip2 inhibits these functions of PABP by dissociating the PABP-poly(A) complex [31]. In this study, we applied UV-induced crosslinking to characterize the composition of the poly(A) mRNP in the absence of PABP. Upon UV irradiation, proteins crosslink to poly(A) when bound in proximity to the photochemically reactive purine rings [33],[71]. In contrast, ionic interactions of proteins with the sugar-phosphate backbone of poly(A) do not fulfill the requirement for crosslinking. In addition, UV irradiation does not cause protein–protein crosslinking. Thus, UV-induced crosslinking is a reliable technique in revealing specific protein–poly(A) interactions. In all cell extracts studied, PABP appeared as the single major poly(A)-binding protein and Paip2 decreased the association of PABP with poly(A) (Figure 1 and [31],[71]). When PABP was depleted from RRL, hnRNP-Q2 became the major poly(A) binding protein by default. PABP and hnRNP-Q2 are presumably the only strong poly(A) binders since RRL depleted of both PABP and hnRNP-Q2 produced no major cross-links (unpublished data). To our knowledge, the first description of a cytoplasmic poly(A) interacting protein (p78X) that is distinct from PABP dates back to 1981 [72]. At the time, the identity of this protein and its function has not been explored. Along with hnRNP-Q2/Q6, p58/59 crosslinked with poly(A) in PABP-depleted Krebs and HeLa extracts (Figure 1B, C). This protein(s) might be similar or identical to the nuclear poly(A)-associated protein p60A with as yet unidentified function [72]. The absence of p58/59 from rabbit reticulocytes, which lack nuclei, favors this possibility. The leakage of p58/59 from the nucleus might have occurred during extract preparation, as these proteins are especially abundant in extracts derived from excessively disrupted cells (unpublished data).
How important is hnRNP-Q2 for mRNA translation and metabolism? Preferential binding to poly(A) distinguishes hnRNP-Q2 from the bulk of general RNA-binding proteins, which do not exhibit sequence specificity [73]. Moreover, hnRNP-Q2 competed with PABP for binding to the poly(A) tail of the mRNA. This competition would be expected to impair multiple functions of PABP in global and mRNA-specific regulation of protein synthesis. In agreement with this prediction, we showed that hnRNP-Q2 inhibits the initiation of translation that requires the PABP/eIF4G complex. In addition, hnRNP-Q inhibited miRNA-mediated deadenylation and repression of mRNAs that are promoted by PABP.
A paramount issue in addressing the competition between hnRNP-Q2 and PABP for poly(A) binding in vivo concerns the relative abundance of these proteins in the cell. The concentration of hnRNP-Q2 in Krebs and RRL translation mixtures is ∼1.7-fold higher than that reported for PABP (240–480 nM versus 140–280 nM, Figure S3, and [25],[54]). Since the affinity of hnRNP-Q2 for poly(A) is ∼6-fold lower than that of PABP, its molar excess over PABP might not be sufficient for efficient competition under standard physiological conditions. However, a significant fraction of PABP might be sequestered into complexes with repressor proteins, such as Paip2. This would not only increase the hnRNP-Q2/PABP ratio but also impair PABP cooperative binding that is important for the stability of the PABP/poly(A) complex. Finally, the cytosolic hnRNP-Q levels are likely elevated in the G2/M phases of cell cycle and under stress conditions [44],[74]. As a result of these rearrangements, conditions for efficient competition from hnRNP-Q could be met. The observed stimulatory effects of hnRNP-Q2 depletion on translation both in vitro and in vivo indicate that the endogenous concentration of hnRNP-Q2 suffices for translational inhibition. Although SDS-PAGE analysis of proteins de novo indicates that hnRNP-Q2 targets global protein synthesis (Figure 7D), it might also differentially affect the translation of specific mRNAs. Pointing to this possibility is the negative regulation of RhoA mRNA translation by the cytoplasmic isoform of hnRNP-Q [36], and the presence of NSAP1/hnRNP-Q in a translational silencing complex that recognizes a specific element in the 3′UTR of inflammatory mRNAs (termed IFN-γ-Activated Inhibitor of Translation, or GAIT, element) [75],[76]. On the other hand, binding of hnRNP-Q to several IRES elements stimulates translation [41]–[45]. In addition, hnRNP-Q can possibly activate IRESs indirectly by reducing competition from the bulk of cellular mRNA. Finally, as shown here and discussed below, the displacement of PABP from the poly(A) tail by hnRNP-Q2 attenuates miRNA-induced deadenylation, decay, and repression of specific mRNAs. Thus, in addition to its function as a general translation repressor, hnRNP-Q might play divergent roles in mRNA-selective translational control.
Translationally repressed mRNAs accumulate in two cytoplasmic foci: processing bodies and stress granules, which serve as sites for mRNA degradation or storage [77],[78]. It is conceivable that once bound by hnRNP-Q2, the mRNA is guided to cytoplasmic granules. Indeed, in neurons, SYNCRIP/hnRNP-Q2 localizes to mRNA granules that are transported along dendrites [79]. In addition, in stressed cells, hnRNP-Q relocalizes to cytoplasmic granules, as evidenced by its co-localization with HSP70, GW182, and TIA-1 marker proteins [74]. In both types of granules mRNA translation is inhibited [2]. It is an intriguing possibility that hnRNP-Q2 plays a role in this inhibition.
Deadenylation and subsequent decrease of target mRNA levels significantly contributes to miRNA-induced reduction of gene expression [67],[70]. PABP interacts with the GW182 proteins, which are essential components of the miRISC [25]. This interaction promotes miRNA-dependent deadenylation, potentially by bringing the poly(A) tail in proximity to miRISC-associated CAF1/CCR4 deadenylase. HnRNP-Q2 markedly impaired PABP-dependent let-7a miRNA-mediated deadenylation in Krebs extract, most probably by partially displacing PABP from the poly(A) complex (Figure 8). In a more physiological context, L929 cells, hnRNP-Q2 depletion augmented the miRNA-dependent degradation and repression of a target mRNA (Figure 9). Interestingly, the expression of RL-6xB reporter was stronger affected by hnRNP-Q2 depletion at the level of protein than mRNA. Thus, it is likely that hnRNP-Q2 also targets the function of PABP in miRNA-mediated translational repression [20]. It is noteworthy that the role of hnRNP-Q2/NSAP1/SYNCRIP in mRNA stabilization is also suggested by its presence in protein complexes that stabilize c-fos and c-myc proto-oncogene mRNAs [38],[80]. Thus, competition from hnRNP-Q provides a novel mechanism by which multiple functions of PABP are regulated. Control of PABP functions by hnRNP-Q2 could play important roles in various biological processes, such as development, virus infection, and human disease.
Recombinant PABP, GST-Paip2, eIF4A, eIF4E, and eIF4B were expressed and purified as described [17],[31],[58],[81]. Native eIF4F was purified from RRL [81]. The proteins were dialyzed against buffer A containing 20 mM Tris-HCl, pH 7.5, 100 mM KCl, 0.1 mM EDTA, 1 mM DTT, and 10% glycerol. Mouse monoclonal anti-hnRNP-Q (clone 18E4) antibody and anti-FLAG antibody used for the preparation of hnRNP-Q2-depleted and mock-depleted Krebs extracts, respectively, were from Sigma. For the description of antibodies used for Western blotting and immunoprecipitation, see corresponding sections below.
The expression vector for mouse hnRNP-Q2 was constructed as follows. The cDNA encoding hnRNP-Q2 (accession number GI:114145481) was obtained by reverse transcription-PCR (RT-PCR) of Krebs-2 cell poly(A)+ RNA using QIAGEN OneStep RT-PCR kit. The hnRNP-Q2 coding DNA fragment was amplified with forward (GATATACCATGGCTACAGAACATGTTAATGGAAATGGTACTGAAGAGCCTATGGATACTACTTCAGCAG) and reverse (GTGGTGCTCGAGTTGTAACAGGTCAGGACCGGCCTCG) primers, designed to generate the 5′ terminal NcoI and 3′ terminal XhoI restriction sites and eliminate an internal NcoI site by introducing a silent mutation (the flanking sequences for cloning purposes and silent mutation are underlined). After digestion with NcoI and XhoI, the hnRNP-Q2 coding DNA fragment was cloned into NcoI-XhoI sites of pET28a (Novagen) to generate a vector for the expression of hnRNP-Q2 with a six-His sequence at the C-terminus (pET28a-hnRNP-Q2-His). The His-tagged hnRNP-Q2 protein was expressed in Escherichia coli and purified to apparent homogeneity by Ni2+-nitrilotriacetic acid agarose chromatography (QIAGEN) using a batch procedure. Briefly, frozen bacterial cells were suspended in suspension buffer (S) (20 mM HEPES, pH 7.5, 5 mM 2-mercaptoethanol, and 10% glycerol) containing 2 M KCl and Complete EDTA-free protease inhibitor cocktail (Roche) and lysed by sonication. Following addition of Triton X-100 (to 0.1% final concentration), cell debris was removed by centrifugation (40,000 g, 1 h, and 4°C). The supernatant was supplemented with 20 mM imidazole, pH 7.5 and applied to Ni2+-NTA agarose resin equilibrated with buffer S containing 20 mM imidazole, 2 M KCl, and 0.1% Triton X-100. The beads were washed first with the same buffer and then with buffer S containing 20 mM imidazole, 0.1 M KCl, and 0.1% Triton X-100. Bound proteins were eluted with buffer S containing 250 mM imidazole and 0.1 M KCl and dialyzed against buffer A.
Plasmids for transcription of Luc mRNAs with 98-nucleotide long poly(A) tails, T3luc(A)+, T7PVluc(A)+, and T7HCVluc(A)+ [82], were linearized with BamHI and transcribed with T3 or T7 RNA polymerase, as appropriate. The templates for transcription of Luc mRNAs with variable poly(A) tails were obtained by PCR using plasmid T3luc [82], forward primer GCTCGAAATTAACCCTCACTAAAGGG, and five different reverse primers, collectively named as (T)nGGATCCCCCGGGCTGC, where (T)n designates oligo(dT) tracts of 0, 15, 30, 45, and 90 nucleotides (the core T3 promoter sequence is underlined). After purification on a Chroma Spin-1000 column (BD Biosciences), the PCR products were transcribed with T3 RNA polymerase. Luc mRNA with the poly(A) tail of ∼250 nucleotides was obtained by polyadenylation of the Luc-A98 mRNA using yeast poly(A) polymerase (USB) as recommended by the manufacturer. Capping of mRNAs was done using the ScriptCap m7G capping system (Epicentre). The integrity of mRNAs was verified by denaturing agarose gel electrophoresis.
Synthetic RNA oligonucleotide (A30, Dharmacon) was 5′ labeled using [γ-32P]ATP and polynucleotide kinase. Prior to use, the probe was purified by centrifugation through a Chroma Spin-10 column. Standard binding reaction mixtures (20 µl) contained 8 fmol (∼40,000 cpm) of 5′-labeled A30, 10 µl of 2× incubation buffer (40 mM HEPES-KOH, pH 7.3, 200 mM KCl, 4 mM MgCl2, 2 mM DTT, 0.1% NP40, 10% glycerol, and 0.2 mg/ml acetylated bovine serum albumin), and 2 µl of hnRNP-Q2 diluted to the appropriate concentrations with buffer A. Following incubation at 30°C for 30 min, the samples were supplemented with 2 µl of 50% v/v glycerol and analyzed by electrophoresis in 7% nondenaturing polyacrylamide gel (prepared with TBE buffer and 5% v/v glycerol) at 4°C. Bands were visualized by autoradiography. The amount of free and bound RNA was determined using a Typhoon Phosphorimager (GE Healthcare).
Krebs extracts, untreated or treated with micrococcal nuclease, were prepared as described previously [34]. Where indicated, the nuclease-treated extracts were depleted or mock depleted of hnRNP-Q2 and PABP as described below. The reaction mixtures (12 µl) included Krebs extracts (50% by volume), salts, essential translation components [34], indicated mRNAs (0.2 µg/ml, unless specified in the figure legends), and proteins. For optimal translation of HCV and PV IRES-containing mRNAs, the concentration of KOAc in the reaction mixtures was increased by 75 mM. Incubation was at 32°C for 1 h. Luciferase activity was determined in 1 µl aliquots of samples using Luciferase assay system (Promega) and Lumat LB 9507 bioluminometer (EG&G Bertold). The relative light units (RLU) reported are averages of three assays with the standard deviation from the mean.
Prior to hnRNP-Q2 depletion, a nuclease-treated Krebs extract was supplemented with salts, amino acids, and energy generating system as described previously [83]. The supplemented Krebs extract or RRL (Promega) were clarified by centrifugation at 10,000 g for 1 min. To couple the antibody against hnRNP-Q to beads, 40 µg (∼20 µl) of anti-hnRNP-Q (18E4) were incubated with a Protein-G Sepharose slurry (GE Healthcare; 150 µl pelleted beads per 0.6 ml of phosphate buffered saline [PBS; 137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, and 2 mM KH2PO4]) at 4°C for 1.5 h while mixing on a rotator. In a control tube, the beads were similarly incubated with 40 µg of an anti-FLAG antibody. Following incubation, the beads were washed by centrifugation, once with PBS containing 1% bovine serum albumin and twice with buffer D (25 mM HEPES-KOH, pH 7.3, 50 mM KCl, 75 mM KOAc, and 2 mM MgCl2) [34]. After the final centrifugation at 2,400 g for 2 min, the bead pellets were suspended in 600 µl of Krebs extract (or RRL), already containing the necessary translation components. After gentle agitation for 1.5 h, the beads were precipitated by centrifugation as above. The supernatants, which constitute the hnRNP-Q2 and mock-depleted extracts, were collected, centrifuged again to remove any residual beads, and frozen at −80°C in small aliquots. For the depletion of PABP, the supplemented Krebs extract or RRL was incubated with the GST-Paip2 protein that was immobilized onto glutathione-Sepharose beads [34]. Mock-depleted extracts were treated with GST alone. For the removal of both hnRNP-Q2 and PABP, the extracts were incubated first with anti-hnRNP-Q and then with GST-Paip2. The efficiency of hnRNP-Q2 and PABP depletion was analyzed by Western blotting.
The poly(A) tail of rabbit globin mRNA (1.5 µg; Gibco BRL, discontinued) was extended in a 50 µl total reaction volume using [α-32P]ATP (60 µCi, 3,000 Ci/mmol, Perkin Elmer) and yeast poly(A) polymerase (1,500 U, USB) as recommended by the manufacturer. Incubation was at 37°C for 30 min. After extraction with a mixture of phenol and chloroform, the RNA was purified by Chroma Spin-100 column chromatography. Reaction mixtures (15 µl) contained 10 µl of RRL (or Krebs extract supplemented with essential translational components), labeled RNA (∼400,000 cpm), and other components as indicated in the figure legends. After incubation at 30°C for 10 min, the samples were applied drop-wise along a line onto a Parafilm-covered glass plate and irradiated with UV light at 254 nm (using a 15-W germicidal lamp with ∼4 cm distance between the lamp and the samples) for 20 min on ice. The samples were collected into tubes already containing 4 µl of an RNase cocktail (1 mg/ml RNase A and 1,000 U/ml [∼0.7 mg/ml] nuclease S7 [Roche]). One µl of 200 mM CaCl2 was then added, and the samples were digested at 37°C for 30 min. Proteins were denatured by adding 40 µl of 1.5× SDS sample buffer and heating at 95°C for 5 min. Samples were analyzed by SDS-PAGE (10% acrylamide, 99∶1 acrylamide/N, N′-methylenebisacrylamide ratio) and autoradiography at −80°C with an intensifying screen.
Primary antibodies were the following: mouse monoclonal anti-hnRNP-Q antibody (clone 18E4, Sigma), rabbit polyclonal anti-PABP antibody [31], rabbit polyclonal anti-Paip2 antibody (Sigma), rabbit polyclonal anti-YB-1 (Abcam), rabbit polyclonal anti-eIF4GI antibody [84], mouse monoclonal anti-eIF4A antibody [85], mouse monoclonal anti-eIF4E antibody (BD Biosciences), mouse monoclonal anti-ribosomal protein S6 antibody (Santa Cruz), and mouse monoclonal anti-β-actin antibody (Sigma). Proteins in the samples were resolved by SDS-PAGE, transferred to a nitrocellulose membrane, and detected using Western Lightning chemiluminescence kit (Perkin-Elmer Life Sciences). Primary antibodies against hnRNP-Q, PABP, eIF4GI, YB-1, and β-actin were used diluted 1∶2,500, 1∶1,000, 1∶1,000, 1∶1,000, and 1∶5,000, respectively. The dilutions of antibodies against Paip2, eIF4E, and ribosomal protein S6 were as per the manufacturers' instructions. Secondary HRP-conjugated anti-mouse or anti-rabbit antibody (GE Healthcare), as appropriate, was used diluted 1∶5,000. Typically, a single membrane was probed, exposed, and stripped before probing with another antibody.
Anti-hnRNP-Q and anti-PABP antibodies that were used for Western blotting were also used for immunoprecipitation of p68. Mouse monoclonal and rabbit polyclonal anti-hsp70 antibodies were from Santa Cruz and Calbiochem, respectively. To generate 32P-labeled p68, PABP-depleted RRL was UV crosslinked with the 32P poly(A) tail-labeled globin mRNA in twenty 15-µl aliquots. The aliquots were combined and treated with RNAses. After the addition of 10% SDS to 0.2% final concentrations, the crosslinked RRL was 10-fold diluted with PBS containing 0.2% NP40. For immunoprecipitation, 0.7 ml portions of the diluted reaction mixture were incubated with antibody-conjugated protein-G Sepharose beads (20 µl) at 4°C for 4 h while mixing on a rotator. The beads were washed three times with 0.2% NP-40 containing PBS and finally with PBS alone. Bound proteins were eluted by heating in SDS-sample buffer and analyzed by SDS-PAGE and autoradiography. UV crosslinked control and PABP-depleted RRL were loaded on the same gel for comparison.
HeLa S3 cells were grown in a 15-cm Petri dish to ∼90% confluence. A polysome profile was obtained after centrifugation of a fresh cellular extract through a 10%–50% sucrose density gradient according to standard methods [86]. Centrifugation was in a Beckman SW41Ti rotor at 35,000 rpm for 2.5 h at 4°C. Optical density at 254 nM was continuously recorded using an ISCO fractionator (Teledyne ISCO, Lincoln, NE). Aliquots of fractions (30 µl) were analyzed by Western blotting using antibodies against hnRNP-Q, PABP, Paip2, YB-1, eIF4A, eIF4E, and 40S ribosomal protein S6.
For 80S ribosome binding studies, 32P-poly(A)-labeled globin mRNA (∼300,000 cpm, 6 ng) was incubated in a total reaction volume of 30 µl with nuclease-treated Krebs extract or KCl (60 mM)-supplemented RRL in the presence of the translation components and 0.6 mM cycloheximide [17]. Where indicated, recombinant hnRNP-Q2 or control buffer were added to the reaction mixtures. Incubation was at 32°C for 15 min. Reactions were stopped by 4-fold dilution with ice-cold HSB buffer [54], and 80S ribosomal complexes were resolved by centrifugation in 5-ml 15%–30% sucrose gradients (Beckman SW55 rotor, 54,000 rpm for 1 h, 45 min at 4°C). Fractions (0.2 ml) were collected from the top of the gradients and analyzed by scintillation counting. 48S complexes were formed in RRL in the presence of GMPPNP (2 mM), MgCl2 (2 mM), and cycloheximide (0.6 mM) [17]. Prior to the addition of mRNA, the reaction mixtures were pre-incubated at 32°C for 2 min. Subsequent incubation with 32P-poly(A)-labeled globin mRNA was at 32°C for 10 min. Reactions were stopped by chilling and diluting 4-fold with buffer K (20 mM Tris-HCl, pH 8.0, 2 mM DTT, 100 mM potassium acetate) containing 5 mM MgCl2 [81]. Total reaction mixtures were applied onto 11-ml 10%–30% sucrose gradients prepared with the same buffer. Centrifugation was in SW41 rotor at 40,000 rpm and 4°C for 3.5 h. Fractions (0.36 ml) were collected from the top of the gradients. Radioactivity in each fraction was determined and expressed as percentage of total recovered counts. The area under the peaks (less background) was used to quantify ribosome binding. Sedimentation profiles of the purified 40S and 60S ribosomal subunits in sucrose density gradients served to confirm the positions of the 80S and 48S initiation complexes.
Uncapped Luc mRNA (Promega) was 3′ poly(A) extended and radioactively labeled at the m7G cap using [α-32P]GTP, S-adenosyl methionine, and vaccinia-virus guanylyltransferase [54]. After purification and oxidation with NaIO4, the 32P cap-labeled RNA was used for crosslinking studies in RRL as described previously [17],[54],[87]–[89].
To analyze mRNA stability in vitro, Cap-Luc-A98 mRNA, uniformly labeled with [α-32P]UTP during transcription (1.6×106 cpm, 40×103 cpm/ng), was translated in Krebs extracts (mock-depleted or hnRNP-Q2-depleted) in a total volume of 100 µl under standard conditions. Fifteen µl aliquots of the reaction mixture were withdrawn at 30-min intervals. Total RNA was deproteinized by phenol-chloroform extraction, separated on a formaldehyde-1% agarose gel, and transferred onto a nylon membrane (Hybond-N; GE Healthcare). Blots were stained with Blot Stain Blue (Sigma) to determine the levels of 28S rRNA (loading control). 32P-labeled Cap-Luc-A98 mRNA was detected by autoradiography. Quantifications of band intensities were carried out using NIH ImageJ software.
The Coomassie-stained protein bands were cut from the gel and treated with trypsin. Tryptic peptides were analyzed at Genome Quebec Innovation Centre using a nano-HPLC system coupled to a 4000 Q TRAP mass spectrometer (Applied Biosystems, Foster City, CA). Peptide identities were determined by searching UniProt database (version 13.8) with restriction to human using Mascot (version 2.1, Matrix Science, London).
Mouse fibroblast-like cell line L929 was purchased from ATCC. The cells were transduced with two shRNAs directed against human hnRNP-Q, shRNA1 (TRCN0000112054), shRNA2 (TRCN0000112053), and a nontargeting control shRNA (SHC002) using a lentivirus transduction system (Sigma-Aldrich) as recommended by the manufacturer. Cells were selected with puromycin (2 µg/ml) for 4 d. One week after infection, the cytoplasmic extracts were prepared and analyzed for down-regulation of hnRNP-Q2 by Western blotting.
Control and hnRNP-Q2 knockdown L929 cells, at ∼90% confluence, were washed with methionine-free DMEM and incubated in methionine-free DMEM supplemented with dialyzed fetal bovine serum (10%; GIBCO), glutamine, and [35S]methionine/cysteine labeling mixture (100 µCI/ml) at 37°C for 30 min. Cells were lysed in SDS-sample buffer, and 35S incorporation into trichloroacetic acid-insoluble material was determined [34]. The values for 35S incorporation were normalized to the amounts of total protein in the samples.
To produce the reporter RNA, a 350 nt DNA fragment of pRL-6xB-A98 containing six target sites for human miRNA let-7a [62] was PCR-amplified using primers T7-3′UTR (encoding the T7 promoter) and Oligo 3R(−) [25]. The resulting PCR product was linearized using the restriction site AgeI immediately downstream the poly(A) tail and used as a template for in vitro transcription. [α-32P]UTP-labeled 6xB-3′UTR RNA was synthesized using the T7 MaxiScript in vitro transcription kit (Ambion) and purified by passing through the Mini Quick Spin RNA column (Roche) [25],[62]. To assay deadenylation, portions (8 µl) of nuclease-treated and supplemented Krebs extract, depleted of either hnRNP-Q2 or both hnRNP-Q2 and PABP, were mixed first with the indicated amounts of recombinant hnRNP-Q2, PABP, or Paip2 and then with 0.1 ng purified [α-32P]UTP-labeled 6xB-3′UTR RNA in a total volume of 10 µl. The reactions were incubated at 30°C for the indicated times, after which RNA was extracted with TRIzol reagent (Invitrogen) and analyzed by 4.5% polyacrylamide-urea gel electrophoresis. The dried gels were analyzed with a Typhoon PhosphorImager (GE Healthcare).
Nearly confluent L929 cells, control and hnRNP-Q2 knockdown, were transfected with 100 ng of pRL or pRL-6xB and 50 ng of pFL in six-well plates using Lipofectamine 2000 (Invitrogen) [69]. Where indicated, 2′-O-Me antisense oligonucleotides complementary to either Let-7a or miR-122a miRNA (Dharmacon) were co-transfected at a final concentration of 90 nM [69]. Cells were split after 24 h. RL and FL activities were measured 48 h posttransfection using Dual-Luciferase Assay Kit (Promega) and their ratio was determined. The results from three estimations are represented as means ± SD. To analyze the levels of RL and RL-6xB mRNAs, RNA was extracted from a spare set of transfected cells by TRIzol and subjected to Northern blotting as described previously [69],[90]. The levels of RL and RL-6xB mRNAs were normalized against a control FL mRNA.
Statistical significance of the differences between means was evaluated using a Student's paired t test with a two-tailed distribution. The differences were considered significant at p<0.05.
|
10.1371/journal.pgen.1006579 | Reduced Insulin/Insulin-Like Growth Factor Receptor Signaling Mitigates Defective Dendrite Morphogenesis in Mutants of the ER Stress Sensor IRE-1 | Neurons receive excitatory or sensory inputs through their dendrites, which often branch extensively to form unique neuron-specific structures. How neurons regulate the formation of their particular arbor is only partially understood. In genetic screens using the multidendritic arbor of PVD somatosensory neurons in the nematode Caenorhabditis elegans, we identified a mutation in the ER stress sensor IRE-1/Ire1 (inositol requiring enzyme 1) as crucial for proper PVD dendrite arborization in vivo. We further found that regulation of dendrite growth in cultured rat hippocampal neurons depends on Ire1 function, showing an evolutionarily conserved role for IRE-1/Ire1 in dendrite patterning. PVD neurons of nematodes lacking ire-1 display reduced arbor complexity, whereas mutations in genes encoding other ER stress sensors displayed normal PVD dendrites, specifying IRE-1 as a selective ER stress sensor that is essential for PVD dendrite morphogenesis. Although structure function analyses indicated that IRE-1’s nuclease activity is necessary for its role in dendrite morphogenesis, mutations in xbp-1, the best-known target of non-canonical splicing by IRE-1/Ire1, do not exhibit PVD phenotypes. We further determined that secretion and distal localization to dendrites of the DMA-1/leucine rich transmembrane receptor (DMA-1/LRR-TM) is defective in ire-1 but not xbp-1 mutants, suggesting a block in the secretory pathway. Interestingly, reducing Insulin/IGF1 signaling can bypass the secretory block and restore normal targeting of DMA-1, and consequently normal PVD arborization even in the complete absence of functional IRE-1. This bypass of ire-1 requires the DAF-16/FOXO transcription factor. In sum, our work identifies a conserved role for ire-1 in neuronal branching, which is independent of xbp-1, and suggests that arborization defects associated with neuronal pathologies may be overcome by reducing Insulin/IGF signaling and improving ER homeostasis and function.
| Sensory neurons sample their environment through highly branched structures termed dendritic arbors or trees. The precise patterning of dendritic arbors is important for the proper functioning of the nervous system, and evidence indicates an involvement of sensory neurons in neuropsychiatric disease such as autism spectrum disorders. The unfolded protein response is a cellular process that ensures and maintains a functional protein-folding environment in the cell’s endoplasmic reticulum, and is compromised in a number of neurodegenerative conditions. Recently, this process has also been implicated in dendrite patterning. We discovered that the function of the unfolded protein response in dendrite patterning is evolutionarily conserved from the roundworm C. elegans to mammals. Specifically, dendrites in both worms and mammals require the function of a conserved enzyme with both kinase and ribonuclease activity, which acts as a sensor for the unfolded protein response. Importantly, we find that loss of this enzyme can be bypassed by reducing the signaling through the insulin-like growth factor receptor. Our findings reveal a new way of bypassing defects in the unfolded protein response during dendrite development.
| During their development neurons can form complex dendritic branching patterns. The specific arbor morphologies of different neuron types are thought to have evolved to mediate the acquisition and processing of distinct inputs [1]. Defective arbor morphologies in brain neurons are a common cellular symptom in many neuropsychiatric and neurodegenerative diseases [2–4].
Dendritic arbor growth requires the accurate orchestration of numerous cellular events that occur concomitantly at a distance from the neuronal cell body and integrate dramatic membrane extension, local protein translation and processing, vesicular transport, shifts in cytoskeleton dynamics and elevated metabolic activity. How neurons control these various processes at the genetic and molecular level remains only partially understood [5–7].
Our understanding of dendrite arbor morphogenesis has advanced significantly through the study of peripheral mechanosensory arbor development in the fly Drosophila melanogaster and the nematode Caenorhabditis elegans. In Drosophila, larval da (dendrite arborization) neurons are grouped into four classes according to the degree of arbor complexity [6, 8]. Screens for da dendrite defects have identified many genes that control arborization, such as transcription factors, membrane receptors and their ligands, integrins, vesicular transport factors and cell adhesion molecules [5, 6, 8]. Recently, the polymodal sensory neuron PVD in C. elegans with its characteristic multidendritic arbor has become a model neuron for the study of dendrite morphogenesis [9, 10]. Genetic work on the formation of the repetitive PVD menorah-shaped dendritic units has identified several genes not implicated before in dendrite morphogenesis, including roles for fusogens [11], the LRR-type receptor dma-1 [12], the fam151 family member mnr-1/menorin [13, 14], the secreted leukocyte cell-derived chemotaxin 2 lect-2/chondromodulin II [15, 16], and the furin-like protease kpc-1 [17–19].
The endoplasmic reticulum (ER) is the primary cellular site for secretory protein and lipid biosynthesis, both of which are essential for proper cellular function. In agreement, disruption of ER homeostasis is associated with pathologies such as neurodegenerative disorders [20–22]. To prevent deleterious outcomes of perturbed ER homeostasis, a cellular program called the Unfolded Protein Response (UPR) is triggered at times of increased load on the ER (i.e. ER stress) to ensure that ER homeostasis is retained regardless of the dynamic nature of cellular demand [23]. In mammalian cells (as well as in C. elegans), the UPR is composed of three pathways that are initiated by distinct ER stress sensors: inositol-requiring enzyme 1 (IRE1), protein kinase RNA (PKR)-like ER kinase (PERK) and activating transcription factor-6 (ATF6). IRE1 is the most ancient of the UPR sensors, being conserved from yeast to humans, and bears both kinase and ribonuclease activities [24]. Upon its activation, IRE1 undergoes autophosphorylation and oligomerization into multimers [25, 26]. In its oligomeric state it removes an intron from xbp-1 (X-box binding protein-1) mRNA through unconventional splicing allowing the translation of an activated form of the XBP-1 transcription factor. This activated transcription factor induces the expression of chaperones, ERAD components and other ER auxiliary factors to rebalance ER capacity [27, 28]. The UPR, and specifically the ire-1/xbp-1 arm of the UPR, is important even under normal physiological conditions (i.e. in the absence of induced ER stress), as perturbations in this pathway impair secretory protein metabolism [29].
Additional xbp-1 independent functions of ire-1 are also known. These include activation of the cell death machinery [30–32], induction of autophagosomes [33], induction of a cellular anti-oxidant response [34] and degradation of ER-localized mRNAs that encode secreted and membrane proteins through the RIDD (regulated Ire1-dependent decay) pathway [35]. Recent in vitro work using the yeast Ire1 has suggested that RIDD activity can be mediated by IRE1 even in its monomeric state [36].
Here, we demonstrate that IRE1’s role in dendrite arborization is conserved during evolution from C. elegans to mammals. We show that in C. elegans ire-1 deficiency elicits a secretory block in the PVD neuron that interferes with the targeting of the DMA-1 receptor to the plasma membrane, strengthening similar results by Wei et al. [37]. We further reveal that this trafficking block, which does not occur in xbp-1 mutants, can be bypassed by reducing insulin/IGF1 signaling to restore normal arbor architecture. Altogether, this work assigns a conserved role for IRE-1 function in neuronal development and demonstrates that activation of alternative ER homeostasis-promoting pathways can counteract and prevent the deleterious consequences of compromised ER homeostasis on neuronal development.
The dendrites of the polymodal somatosensory PVD neurons are stereotypically patterned, by the consecutive branching of secondary, tertiary, and quaternary branches from primary dendrites that exit the PVD cell body on either side both in an anterior and in a posterior direction (Fig 1A). In concordance with a recent report [37], we isolated a mutant allele of ire-1, which encodes the C. elegans homolog of the inositol requiring enzyme 1 (IRE1) in a screen for genes required for PVD morphogenesis [13]. The ire-1(dz176) allele changes Glycine 708, a residue that is located in an alpha helix of the kinase domain and conserved from yeast to humans (Fig 1A and 1B). The PVD phenotypes were shared with another missense allele (zc14), which also changed a perfectly conserved G723 in the kinase domain, as well as the deletion allele ok799 (Fig 1A–1C). Mutant phenotypes were transgenically rescued by both a wild-type copy of ire-1 (S1A–S1C Fig) and expression of a cDNA pan-neuronally or in PVD neurons, but not in the intestine or hypodermis (skin) (S1D Fig). These findings complement previous mosaic studies [37], and together strongly argue for a cell-autonomous function of ire-1.
As reported previously, ire-1 mutants formed dendrites with quaternary branches only in the area proximal to the PVD cell body, and gradually became less developed as their distance from the cell body increased both anteriorly and posteriorly (Fig 1A) [37]. We extended these observations in morphometric analyses, which showed a reduction both in the number and aggregate length of secondary, tertiary and quaternary branches in the presumptive ire-1(ok799) null mutant (Fig 1D and 1E; S1E Fig). In addition, we discovered a self-avoidance defect in ire-1 mutants, where adjacent tertiary dendrites failed to retract upon mutual contact, thereby eliminating the characteristic gaps between them (Fig 1F).
Since the role of IRE1 in dendrite patterning in mammals had never been addressed before, we investigated whether IRE1 serves an evolutionarily conserved function during dendrite patterning in mammals. We studied dendrite morphogenesis in dissociated rat hippocampal cultures and measured changes in dendrite length and complexity after 8 and 12 days in vitro (DIV), a time period during which dendrites undergo dynamic growth (Fig 2A). During this time window, neurons in culture were either treated with vehicle or the IRE1-specific inhibitor 4μ8C [38]. Vehicle-treated neurons showed the expected developmental increase in total dendritic branch length from 8 to 12DIV (Fig 2B; 8DIV, 897 +/- 50 μm, n = 42 vs. 12DIV+veh, 1308 +/- 75 μm, n = 42; p<0.0001). In contrast, neurons treated from 8DIV onwards with 50 μM of the specific IRE1 RNAse inhibitor 4μ8C did not show this developmental increase in total dendritic branch length (Fig 2B, 8DIV vs. 12DIV+4μ8C, 893 +/- 66 μm, n = 41; p = 0.99). In neurons treated with the IRE1 inhibitor, there was a trend towards fewer dendrite tips as compared with vehicle treated neurons (Fig 2C; 12DIV+veh, 22 +/- 1.1 tips vs. 12DIV+4μ8C, 18 +/- 1.6 tips; p = 0.078), consistent with the correlation of shorter total dendritic branch length with fewer dendritic tips [39]. IRE1 inhibition did not restrict the developmental increase in average dendritic branch length, supporting the notion that this aspect of dendrite differentiation was not impaired (Fig 2D; 8DIV, 26 +/- 1.1 μm vs. 12DIV+veh, 36 +/- 1.6 μm; p<0.0001; 8DIV vs. 12DIV+4μ8C, 33 +/- 2.3 μm, p<0.001). Importantly, the effect of IRE1 inhibition was specific to higher order branches and did not alter the number of primary branches (Fig 2E), similar to the effects observed in PVD dendrites of ire-1 mutants in C. elegans (Fig 1D and 1E). We conclude that IRE-1 serves a conserved role in dendritic dendrite morphogenesis under normal physiological conditions, and in the absence of external induction of ER stress. Thus, our studies in rats and C. elegans provide the first example for a conserved developmental function of the ire1 stress sensor in neural development. This adds to a growing body of literature, based on knockout approaches in mice, that show functions for the unfolded protein response during liver development [40–42], as well as in the development of antibody-producing B cells [43] and secretory cells of the pancreas [44].
The IRE-1 protein is composed of a luminal unfolded protein sensor domain and a cytosolic bifunctional active site, comprising a kinase and a ribonuclease domain. Our mutants in the kinase domain, as well as mutants identified by Wei et al. [37] suggested that both domains may be important for ire-1 function. To further investigate this notion, we generated mutant versions of IRE-1, defective in each of these domains, and conducted rescue experiments in ire-1 mutants. We found that expression of a mutant where the ER luminal domain, thought to serve as an unfolded protein receptor [45], had been replaced by red fluorescent protein (mCherry), rescued PVD morphology in ire-1 mutant animals, although not as efficiently as the full length transgene (Fig 3D). In contrast, expression of the ribonuclease-deficient mutant version IRE-1K853A, affecting a highly conserved residue in the putative nuclease active site [46] and completely devoid of any detectable xbp-1 splicing activity (Fig 3E), failed to rescue PVD arbor morphology in ire-1 mutant animals (Fig 3D). This implies that IRE-1 nuclease activity is necessary for dendrite morphogenesis. In addition, we expressed IRE-1L589G, an IRE-1 transgene harboring a mutation analogous to the yeast ire1p mutation L745G, which alters the specificity of the ATP binding site in the kinase domain of the protein [47]. In contrast to the yeast studies, the ribonuclease activity of IRE-1L589G appeared intact in an xbp-1 splicing assay (Fig 3E). We found that expression of IRE-1L589G also rescued the arborization defects in PVD sensory dendrites (Fig 3D). Collectively, our rescue studies show that PVD development requires the ribonuclease activity of IRE-1. This conclusion is consistent with the defective PVD arborization phenotype previously observed in ire-1(wy762) mutants, in which a conserved residue in the endoribonuclease domain of the protein has been altered [37]. In addition, kinase activity is likely required, because three mutant alleles of ire-1 (dz176, zc14, this study; wy782, [37]) that result in substitutions of distinct conserved residues in the kinase domain of the protein, display a defective PVD arborization phenotype.
In addition to IRE-1, metazoans have at least two distinct additional sensors of ER stress, the pek-1/PERK kinase and the atf-6/ATF6 transcription factor [48]. Interestingly, PVD development proceeds normally in pek-1/PERK and atf-6/ATF6 mutants, demonstrating that they do not individually serve a critical role in PVD dendrite morphogenesis, and pointing at a unique function of IRE-1 (Fig 3A and 3C)[37].
To gain insight in the downstream effectors of IRE-1 signaling, we focused on the processing of xbp-1 mRNA through unconventional splicing by IRE-1, the best known activity of IRE-1 [27, 28]. Interestingly, two different xbp-1 mutant alleles, zc12 and tm2457 displayed a PVD arbor that was indistinguishable from wild type animals (Fig 3B and 3C), suggesting that IRE-1 can function through xbp-1-independent activities in patterning PVD dendrites.
Known xbp-1-independent functions of ire-1 include activation of the TRAF2 and JNK kinase signaling cascade [30, 49], and degradation of ER-localized mRNAs that encode secreted and membrane proteins through the RIDD (regulated Ire1-dependent decay) pathway [35]. We found that PVD arborization remained normal upon depletion of the C. elegans TRAF2 homolog trf-1 or concomitant depletion of all three C. elegans jnk-1-related kinases (Fig 3F) suggesting that neither pathway plays non-redundant roles in PVD morphogenesis. To directly explore whether RIDD is the mechanism by which IRE-1 controls PVD arborization, we sought another way to maintain RIDD activity in ire-1 mutants while compromising xbp-1 splicing activity. A mutation in the yeast yIre1 protein, R1087D, uncouples the two nuclease activities of ire1p in yeast by impairing xbp-1 splicing while leaving RIDD activity intact [36]. The analogous mutation in worms, IRE-1R882D, failed to rescue PVD architecture (Fig 3D). Since xbp-1 function is not required for PVD morphogenesis, we suggest that C. elegans IRE-1R882D mutant protein may not discriminate between xbp-1-related and unrelated nuclease activities. Thus, among the known xbp-1 independent activities of IRE-1, RIDD remains the most likely to mediate PVD dendrite arborization. This conclusion supports experiments where mosaic knock out of an essential xrn-1 RNA endonuclease, believed to be part of the RIDD pathway, produced low penetrance defects in PVD neurons [37].
Recently, it was shown that even under normal growth conditions (i.e. without artificially-induced ER stress) ire-1 mutants display defects in the metabolism of secretory proteins [29]. One central protein located on the cell membrane of PVD and essential for proper dendrite branching is the DMA-1 leucine rich repeat transmembrane receptor [12]. In concordance with a recent report [37] we found that a DMA-1::GFP reporter primarily localized to the cell body of ire-1 mutant animals (Fig 4B and 4E). In contrast, in wild-type animals, the DMA-1::GFP reporter localized both to the cell body as well as to the entire PVD arbor, throughout the primary, secondary, tertiary and quaternary branches (Fig 4A and 4E). Importantly, although the primary branch of the PVD dendrite is always present and extends along the body of ire-1-deficient animals (Fig 1A), DMA-1::GFP expression was restricted to the cell body and was not detected on the plasma membrane of the primary branch of PVD (Fig 4B). This suggests that DMA-1 is specifically required for patterning of the secondary, tertiary and quaternary branches in ire-1-deficient animals. This further suggests that the DMA-1::GFP localization defect in ire-1 mutants precedes the PVD patterning defect. Collectively, these observations suggest that DMA-1 fails to shuttle properly through the secretory pathway, resulting in patterning defects of higher order branches of the PVD dendrite.
Intriguingly, xbp-1 mutants displayed completely normal DMA-1::GFP staining of the entire PVD dendrite similar to wild-type animals (Fig 4C), and did not accumulate DMA-1::GFP in the soma like ire-1 mutants (Fig 4E). The proper localization of DMA-1::GFP in xbp-1 mutants contrasts with its mislocalization in ire-1 mutants (Fig 4A–4C and 4E), but is consistent with the PVD arborization architecture seen in the respective mutants (Fig 3A–3C). This finding was surprising, given that loss of xbp-1 has been shown to perturb ER homeostasis and interfere with secretory protein metabolism [29]. Based on the trafficking defects in ire-1 mutant animals, and the finding that the basal activity of the UPR in PVD itself is largely dependent on DMA-1 expression, it has been suggested that the failure of DMA-1::GFP to reach the plasma membrane is a consequence of a folding challenge of DMA-1 itself [37]. However, we point out that although in many cases functional xbp-1 is also required for the trafficking and maturation of other secreted and transmembrane proteins [29, 50], it was not required for DMA-1 trafficking to the plasma membrane, and PVD morphogenesis is normal in xbp-1 mutants. Thus, DMA-1 can traffic to the plasma membrane and support PVD dendrite morphogenesis even under the unfavorable proteostatic conditions in the ER of xbp-1-deficient animals. This suggests that DMA-1 may not have an intrinsic tendency to fold improperly and that the DMA-1 trafficking defect is more likely a reflection of a general overload and perturbed function of the ER in the PVD neuron that lacks ire-1. A possible explanation for the differences between ire-1 mutants and xbp-1 mutants is that ER homeostasis and function is less compromised in xbp-1-deficient animals compared to ire-1-deficient animals [29]. Thus, the DMA-1-dependent activation of the UPR in PVD suggested by Wei et al. [37] may be an indirect consequence of DMA-1 promoting dendrite morphogenesis and expansion, both of which require the synthesis of membranes proteins and lipids and impose a significant biosynthetic load on the ER. This ‘capacity model’ is also consistent with the observation that overexpression of spliced xbp-1 or its target, the ER-localized heat shock protein HSP-4/BiP/grp78 can bypass the requirement for ire-1 and rescue the morphological defects and the DMA-1::GFP secretion defects in ire-1 mutants [37]. Altogether, our report adds to a growing number of recent works delineating an xbp-1-independent branch of the ire-1 pathway [29, 35, 40, 51, 52].
If the failure to form menorahs in ire-1 mutants is a result of a block in the secretory pathway in the PVD neuron then conditions that release the secretory block in ire-1 mutants should restore PVD arborization. One way to overcome the secretory block in ire-1 mutants is by activating the FOXO transcription factor DAF-16, which is inhibited by the insulin/IGF-1 signaling (IIS) pathway [52]. Indeed, reducing IIS in animals through a mutation in their daf-2 gene, the only insulin-like growth factor receptor in C. elegans, resulted in reduced accumulation of DMA-1::GFP in the cell body and redistribution to the plasma membrane of PVD in ire-1 mutants (Fig 4D and 4E). Consistent with the restored localization of DMA-1::GFP expression pattern in the PVD neuron, we found that PVD dendrite morphogenesis defects in ire-1; daf-2 double mutants were completely reversed and PVD arbors of double mutants were indistinguishable from wild type animals (Fig 4F). This finding was further corroborated by morphometric analyses. We found that the reduced length of secondary, tertiary and quaternary branches in ire-1 mutants was suppressed upon reduced DAF-2/IIS signaling (Fig 4G). Moreover, this suppression was largely (although not completely) dma-1-dependent, because dma-1 appeared epistatic in a dma-1; ire-1; daf-2 triple mutant (Fig 4G).
The physiological consequence of reduced DAF-2/IIS signaling, including improving ER homeostasis in ire-1-deficient animals [52], in many cases depends on the activation of the transcription factor DAF-16/FOXO. We found that daf-16; ire-1; daf-2 triple mutant animals showed the same frequency of PVD defects as ire-1 single mutants, indicating that the suppression of defects in ire-1 mutants by loss of daf-2 insulin signaling was entirely dependent on daf-16 activation (Fig 4F). In other words, the defects in dendrite morphogenesis of ire-1 mutants can be rescued by compromising DAF-2/IIS signaling in a daf-16/FOXO-dependent manner.
Our finding that trafficking of a DMA-1::GFP reporter is restored in daf-2/IIS mutants suggests that (1) different approaches can be used to relieve the secretory block in ire-1 mutants, and (2) are consistent with previous observations that attenuation of IIS can result in favorable effects on proteostasis, ER homeostasis, organismal health and survival in C. elegans, as well as other organisms [53, 54]. Similarly, activation of the IIS regulated transcription factor DAF-16/FOXO3A in ire-1-deficient cells can bypass the requirement of the canonical ire-1/xbp-1 pathway for the maintenance of ER homeostasis, and improve both ER homeostasis and restoration of normal secretory protein trafficking in worms and mammalian cells [52]. Thus, our findings may provide a mechanistic explanation for observations in several studies showing that neurons grow and function better under reduced IIS conditions [55–57], and expands this notion to include dendritic arbor morphogenesis. Since the improvement on DMA-1::GFP trafficking and dendrite morphology were dependent on activation of the DAF-16/FOXO transcription factor, the activation of this pathway by alternative cues including starvation as well as a variety of cytotoxic stresses (e.g. heat-shock and oxidative stresses), which directly or indirectly activate DAF-16, hold the potential to recover PVD dendrite morphogenesis in the absence of a properly functioning UPR.
In summary, our results establish that the function of the IRE-1 UPR sensor in neuronal patterning is conserved from invertebrates to mammals. Our findings demonstrate that promoting ER homeostasis, e.g. by reducing IIS, can overcome morphological defects in neuronal patterning. This underscores the importance of discovering and investigating new approaches that can bypass excessive ER stress. Given the conservation of the role of the UPR in dendrite branching and morphogenesis from C. elegans to mammals, as well as the conservation of the proteostasis-promoting effects of the IIS pathway, these findings may offer novel approaches for treatment of neurodegenerative disorders.
Worms were grown on OP50 Escherichia coli-seeded nematode growth medium plates at 20°C. Strains used in this work include: N2 (wild type reference), ire-1(dz176), ire-1(ok799), ire-1(zc14), xbp-1(tm2457), xbp-1(zc12), pek-1(ok275), atf-6(ok551), daf-2(e1370), daf-16(mu86), trf-1(nr2014), kgb-1(um3) kgb-2(gk361) jnk-1(gk7). PVD neurons were visualized by the integrated transgene wdIs52 (Is[F49H12.4::GFP]). Transgenic strains for cell-specific rescue were established by injecting the respective plasmids at 5–10 ng/μl together with rol-6(su1006) or Pttx-3::mCherry (labeling the interneuron AIY) at 50 ng/μl as an injection marker into ire-1(ok799); wdIs52. The PVD::DMA-1::GFP translational fusion was a kind gift of K. Shen (Stanford U, California). For a complete strain list see Supporting Information.
The ire-1 cDNA was amplified with gene specific primers from a N2 mixed stage cDNA sample and cloned KpnI/SphI downstream of the Pttx-3promB regulatory element [58]. For the cell specific heterologous rescue the ire-1 cDNA was placed under control of the Pdpy-7 (hypodermis-specific), Pmyo-3 (muscle), Pges-1 (intestine), Prgef-1 (pan-neuronal) or Pser-2prom3 promoter (PVD/OLL specific). For further details see Supporting Information.
On day 1 of adulthood, animals were collected for RNA extraction, purification and reverse transcription, using random 9-mers and standard protocol. A primers set encompassing the noncanonical intron of the xbp-1 transcript was used, giving rise to two PCR products of amplified spliced and unspliced xbp-1 transcript (primers: 5’- TCCGCTTGGGCTCTTGAGATGTTC-3’ and 5’-TGTCGTCGTCGGAGGAGAGGATCG- 3’). PCR products were visualized on a 2% agarose gel stained with ethidium bromide.
Images of immobilized animals (1–5 mM levamisol, Sigma) were captured using either a Zeiss Axioimager Z1 Apotome at 40X, where Z stacks were collected and maximum projections were used for imaging of dendrites, or with a CCD digital camera using a Nikon 90i fluorescence microscope at 20X magnification. For DMA-1::GFP signal quantification the NIS element software was used to quantify sum and mean fluorescence intensity as measured by intensity of each pixel in the selected area.
Hippocampal neurons were prepared from rats at E18 as previously described [59] with modifications. In brief, dissected hippocampi were incubated in 0.05% trypsin at 37°C for 20 minutes (Invitrogen 25300054) and plated at a density of 60,000 cells per 12 mm coverslip coated with poly-l-lysine (Sigma P1274). Cells were incubated in a cell culture incubator maintained at 37°C with 5.0% CO2. Cytosine arabinoside (ara-c, Sigma C1768) was added at a final concentration of 2 μM at 2 days in vitro (DIV) to prevent glia cell overgrowth before being replaced with Neurobasal without ara-c at 4DIV. Neurons were transfected at 5-6DIV using Lipofectamine LTX and Plus Reagent (ThermoFisher). For cytoplasmic labeling of neurons to visualize dendrites, 60,000 cells were transfected with 0.25 μg pCAGGS-mCherry [60]. Neurons were treated at 8DIV with IRE-1 RNAse inhibitor 4μ8C (EMD Millipore 412512). 4μ8C was first dissolved in DMSO (Invitrogen) and diluted in supplemented Neurobasal. Diluted 4μ8C at 100 μM was added to cultures at 1:1 with conditioned neural media with final concentrations of 50 μM 4μ8C and 0.5% DMSO. Additional 4μ8C was added to neurons at 10DIV resulting in a final concentrations of 37.5 μM 4μ8C and 0.5% DMSO. Vehicle neurons were treated in identical ways using media containing DMSO only.
At 8 or 12 DIV coverslips with neurons were quickly washed two times with PBS, followed by fixation for 15 min with 4% PFA / 4% sucrose in PBS at RT. Cells were blocked and permeabilized with 3% horse serum and 0.05% Triton X-100 in PBS for 1 h at RT. Cells were incubated with antibodies against mCherry (Rockland 600-401-379) at 4°C overnight. The next day cells were washed three times with PBS and labeled with AlexaFluor-conjugated secondary antibodies (Invitrogen; 1:500) for 1 hr at RT. Cells were washed three times with PBS, stained with DAPI, and mounted on slides Aqua-Mount mounting media (Thermo Scientific). Tiled images of dendritic arbors were acquired using a Keyence BZ-X710 Fluorescence Microscope equipped with a Nikon 60X oil-immersion 1.40 NA objective. Merged composite images of the individually acquired tiled images were generated using Keyence software. Dendritic arbors were traced using the NeuronJ plugin for ImageJ [61]. Total dendritic branch length was calculated as the sum of the length of all dendrites. Average dendrite branch length is average length of each dendritic branch excluding primary dendrites, as primary dendrite lengths are highly variable across neurons. All images were acquired and all analysis was performed with the experimenter blind to conditions. Data analysis was performed using GraphPad Prism 6.
Error bars represent the standard error of the mean (SEM) of at least 3 independent experiments. P values were calculated using the unpaired Student's t test, or one-way ANOVA with the Tukey correction for multiple comparisons (GraphPad Prism 6).
|
10.1371/journal.pcbi.1003256 | Reassessing Google Flu Trends Data for Detection of Seasonal and Pandemic Influenza: A Comparative Epidemiological Study at Three Geographic Scales | The goal of influenza-like illness (ILI) surveillance is to determine the timing, location and magnitude of outbreaks by monitoring the frequency and progression of clinical case incidence. Advances in computational and information technology have allowed for automated collection of higher volumes of electronic data and more timely analyses than previously possible. Novel surveillance systems, including those based on internet search query data like Google Flu Trends (GFT), are being used as surrogates for clinically-based reporting of influenza-like-illness (ILI). We investigated the reliability of GFT during the last decade (2003 to 2013), and compared weekly public health surveillance with search query data to characterize the timing and intensity of seasonal and pandemic influenza at the national (United States), regional (Mid-Atlantic) and local (New York City) levels. We identified substantial flaws in the original and updated GFT models at all three geographic scales, including completely missing the first wave of the 2009 influenza A/H1N1 pandemic, and greatly overestimating the intensity of the A/H3N2 epidemic during the 2012/2013 season. These results were obtained for both the original (2008) and the updated (2009) GFT algorithms. The performance of both models was problematic, perhaps because of changes in internet search behavior and differences in the seasonality, geographical heterogeneity and age-distribution of the epidemics between the periods of GFT model-fitting and prospective use. We conclude that GFT data may not provide reliable surveillance for seasonal or pandemic influenza and should be interpreted with caution until the algorithm can be improved and evaluated. Current internet search query data are no substitute for timely local clinical and laboratory surveillance, or national surveillance based on local data collection. New generation surveillance systems such as GFT should incorporate the use of near-real time electronic health data and computational methods for continued model-fitting and ongoing evaluation and improvement.
| In November 2008, Google Flu Trends was launched as an open tool for influenza surveillance in the United States. Engineered as a system for early detection and daily monitoring of the intensity of seasonal influenza epidemics, Google Flu Trends uses internet search data and a proprietary algorithm to provide a surrogate measure of influenza-like illness in the population. During its first season of operation, the novel A/H1N1-pdm influenza virus emerged, heterogeneously causing sporadic outbreaks in the spring and summer of 2009 across many parts of the United States. During the autumn 2009 pandemic wave, Google updated their model with a new algorithm and case definition; the updated model has run prospectively since. Our study asks whether Google Flu Trends provides accurate detection and monitoring of influenza at the national, regional and local geographic scales. Reliable local surveillance is important to reduce uncertainty and improve situational awareness during seasonal epidemics and pandemics. We found substantial flaws with the original and updated Google Flu Trends models, including missing the emergence of the 2009 pandemic and overestimating the 2012/2013 influenza season epidemic. Our work supports the development of local near-real time computerized syndromic surveillance systems, and collaborative regional, national and international networks.
| Influenza remains a paradox for public health: While influenza epidemics are expected seasonally in temperate climates, their exact timing and severity remain largely unpredictable, making them a challenge to ongoing preparedness, surveillance and response efforts [1]. Surveillance efforts for influenza seek to determine the timing and impact of disease through characterizing information on reported illnesses, hospitalizations, deaths, and circulating influenza viruses [2]. Since establishment of the first computerized disease surveillance network nearly three decades ago [3]–[5], the use of information and communications technology for public health disease monitoring has progressed and expanded. During the last decade, the use of electronic syndromic surveillance systems have allowed for automated, detailed, high volume data collection and analysis in near-real time [6]–[9]. In parallel, novel approaches based on influenza-related internet search queries have been reported to yield faster detection and estimation of the intensity of influenza epidemics [10]–[16]. The public health utility of such systems for prospective monitoring and forecasting of influenza activity, however, remains unclear [17]–[21], particularly as occurred during the 2009 pandemic and the 2012/2013 epidemic season [22]–[24].
In November 2008, Google began prospectively monitoring search engine records using a proprietary computational search term query model called Google Flu Trends (GFT) to estimate national, regional and state level ILI activity in the United States (US) [12]. The goal of GFT was to achieve early detection and accurate estimation of epidemic influenza intensity [13]. The original GFT model was built by fitting linear regression models to weekly counts for each of the 50 million most common search queries, from the billions of individual searches submitted in the US between 2003 and 2007 [13]. An automated query selection process identified the exact text searches that yielded the highest correlations with national and regional influenza-like-illnesses (ILI) surveillance in the US during the period of model fitting; the top scoring 45 search terms constituted the original GFT ILI search definition.
The GFT search algorithm was revised in the autumn of 2009, following the emergence and rapid spread of the pandemic A/H1N1pdm09 influenza virus in the US, which had gone wholly undetected by the GFT system. The updated GFT model used surveillance data from the first 20 weeks of the pandemic and a qualitative decision process with less restrictive criteria for additional ILI-related search terms to be included [14]. By September 2009 the historical GFT model was replaced with retrospective estimates from the revised algorithm. Currently, the updated GFT model provides real-time estimates of influenza intensity at three geographic scales in the US: national, state and select local cities, as well as estimates for many countries worldwide [16].
The original and updated GFT models have both shown high retrospective correlation with national and regional ILI disease surveillance data [13], [14]; however, the prospective accuracy of this surveillance tool remains unclear, even though GFT estimates are used in forecasting models for influenza incidence [15], [20], [21]. We present a comparative analysis of traditional public health ILI surveillance data and GFT estimates for ten influenza seasons to assess the retrospective and prospective performances of GFT to capture season-to-season epidemic timing and magnitude.
We compared weekly ILI and GFT data from June 1, 2003 through March 30, 2013, a period of ten influenza seasons which included a range of mild and moderately severe seasonal influenza epidemics as well as the emergence of the first influenza pandemic in over forty years. The surveillance systems were assessed at three geographical levels: entire US, Mid-Atlantic region (New Jersey, New York and Pennsylvania) and New York City.
All public health surveillance data used in the study came from systems operating prospectively on a daily or weekly basis throughout the study period [2], [25]–[27]. Nationwide and regional ILI surveillance data were compiled from the US Centers for Disease Control and Prevention (CDC) sentinel ILI-Net surveillance system, which includes sources ranging from small physician practices to large electronic syndromic surveillance networks [2]. The CDC ILI-Net system is publically available each week, typically on Friday for the previous week ending Saturday during the respiratory season (October to May), with a recognized reporting lag of 1–2 weeks [2], [13]. Local ILI data came from the New York City Department of Health and Mental Hygiene (DOHMH) emergency department (ED) syndromic surveillance system, which is collected and analyzed daily, with a reporting lag of about one day [25]–[27]. In each system, all weekly public health surveillance ILI proportions were calculated as total ILI visits divided by all visits each week.
Internet search query data came from the original [13] and updated GFT models [14], using weekly estimates available online [16] from both the periods of retrospective model-fitting (4 seasons for the original model and 6 seasons for the updated model) and prospective operation for both models (1 season and 4 seasons, respectively; Table 1). Finalized weekly GFT estimates were publically available each Sunday for the previous week, with a reporting lag of about one day. The original and updated GFT models used scaled measures of ILI-related searches to be directly comparable to the weighted ILI proportions from the CDC ILI-Net system [2], [13], [14], [16] (Figure 1). For additional details on data sources, see Supporting Information.
All observed ILI weekly proportions were analyzed with a traditional Serfling regression approach to establish weekly expected baselines and estimate the “excess” ILI proportions attributable to influenza and identify epidemic periods ([28]–[33]; Supporting Information). The GFT system presents ILI search query estimates as a qualitative measure of influenza activity on a scale ranging from “minimal” to “intense” each week [16]; neither GFT model provided quantitative measure for detection or estimation of impact [13], [14]. For all public health surveillance and GFT estimates we assessed two epidemiological criteria to characterize influenza outbreaks: epidemic timing and intensity.
Timing was based on estimates of epidemic onset and peak week for each season and ILI surveillance system. The onset each season was defined as the first of consecutive weeks exceeding the surveillance threshold (upper limit of the 95% confidence interval of the Serfling baseline). The peak week was identified as the week with the greatest proportion of ILI visits each season or epidemic (Table 2).
For each data source and season we assessed epidemic intensity by determining the proportion of excess ILI for peak weeks and by summing the weekly excess ILI proportions for each epidemic period as a measure of cumulative ILI intensity for each season and epidemic. All Serfling regression confidence intervals represented the upper and lower 95% limit, calculated as the predicted non-epidemic baseline ±1.96 standard deviations [28]–[33]. We calculated the ratio of excess GFT divided by excess ILI at each geographic level for each epidemic (Table 3), with a constant ratio indicating consistent influenza monitoring by GFT for the period.
To further evaluate the week-to-week accuracy and timing of GFT and potential asynchrony with traditional ILI surveillance, we calculated Pearson correlations in the national, regional and local datasets, following the original methods used in the development [13] and evaluation of GFT [14]. Original and updated GFT model estimates were assessed for the periods of retrospective model-fitting and prospective monitoring (Table 2), and for specific epidemic seasons (Table 4). We measured cross-correlations at negative and positive lags for each influenza season to identify the corresponding lead or lag with the highest correlation values between GFT and traditional ILI systems, indicating the degree of shift in the timing of the GFT trends compared to ILI surveillance.
While correlations are useful to assess GFT [14], they only provide a measure of relative correspondence between ILI and internet search systems, and do not provide an indication of the nature of the relationship between the trend estimates or the observed lags. As a complementary measure, we compared the regression slope of public health ILI data with GFT estimates during retrospective model-fitting and prospective periods, and for specific seasons. For further details, see Supporting Information.
During the study period, June 2003 to March 2013, over 4.5 million ILI visits out of 230 million total outpatient sentinel physician visits were reported nationwide to the CDC ILI-Net surveillance network, of which 16.5% were from the Mid-Atlantic surveillance region. In New York City, over 780,000 ILI and 38 million total ED visits were recorded in the DOHMH syndromic surveillance system, with coverage increasing from 88% of all ED visits that occurred citywide during 2003/2004 to >95% of all visits since 2008. The weekly proportion of ILI visits and GFT estimates showed similar seasonal and epidemic patterns across the three regional scales, though with notable differences between retrospective and prospective periods (Figure 1; Table 1). Specifically, during prospective use the original GFT algorithm severely underestimated the early 2009 pandemic wave (shaded 2009 period, Figure 1), and the updated GFT model greatly exaggerated the intensity of the 2012/2013 influenza season (shaded 2012/2013 period, Figure 1).
Historical estimates from the original GFT model were based on the model-fitting period from September 28, 2003 to March 17, 2007; the system was evaluated during March 18, 2007 to May 11, 2008, and has run prospectively since then. The week-to-week GFT estimates during the model-fitting period were highly correlated with ILI surveillance data at the national (R2 = 0.91), regional (Mid-Atlantic, R2 = 0.79) and state/local level (New York, R2 = 0.89; Table 1). Similarly, GFT estimates were highly correlated with CDC ILI surveillance at the national and regional levels during the validation period [13], and remained high through the period of prospective use prior to the emergence of the 2009 A/H1N1 pandemic, from May 12, 2008 to March 28, 2009 (R2≥0.75; Table 4). Seasonal and epidemic onset and peak weeks varied considerably during the period (Table 2). Estimation of excess ILI visits and GFT search query fractions were also well correlated on a week to week basis during this period (Supporting Tables; Figure 2).
In late-April 2009, detection of novel A/H1N1 influenza in an outbreak in Queens, New York, was immediately followed by a spike in ILI surveillance data across much of the nation during the week ending May 2, 2009 [2]. Mid-Atlantic States and New York City experienced a substantial spring pandemic wave (Figure 1B,C), unlike many other regions of the US [2]. Despite recognized pandemic activity, the national GFT estimates were below baseline ILI levels for May–August 2009, indicating no excess impact (red line, shaded 2009 period, Figure 1A). The correlations between the surveillance ILI and GFT estimates, however, were very high during this period at the US level for observed (R2 = 0.91) as well as estimated excess values (R2 = 0.81; Figure 2A). At the Mid-Atlantic level, correlations were lower for observed (R2 = 0.51), but still high for estimated excess values (R2 = 0.80), while the slope of the linear relationship between the two surveillance systems was near zero (slope = 0.11), indicating that there was little or no excess ILI estimated by GFT (Figure 2B). The discrepancy at the Mid-Atlantic level was exacerbated for New York City, where the pandemic impact was greater than any other epidemic that decade, while the original GFT estimates remained near expected baseline levels for the entire period (R2 = 0.78). Accordingly, the slope of the GFT regression against ILI was near zero (slope = 0.05), indicating that GFT data did not accurately measure the intensity of the pandemic (Figure 2C). Taken together, the original GFT model missed the spring 2009 pandemic wave at all levels (Figure 1), providing incidence estimates 30–40 fold lower than those based on ILI surveillance (Table 3).
The original and updated GFT estimates appeared very similar during the pre-pandemic period 2003–2009, but diverged considerably by May 2009 (red and blue lines, Figure 1). Like the original GFT model, the updated GFT estimates during the model-fitting period were highly correlated with CDC ILI surveillance at the national and regional levels (R2≥0.77, Table 1). In contrast for New York City, the updated GFT estimates were less well correlated with local ILI syndromic surveillance data during this period (R2 = 0.51, Table 1). Of particular interest is the retrospective characterization of the 2009 pandemic by the updated GFT algorithm, which tracked the spring wave very well at the national level, but underestimated the magnitude at the regional level by nearly 30%, and at the New York City level by 70% (Figure 1; Table 3).
In September 2009, the updated GFT algorithm began running prospectively, providing estimates that tracked CDC ILI surveillance data well for the remainder of 2009, a period in which most pandemic A/H1N1 infections occurred. Updated GFT estimates were highly correlated with ILI surveillance at the national (R2 = 0.98), and regional (R2 = 0.92) levels (Figure 1A–B; Table 4). Mid-Atlantic ILI surveillance, however, demonstrated two peaks, consistent with different timing of pandemic waves in states within the region (Figure 1B). For New York City, the updated GFT estimates and ILI surveillance were less well correlated when measured directly (R2 = 0.51), though highly correlated when lagged by three weeks (R2 = 0.89), showing the updated GFT model estimates for the fall 2009 pandemic wave to increase and peak 3 weeks earlier than ILI surveillance (Figure 1C; Table 4). Overall, GFT underestimated the cumulative ILI incidence of the main pandemic period, May–December 2009, by 52% for New York City (25% for the broader region), with non-overlapping confidence intervals between the GFT and ILI surveillance systems (Table 3).
Correlations between the updated GFT model and ILI data during the first two years of prospective post-pandemic surveillance were high at the national level during the 2010/2011 (R2 = 0.95) and 2011/2012 (R2 = 0.88) seasons (Table 4). At the regional level, there was high correlation in 2010/2011 (R2 = 0.83) with a slight underestimation of incidence, and low correlation in 2011/2012 (R2 = 0.37) with a slight overestimation of ILI incidence (Figure 1B). At the New York City level, updated GFT estimates for 2010/2011 were reasonably well correlated with observed ILI (R2 = 0.74), though with ILI surveillance increasing and peaking earlier (Figure 1C), and showing an improved lagged correlation (R2 = 0.80, lagged 1 week; Table 4).
For the relatively early and moderately severe 2012/2013 epidemic season, observed GFT estimates greatly overestimated the initial onset week and magnitude of the outbreak at all three geographical levels (Figure 1; Table 2). The correlations between the updated GFT model estimates and ILI surveillance, however, were very high at all levels (R2≥0.86, Table 4). GFT model estimates of epidemic intensity were far greater than ILI surveillance data at the national (268%), regional (208%) and local (296%) levels (Table 3). Accordingly, the slopes of the weekly regression of ILI surveillance against GFT estimates during 2012/2013 (United States, slope = 1.91; Mid-Atlantic, slope = 2.29; New York City, slope = 2.63) were far greater than those for other epidemic and pandemic seasons (Figure 3), and substantially different from a slope of 1 (p<0.05).
Following Google's development of GFT in 2008, and the considerable excitement generated by the original publication and release of the online tool [12], [13], [16], concerns were raised regarding the tenuous relationship between internet searches and the presentation of illness to clinical or emergency medical providers [17]. We used clinical ILI surveillance data at local, regional and national scales as a proposed “ground truth” to test the ability of GFT to perform as a timely and accurate surveillance system in the US. We identified substantial errors in GFT estimates of influenza timing and intensity in the face of pandemic and seasonal outbreaks, including prospectively missing the early wave of the 2009 pandemic and overestimating the impact of the 2012/2013 epidemic. Although we are not the first to report issues in GFT estimates for seasonal and pandemic influenza [22], our study is the first to carefully quantify the performance of this innovative system over a full decade of influenza activity and across three geographical scales.
The 2009 A/H1N1 pandemic is a particularly important case study to test the performance of GFT, with its unusual signature pandemic features of out-of-season activity in the spring of 2009, atypical (young) age pattern of cases, recurring waves and substantial geographic heterogeneity [34]–[38]. Immediately following the spread of the pandemic virus in the US, public health officials and electronic surveillance networks found that local and state level surveillance data did not correspond with estimates provided by the original GFT model, particularly in some urban areas and harder hit regions of the Northeastern and Midwestern US [18], [39]. Clearly, the original GFT algorithm was not able to track sentinel ILI patterns that deviated from the influenza seasons that occurred during the model-fitting period. Even after the GFT algorithm was revised in September 2009, we have shown that the retrospective estimates for the spring 2009 pandemic wave were still not in agreement with regional and local surveillance. Further, the updated GFT model that has been used prospectively failed to accurately capture the autumn 2009 pandemic wave in New York City, presenting the outbreak three weeks before it actually occurred. This assessment echoes earlier concerns regarding the timeliness and accuracy of internet search data for public health monitoring at the local level [17] and during the early wave of the 2009 pandemic [18]. To have missed the early wave of the 2009 pandemic is a serious shortcoming of a surveillance system – as these are times when accurate data are most critically needed for pandemic preparedness and response purposes.
Although the GFT system provided relatively accurate estimates during post-pandemic years which were characterized by mild influenza activity, it overestimated the 2012/2013 epidemic by 2–3 fold relative to traditional ILI surveillance systems, across national, regional and local geographical levels in the US (see also [22]). While the intensity of the 2012/2013 influenza season was roughly comparable to the 2003 A/H3N2-Fujian epidemic as measured by traditional surveillance and assessed by CDC as “moderately severe” [2], the 2012/2013 season was scored by the GFT tool as by far the most severe epidemic in over a decade.
A limitation of our study is its focus on US systems. Many international syndromic, physician consultation, laboratory and internet survey surveillance systems provide rapid, detailed and accurate influenza-related surveillance [3]–[5], [40]–[48]. These systems allowed for development of GFT search query algorithms which were trained to mimic the specific regional influenza-related patterns [16]. While international GFT search query estimates are publically available earlier than many government run surveillance systems, it is important to note that public health data typically undergo monitoring for data quality and investigation prior to public release. It is also important to note that GFT has been set up where robust surveillance systems already exist, providing ILI search query data for populations that are already under surveillance.
An additional limitation of our study is the imperfect nature of our assumed “ground truth” surveillance. Our study sought to assess the ability of GFT to estimate physician consultation and syndromic ILI surveillance patterns, not necessarily the true incidence of influenza infection and illness. We recognize that physician sentinel and syndromic data can be biased, particularly during periods of heightened public health concern. This has been well described in a study of online survey data and health-seeking behavior during the two waves of the 2009 pandemic in England [48]. This recognized bias highlights the need for multiple sources of surveillance information in the community.
In a previous evaluation of GFT, the authors and engineers at Google and the US CDC concluded that their original GFT model had “performed well prior to and during” the 2009 pandemic, when assessed as simple correlations at national and regional levels [14]. Regarding this measure of performance, however, we found the use of simple correlation to be inadequate, as values greater than 0.90 often occurred during periods when critical metrics such as peak magnitude and cumulative ILI revealed that the GFT models were actually greatly under- or over-estimating influenza activity. Our study demonstrates that simple correlation measures can mischaracterize the performances of a novel surveillance system, and instead we recommend the use of additional and alternative metrics based on estimates of onset and peak timing and cumulative intensity of influenza epidemics.
Because the search algorithm and resulting query terms that were used to define the original and updated GFT models remain undisclosed, [13], [14], it is difficult to identify the reasons for the suboptimal performance of the system and make recommendations for improvement. Concerns were raised early-on that the data-mining nature of GFT might over-fit the historical data and introduce bias in prospective use [17]. After the original GFT model missed the spring 2009 pandemic wave – an outbreak with different timing and characteristics than the outbreaks present in the retrospective model-fitting period – the GFT algorithm was modified, potentially addressing the possible over-fitting issue. The revised GFT model, however, appeared to be susceptible to bias in the opposite direction, possibly due to changes in health information searching and care seeking behavior driven by the media. Further, important epidemiologic information such as patient age, location, illness complaint or clinical presentation remain un-available in GFT (an adult person could be performing a search on behalf of a sick minor in another state). In contrast, public health information systems are less prone to such biases, as they collect demographic and geographic data as well as additional health outcomes, which can be used to investigate atypical signals.
Ultimately, public health actions are taken locally. As such, the accuracy and timeliness of local disease surveillance systems are critical; as is the utility of the information in supporting decisions. The additional detail in local syndromic ILI surveillance data, and its direct link to individuals seeking care, facilitates public health action. Computerized surveillance, such as the New York City syndromic chief complaint ED system, can accurately capture the impact of influenza activity [25], [26]. In the present study, we have shown that these systems are more accurate than, yet equally timely as the GFT tool, which indicates the need for further research and support for computerized local disease surveillance systems.
We believe there is a place for internet search query monitoring in disease surveillance, and for continued research and development in this area [13]–[21], [49]–[58]. For now, in the US CDC's national and regional ILI surveillance data remain the “ground truth” source of influenza activity at national and regional levels, but timeliness, detail and coverage remain issues. Thus, we believe there is a broader need for electronic clinically-based disease surveillance at the local level, similar to the ED system in place in New York City [25]–[27], and for collaborative and distributed networks connecting these systems for research and practice [39], [58]–[60]. Careful evaluation of the strengths and limitations of GFT and other innovative surveillance tools should be expanded to encompass a range of developed and developing country settings, following the approach proposed here, in order to improve local, regional and global outbreak surveillance methods and inform public health responses. The way forward using high volume search query data such as GFT may be through integration of near-real time electronic public health surveillance data, improved computational methods and disease modeling – creating systems that are more transparent and collaborative, as well as more rigorous and accurate, so as to ultimately make them of greater utility for public health decision making.
|
10.1371/journal.pgen.1005135 | Turning Saccharomyces cerevisiae into a Frataxin-Independent Organism | Frataxin (Yfh1 in yeast) is a conserved protein and deficiency leads to the neurodegenerative disease Friedreich’s ataxia. Frataxin is a critical protein for Fe-S cluster assembly in mitochondria, interacting with other components of the Fe-S cluster machinery, including cysteine desulfurase Nfs1, Isd11 and the Isu1 scaffold protein. Yeast Isu1 with the methionine to isoleucine substitution (M141I), in which the E. coli amino acid is inserted at this position, corrected most of the phenotypes that result from lack of Yfh1 in yeast. This suppressor Isu1 behaved as a genetic dominant. Furthermore frataxin-bypass activity required a completely functional Nfs1 and correlated with the presence of efficient scaffold function. A screen of random Isu1 mutations for frataxin-bypass activity identified only M141 substitutions, including Ile, Cys, Leu, or Val. In each case, mitochondrial Nfs1 persulfide formation was enhanced, and mitochondrial Fe-S cluster assembly was improved in the absence of frataxin. Direct targeting of the entire E. coli IscU to ∆yfh1 mitochondria also ameliorated the mutant phenotypes. In contrast, expression of IscU with the reverse substitution i.e. IscU with Ile to Met change led to worsening of the ∆yfh1 phenotypes, including severely compromised growth, increased sensitivity to oxygen, deficiency in Fe-S clusters and heme, and impaired iron homeostasis. A bioinformatic survey of eukaryotic Isu1/prokaryotic IscU database entries sorted on the amino acid utilized at the M141 position identified unique groupings, with virtually all of the eukaryotic scaffolds using Met, and the preponderance of prokaryotic scaffolds using other amino acids. The frataxin-bypassing amino acids Cys, Ile, Leu, or Val, were found predominantly in prokaryotes. This amino acid position 141 is unique in Isu1, and the frataxin-bypass effect likely mimics a conserved and ancient feature of the prokaryotic Fe-S cluster assembly machinery.
| Frataxin was discovered because mutations in the corresponding gene cause the neurodegenerative disease Friedreich’s ataxia. The finding that frataxin protein physically associates with scaffold proteins Isu1/IscU places it squarely in the pathway of Fe-S cluster assembly. Fe-S clusters are essential cofactors for many proteins involved in cellular respiration, DNA repair, translation and other processes. Frataxin is conserved throughout evolution, being present in eukaryotes such as yeast and human and in some prokaryotes including E. coli. However, differences exist between the eukaryotic and prokaryotic forms of frataxin. The eukaryotic forms are critical for Fe-S cluster assembly whereas prokaryotic forms are more dispensable. We found that a key to this difference is a single amino acid in the scaffold protein Isu1 at position 141. Changes of the eukaryotic amino acid, Met, to prokaryotic amino acids, Ile, Leu, Cys, or Val, rendered mitochondria more frataxin-independent. No other changes were able to replicate this effect. Thus, Isu1 containing Met at position 141 may have coevolved with frataxin in eukaryotes, conferring frataxin-dependence. In contrast, the appearance of other amino acids at this position may have rendered prokaryotic cells less dependent on frataxin.
| Frataxin is a highly conserved protein that is found in both prokaryotic and eukaryotic organisms [1]. The protein was originally identified based on its connection to Friedreich’s ataxia, which is an inherited neurodegenerative and cardiodegenerative disease resulting from a deficiency of frataxin [2]. Recently a mitochondrial Fe-S cluster assembly protein complex was identified consisting of frataxin in association with the cysteine desulfurase Nfs1, the small eukaryote-specific protein Isd11, and the scaffold protein Isu1 [3,4,5]. This protein complex serves to synthesize Fe-S cluster intermediates on Isu1 for subsequent transfer to the myriad proteins that use Fe-S cluster cofactors [6,7]. Iron-sulfur cluster intermediates contain iron and sulfur bound to the Isu1 scaffold in a [Fe2S2] configuration [8]. Although the precise function of frataxin has not been defined, it probably plays a role in sulfur and/or iron donation to Fe-S cluster intermediates [9,10].
The entire process of Fe-S cluster biogenesis is highly conserved between eukaryotic mitochondria and prokaryotic organisms [6,7]. Similar components are present, and in many cases the corresponding proteins function as orthologs. Sulfur for Fe-S cluster synthesis is derived from cysteine via the action of a cysteine desulfurase (Nfs1 in yeast, IscS in bacteria). This enzyme binds the amino acid cysteine in a substrate-binding site via the pyridoxal phosphate (PLP) cofactor [8]. The bound substrate is subjected to nucleophilic attack by an active site cysteine present on a moveable loop of the protein, forming a persulfide (e.g. Nfs1-S-SH in yeast). The persulfide sulfur is then transferred to recipients including Isu1 and used in building Fe-S clusters [11]. Frataxin (Yfh1 in yeast, CyaY in E. coli) interacts with the cysteine desulfurase and may be involved in regulating the enzyme activity. Whereas a positive regulatory effect has been observed for the yeast or human proteins [12,13], a negative regulatory effect has been observed for the E. coli homolog [14], and this difference has still not been explained [15]. Isd11 is a small accessory subunit that interacts with the eukaryotic Nfs1 and is necessary for its cysteine desulfurase activity [16]. However, Isd11 is eukaryote specific, being entirely absent from prokaryotic lineages [17]. Iron combines with sulfur on the scaffold protein to form Fe-S cluster intermediates. The scaffolds (Isu1 in yeast, IscU in E. coli) are highly homologous proteins [18]. The iron donation step is poorly characterized, and frataxin has also been implicated in this step. Both yeast and E. coli frataxins bind iron with low affinity on acidic residues in vitro and interact with their respective scaffold proteins in vitro and in vivo, and thus they may participate in iron donation [19,20]. Electrons provided by ferredoxin (Yah1 in yeast, Fdx in E. coli) are needed for reduction of iron or sulfur during Fe-S cluster intermediate formation [21]. Following formation of the intermediate on Isu1 or IscU, coordinated by three critical cysteines in the protein backbone, Hsp70 chaperones and cochaperones (Ssq1 and Jac1 in yeast, HscA and HscB in E. coli) interact with the scaffolds, mediating Fe-S cluster transfer in an ATP-dependent manner [22].
In terms of phenotypes resulting from frataxin deficiency, however, eukaryotes and prokaryotes show major differences. Total lack of frataxin is lethal in humans and metazoans [4]. Deletion of the YFH1 gene in yeast is associated with extremely deleterious effects, including slow growth, oxidant sensitivity, heme deficiency and lack of Fe-S clusters [23,24]. In addition, frataxin deficiency is associated with a curious iron homeostatic phenotype characterized by constitutive and unregulated cellular iron uptake. Within the cell iron accumulates in mitochondria in the form of biologically unavailable ferric phosphate nanoparticles. This constellation of findings apparently results from defective Fe-S proteins in the iron-sensing machinery [25,26]. In contrast to the yeast mutants, the effects of frataxin deletion in E. coli are mild. The bacterial deletion strain shows normal growth and does not exhibit iron homeostatic abnormalities or sensitivity to oxidative stress, although in one report the protein level for respiratory complex I was reduced [27].
A spontaneously occurring mutation in a frataxin-deleted yeast strain was found to effectively bypass the severe Δyfh1 phenotypes, restoring normal growth, Fe-S cluster protein levels, iron homeostasis, heme synthesis, and oxidative stress resistance. The effect was conferred by the Met to Ile change of amino acid 141 in the scaffold protein Isu1 [28]. The altered Isu1 was able to bind and activate the Nfs1 cysteine desulfurase in the absence of frataxin, thus providing a possible explanation for the bypass activity [29,30]. Interestingly, isoleucine is the amino acid utilized by E. coli in the homologous position of IscU. Thus in yeast lacking frataxin, the Met to Ile change in Isu1, by substituting the E. coli amino acid at this position, effectively rendered yeast more frataxin independent and more "prokaryote like".
Here we have delved more into the genetics of this frataxin-bypass phenomenon, finding more prokaryotic features of Isu1 bypass mutants. Randomly selected Isu1 bypass mutants were confined to a single amino acid position, and the amino acids conferring bypass were all present in homologous prokaryotic proteins. The prokaryotic homologs were identified both in organisms with frataxin and in organisms without frataxin, underscoring the frataxin-independence associated with these particular scaffold mutants. The entire E. coli IscU (targeted to mitochondria with a leader sequence) conferred more frataxin-independence, whereas a reverse substitution, in which the eukaryotic amino acid Met was introduced at the same position, conferred more frataxin-dependence. Examination of the set of Isu1/IscU sequences in available databases also suggests that Isu1-Met appears almost exclusively in eukaryotes and likely coevolved with frataxin. The substituted Isu1-Ile probably mimics a conserved and ancient feature of the prokaryotic Fe-S cluster assembly complex, relieving the frataxin requirement.
The frataxin-bypass mutant was isolated as a spontaneously arising clone of more rapidly growing cells in a background of a frataxin-deleted haploid yeast strain (Δyfh1) [28]. The bypass activity was conferred by a single point mutation (ATG to ATA), changing the amino acid of codon 141 of ISU1 from Met to Ile. However yeast S. cerevisiae carries two redundant genes coding for Fe-S cluster scaffolds, ISU1 and ISU2 [31]. The initial presentation of the suppressor or bypass mutant suggested that it was genetically dominant, because it was able to bypass Δyfh1 even in the presence of a normal copy of ISU2.
In order to evaluate this genetic dominance further, matched Δyfh1 strains were compared, one expressing a single copy of substituted ISU1 and deleted ISU2, called Δyfh1 [ISU1-Ile] (Table 1) and another expressing both ISU1 and ISU2 in addition to the plasmid-borne substituted Isu1-M141I, called Δyfh1 [ISU1-Ile] ISU1 ISU2 (Table 1). As assessed by colony formation, both strains showed improved growth compared with the Δyfh1 control (Fig 1A, compare rows 3 and 4 to row 2), although the single copy [ISU1-Ile] showed slightly better Δyfh1 suppression activity as assessed by growth (Fig 1A, compare rows 3 and 4), and other phenotypes such as iron uptake and Fe-S cluster levels. Thus, a high degree of genetic dominance of the Isu1 Met to Ile substitution was observed in producing reversal of Δyfh1 phenotypes.
ISU1 and ISU2 encode highly homologous proteins, with 83% amino acid identity. Isu1 and Isu2 proteins are functionally redundant, although the endogenous expression level of Isu1 is roughly seven times higher than that of Isu2, and Isu1 represents most of the cellular Isu protein in a wild-type strain [32]. Significantly the critical Met (amino acid 141 in Isu1 or 133 in Isu2) is present in both proteins. The adjacent cysteine that functions as an Fe-S cluster ligand and the adjoining frataxin-binding motif are also present in both proteins [33]. ISU2 and ISU2-Ile, in which the Met-133 was changed to Ile, were experimentally evaluated. Results show that ISU2-Ile conferred frataxin-bypass activity in a YFH1 shuffle strain (S1 Fig). Bypass activity was conferred whether it was expressed from the native ISU2 promoter or from the ISU1 promoter (S1 Fig). Thus, genetic dominance for frataxin-bypass by ISU2-Ile was observed, similar to that of ISU1-Ile. Furthermore, the ISU2 and ISU2-Ile constructs were also able to support growth of the GAL1-ISU1/Δisu2 strain indicating that they were functional as Isu. The reason that only the ISU1-Ile was isolated and ISU2-Ile was not isolated in the original screen for Δyfh1 suppressors is probably due to the lack of saturation of this genetic screen.
All 20 possible amino acids were substituted at position M141 in a single copy plasmid-borne ISU1. Each mutant form of ISU1 was tested for frataxin-bypassing activity by transforming the plasmids into a YFH1 shuffle strain, followed by counterselection with fluoroorotic acid on raffinose medium to remove the covering plasmid (Fig 1B, plates 1–3). After counterselection, robust growth was noted for positive control YCplac22-YFH1, and slow growth was noted for the empty plasmid, YCplac22 (Fig 1B, plate 1, “YFH1” versus “-”), consistent with the important role of frataxin for normal growth. The ISU1-Ile plasmid with the Met to Ile change restored robust growth in the absence of YFH1 (Fig 1B, plate 1, “I”). Similarly, plasmids substituted at the same amino acid position and containing ISU1-Leu, ISU1-Cys, and ISU1-Val, conferred improved growth (Fig 1B, plates 1–3). These ISU1 alleles thus carried frataxin-bypass activity and were genetically dominant, given that wild-type genomic copies of ISU1 and ISU2 were still present.
Fe-S cluster assembly scaffold function was tested separately. The collection of ISU1 alleles was transformed into the GAL1-ISU1/Δisu2 strain. In this strain the redundant ISU2 was deleted and ISU1 was placed under control of GAL1, a galactose-dependent promoter. When the transformants were shifted to a non-inducing carbon source (i.e. raffinose-based medium), only plasmid-borne ISU1 was expressed, allowing scaffold function for each allele to be scored. The wild-type ISU1 (Fig 1B, plate 4, M) supported normal growth, and a large set of ISU1 plasmids with amino acid substitutions also supported normal growth, indicating that these were highly functional ISU1 proteins (Fig 1B, plates 4–6, I, L, A, N, C, F, Y, S, V, G, T and H). Importantly, all of the substitutions conferring frataxin-bypass activity (Fig 1B, plates 1–3, I, L, C, and V) were also functional ISU1 scaffold proteins. Another set of substitutions was partially functional as shown by slowed growth (Fig 1B, plate 5, D, E and W), and these mutants tended to accumulate iron, indicating that they were hypomorphic mutants in the Fe-S cluster assembly pathway. A small set of substituted alleles was completely non-functional and did not grow at all on the raffinose plates (Fig 1B, plates 4–6, R, Q, K and P).
The E. coli ortholog, IscU, was studied by NMR methodology [34] and found to assume interconverting disordered or structured conformations. When a conserved amino acid in the primary sequence, Asn 123 in the Isu1 numbering, was substituted with Asp, the conformation of the mutant protein was shown to be shifted preferentially to the disordered form. Alternatively when Asn 123 was substituted with Ala, the conformation was shifted to the structured form [34]. We wondered if the frataxin-bypassing activity could be related to one or the other of these conformations. The Asp change (N-D; presumably disordered) and the Ala change (N-A; presumably structured) were introduced into YCplac22-ISU1 and the respective plasmids were transformed into the YFH1 shuffle strain. However, neither conferred any bypass activity (Fig 1B, plate 3, N-D and N-A). The Asp 123 form conferred growth to the GAL1-ISU1/Δisu2 strain indicating efficient scaffold function, but the Ala 123 form was poorly functional in this assay (Fig 1B, plate 6, N-D versus N-A).
Isu1 with the Met to Ile substitution, ISU1-Ile, exhibited both frataxin-bypassing activity and scaffold activity. Therefore the question arose of whether both activities must necessarily reside in the same molecule. Alternatively, the ISU1-Ile might act on wild-type ISU1 protein, stimulating it to perform as the primary scaffold. The existence of a protein complex containing the suppressor ISU1-Ile protein and the wild-type ISU1 protein could provide a physical context for such stimulation to occur [35,36]. To begin to address this question genetically, the three cysteine residues of Isu1, presumed to be Fe-S cluster ligands [31], were individually replaced by alanine. The C69A, C96A and C139A forms of Isu1 were introduced into strain GAL1-ISU1/Δisu2, and they were unable to support growth on their own in non-inducing glucose-containing medium. Double mutants were then constructed in which the critical cysteine substitutions were combined with the suppressor ISU1-Ile change of residue 141 in the same molecule. The mutated ISU1 alleles, C69A-M141I, C96A-M141I and C139A-M141I, were tested for frataxin-bypass activity in the YFH1 shuffle strain, and no bypass activity was observed (Fig 1C, plate 2). Likewise the mutated ISU1 alleles were tested for scaffold activity by expression in the GAL1-ISU1/Δisu2 strain in glucose, and no complementation was observed (Fig 1C, plate 4). The frataxin-bypass activity of the Cys substitution of residue 141 raised the possibility that the cysteine at this location was functioning as an alternative Fe-S cluster ligand instead of Cys 139. In a genetic experiment, Cys 139 was replaced with alanine in the presence of the M141C substitution. However, in this case, neither scaffold activity nor bypass activity was observed (Fig 1C, plates 2 and 4, number 7), making it unlikely that the Cys substituted at position 141 can function as an alternative Fe-S cluster ligand. In summary, substitutions that abrogated Isu1 scaffold function by altering the Fe-S cluster ligands also abrogated frataxin-bypass function.
A mutant allele of NFS1, nfs1-14, was identified because of its effects on iron homeostasis, and it was later found to be associated with decreased cysteine desulfurase activity in mitochondria [16]. The cells carrying this mutant allele were viable and grew well in most media despite the low activity of cysteine desulfurase. However, the combination of nfs1-14 and Δyfh1 was synthetically lethal (Fig 1D, YCp). YFH1 (Fig 1D, YCp-YFH1) introduced into the nfs1-14 Δyfh1 strain restored growth. However, although ISU1-Ile was able to bypass Δyfh1 alone, it was unable to rescue the nfs1-14 Δyfh1 double mutant (Fig 1D, YCp-ISU1-Ile). These data suggest that a threshold level of cysteine desulfurase activity is required for the frataxin-bypass activity of ISU1-Ile. Recent biochemical data indicate that frataxin stimulates the cysteine desulfurase activity of Nfs1 [13]. Furthermore, the suppressor ISU1-Ile acts in a similar fashion to stimulate Nfs1 even in the absence of frataxin, perhaps explaining its bypass activity [29]. Thus it makes sense that in the absence of a sufficiently active Nfs1, bypass of Δyfh1 by the Ile-substituted ISU1 will not occur.
The frataxin-bypassing activity of ISU1 protein with changes of Met 141 to Cys, Ile, Val or Leu was shown by growth enhancement of the YFH1 shuffle strain. In this strain, the chromosomal ISU1 and ISU2 were still intact (Fig 1B). To examine and compare these effects in more detail, matched strains were constructed in which each of the bypassing ISU1 alleles was expressed in the absence ISU2 (Table 1). The starting strain was designated YFH1 [ISU1], indicating a YFH1 positive strain with plasmid carrying wild-type ISU1 (Met at position 141) (Table 1). Another strain was designated Δyfh1 [ISU1], indicating a strain deleted for YFH1 and carrying wild-type ISU1. Similarly, Δyfh1 [ISU1-Cys], Δyfh1 [ISU1-Ile], Δyfh1 [ISU1-Leu], and Δyfh1 [ISU1-Val], were used to designate deletions of YFH1 with the indicated ISU1 mutant alleles (Table 1). The growth of the YFH1 [ISU1] strain was more rapid than the Δyfh1 [ISU1] strain as shown by the larger colony size (Fig 2A, upper panel, rows 1 and 2). Hydrogen peroxide exposure exacerbated the Δyfh1 phenotype [37], further slowing growth and viability as a consequence of increased oxidative stress (Fig 2A, lower panel, rows 1 and 2). The ISU1 alleles with Cys, Leu, or Val at position 141, similar to the ISU1 M141I allele, improved growth of Δyfh1 under these conditions, including in the presence of hydrogen peroxide, although not quite to levels in the frataxin plus strain (Fig 2A, upper and lower panels, rows 1–6).
In terms of mitochondrial proteins, immunoblotting with anti-frataxin antibody confirmed the correctness of the genetic assignments: the YFH1 strain expressed frataxin (Fig 2B, Y-M, lane 1) and the Δyfh1 strains did not (Fig 2B, lanes 2–6). The ISU1 alleles were all expressed in mitochondria. The abundance of Isu1 protein was increased in the Δyfh1 [ISU1] strain in the presence of the ISU1-Met allele (Fig 2B, Δ-M, lane 2), but in the presence of the bypassing ISU1 alleles, Isu1 expression returned to normal, similar to the YFH1 positive strain (Fig 2B, lanes 3–6). The changes in protein abundance were several fold, and most likely attributed to Aft1/2 transcriptional effects and protein stability effects [32]. Nfs1 protein levels were unchanged in all the strains consistent with the lack of regulation of the protein level (Fig 2B), although cysteine desulfurase activity was strongly regulated (see below).
Cytochrome c, a heme protein, was undetectable in the Δyfh1 [ISU1] strain (Fig 2B, lane 2) compared with the wild-type (Fig 2B, lane 1). The various suppressor strains recovered significant levels of cytochrome c (Fig 2B, lanes 3–6). Cytochrome c is regulated transcriptionally and post-transcriptionally by heme availability [38]. The dramatic changes in protein abundance probably reflect changes in heme synthesis and steady state heme levels.
Iron homeostasis was examined using steady state labeling of growing cells with radioactive 55Fe. In YFH1 [ISU1] strain, cellular iron levels were appropriately regulated, whereas in the mutant Δyfh1 [ISU1] strain, by comparison, excess iron accumulated (Fig 2C, Cellular iron, Y-M versus Δ-M). The presence of the suppressor ISU1 mutant alleles, ISU1-Cys, ISU1-Ile, ISU1-Leu, or ISU1-Val, restored cellular iron levels towards normal (Fig 2C, Cellular iron, Δ-C, Δ-I, Δ-L, Δ-V). The radiolabeled cells were subjected to subcellular fractionation, separating mitochondrial and post-mitochondrial fractions (Fig 2C, Mitochondrial iron and Post-mito supernatant iron). The iron quantitation for these fractions resembled the patterns for whole cell iron. The radiolabeled mitochondrial iron showed features that were dependent on YFH1. In the YFH1 [ISU1] mitochondria, the predominant portion was solubilized by exposure to non-ionic detergents such as Triton X-100, whereas in the Δyfh1 [ISU1] mitochondria, proportionally more iron was recovered in the pellet following centrifugation in the presence of detergent (Fig 2C, Mitochondrial iron distribution, Y-M versus Δ-M). Most likely this effect reflects the accumulation of nanoparticles of ferric phosphate that is a hallmark feature of Δyfh1 mitochondria [25]. In mitochondria from the different suppressor strains the amount of insoluble iron was decreased compared with the control deletion strain Δyfh1 [ISU1], indicating an improvement in the biochemical properties of mitochondrial iron pools (Fig 2C, Mitochondrial iron distribution, Δ-M versus Δ-C, Δ-I, Δ-L, Δ-V).
Aconitase (Aco1), an abundant [Fe4S4] protein of mitochondria, was evaluated. An in- gel assay showed the aconitase enzyme activities in mitochondrial lysates. The control strain YFH1 [ISU1] (Fig 3A, upper panel, lanes 1 and 7, Y-M) showed concentration dependent activity. The frataxin minus strain Δyfh1 [ISU1] (Fig 3A, upper panel, lanes 2 and 8, Δ-M) had no detectable activity regardless of the concentration of lysate in the assay. The various suppressors (Fig 3A, upper panel, Δ-C, Δ-I, Δ-L, Δ-V, lanes 3–6 and lanes 9–12) recovered significant activity, although not entirely to wild-type levels. Aconitase protein (Fig 3A, lower panel) as detected by immunoblotting was present in YFH1 [ISU1] control (lanes 1 and 7), and was decreased in the frataxin-minus Δyfh1 [ISU1] (lanes 2 and 8) mitochondria, perhaps because the apoprotein was turned over more rapidly by mitochondrial proteases. Aconitase protein levels were restored in the suppressor strains (Fig 3A, lower panel, lanes 3–6 and lanes 9–12), consistent with recovery of enzyme activity.
The persulfide-forming activity in these mitochondria was tested as an indication of their cysteine desulfurase activity. Isolated and intact mitochondria were depleted of endogenous nucleotides and NADH by incubation at 30°C for 10 min, thereby blocking Fe-S cluster biogenesis without disrupting cysteine desulfurase activity [16]. After labeling with 35S-cysteine, mitochondrial proteins were separated by non-reducing SDS-PAGE, and the persulfide covalently bound to Nfs1 was visualized by autoradiography (Fig 3B, Nfs1-S-35SH). The signal was absent in nfs1-14 mitochondria, which were unable to form significant Nfs1 persulfide because of a hypomorphic mutation in Nfs1 (Fig 3B, lane 13) [16]. The specificity of the Nfs1-persulfide signal was further confirmed by immunoprecipitation with anti-Nfs1 antibody [16]. Several other radiolabeled bands were detected in mitochondria (Fig 3B), some of which were present in the nfs1-14 control and some of which were absent. Thus these background bands could be attributed to direct binding of 35S-cysteine, persulfide transfer from Nfs1 which was incompletely blocked, or binding of other reactive cysteine persulfides to mitochondrial proteins. The YFH1 [ISU1] mitochondria (Fig 3B, lanes 1 and 2, Y-M) had significantly more Nfs1 persulfide than the frataxin-minus mutant Δyfh1 [ISU1] (Fig 3B, lanes 3 and 4, Δ-M). Each of the suppressor mutants (Fig 3B, Δ-C, Δ-I, Δ-L, and Δ-V; lanes 5–12) recovered persulfide-forming activity in mitochondria. Differences among the suppressor alleles were not apparent, and each one provided rescue of the persulfide-forming activity that was deficient in Δyfh1 [ISU1] mitochondria.
Next, 35S-cysteine labeling of isolated intact mitochondria was performed in the presence of added ATP, GTP, NADH and iron, thereby permitting multiple cycles of Fe-S cluster formation to occur [16]. Soluble proteins from these mitochondria were separated on native gels, and autoradiography was used to detect newly synthesized Fe-S clusters on aconitase (Fig 3C, Aco1 [Fe-35S]). In the”wild-type” YFH1 [ISU1] mitochondria (Fig 3C, lanes 1 and 2, Y-M) a strong signal was observed, whereas in the frataxin-minus Δyfh1 [ISU1] (Fig 3C, lanes 3 and 4, Δ-M) no signal was present, consistent with the important role of frataxin in mitochondrial Fe-S cluster assembly. In the various suppressor mutants, although they still lacked frataxin, Fe-S cluster synthesis on Aco1 was restored to 24–31% of the YFH1 control as assessed by densitometry and correction for the total amount of protein loaded (Fig 3C, lanes 5–12).
A library of randomly mutated ISU1 plasmids was generated by error-prone PCR and transformed into a YFH1 shuffle strain also deleted for ISU1. The colonies appeared uniform (Fig 4A, plate 1). The diversity of this library was confirmed by sequencing the inserts from randomly selected colonies. Of 42 colonies evaluated in this way, inserts were identified that included 72 amino acid changes in Isu1, which were well distributed throughout the coding region (S1 Table, controls). Transformants were replicated to cycloheximide plates, counterselecting against the YFH1-containing plasmid and uncovering the Δyfh1 phenotype. Most colonies grew slowly on glucose or not at all on raffinose, but a few colonies exhibited robust growth (Fig 4A, plates 2 and 3). Plasmid DNA rescued from these more rapidly growing Δyfh1 colonies was sequenced and in most cases was found to contain single nucleotide changes in ISU1 conferring substitutions of residue 141 of the coding sequence. In some cases, YFH1 sequences were found to have recombined into the ISU1 plasmid and these clones were discarded. The PCR randomization was then repeated with different ISU1 templates starting with Y141, H141 or F141, and a large number of colonies (approximately 36,950 representing 17,588 amino acid changes in Isu1) was screened in this way. The “hits” with frataxin-bypass activity included amino acid changes of M141 to Ile, Cys, Leu, and Val, sometimes in combination with other amino acids changes and sometimes alone (Fig 4B and S1 Table). However, no other amino acid change or combination of changes was able to confer frataxin-bypass activity. All possible amino acid changes in ISU1 were not sampled, and only single nucleotide changes at position 141 were selected. The failure to find amino acid substitutions with bypass activity at other locations in the Isu1 protein may derive from the many constraints on this essential scaffold protein, which must interact with multiple partner proteins and perform multiple functions, such as stimulating Nfs1 activity, coordinating Fe-S clusters and transferring Fe-S clusters [39]. Based on these results, the possibility of finding another ISU1 mutant with frataxin-bypass activity, while not entirely ruled out, seems unlikely.
The Met to Ile substitution in yeast ISU1 that conferred frataxin-bypass activity did so by altering the methionine at position 141 (107 in the signal-cleaved mature protein) to the amino acid isoleucine used in the E. coli IscU in the corresponding position (I108 in IscU). Interestingly, deletion of the homologous frataxin gene cyaY in E. coli gave milder phenotypes in that organism than in yeast. Slowed growth, iron accumulation, and oxidative stress sensitivity were not observed in the E. coli knockout [27] in contrast to the yeast knockout. A series of species cross-complementation studies was undertaken in order to test the relative frataxin dependence or independence of yeast expressing the entire E. coli IscU protein targeted to mitochondria. For comparison, IscU protein was reverse engineered to place Met at position 108, as in the eukaryotic Isu1. The IscU protein (authentic E. coli protein with Ile) and IscU-Met (substituted E. coli protein with Met) separately were fused to the leader sequence of yeast mitochondrial protein CoxIV. Each of these fusion constructs was transformed into the GAL1-ISU1/Δisu2 strain.
The E. coli IscU proteins, with the authentic isoleucine or with the substituted methionine, were able to function in yeast as indicated by complementation of GAL1-ISU1/Δisu2 cells. Each of these complemented strains grew well, indicating that the E. coli IscU or IscU-Met targeted to yeast mitochondria could function as the only Fe-S cluster assembly scaffold in the cell. No difference between iscU or iscU-Met expressing cells was noted in terms of growth (Fig 5A, rows 1 and 2). Frataxin was deleted in these strains, and a striking growth phenotype was observed. Both the Δyfh1 [iscU] or Δyfh1 [iscU-Met] could be maintained under an argon atmosphere with subtle differences in colony size on agar plates (Fig 5A upper panel). However, following air exposure, Δyfh1 [iscU] continued to grow with a doubling time of 2.5 h, whereas Δyfh1 [iscU-Met] progressively slowed until the doubling time reached 8.5 h in defined raffinose-based medium (Fig 5A, compare top panel for argon growth to bottom panel for aerobic growth in rows 3 and 4). This air/oxygen dependent growth inhibition was much more severe for the Δyfh1 [iscU-Met] strain than for the matched Δyfh1 [ISU1], carrying Δyfh1 and yeast ISU1-Met. Perhaps the hybrid yeast-E. coli Fe-S cluster assembly machinery is particularly oxygen sensitive in the absence of frataxin. One of the functions of frataxin could be to shield the Fe-S cluster assembly machinery from oxygen [10].
Isolated mitochondria were examined for protein expression by immunoblotting. The yeast Isu1 migrated as a single band of about 14 kDa in mitochondria (Fig 5B, lane 1). The CoxIV fusions with the E. coli proteins reacted strongly with antibody raised against the yeast Isu1, giving rise to a slower migrating doublet (Fig 5B, lanes 2–7). The slower migrating band co-migrated with the bacterial expressed and purified IscU at about 17 kDa, and therefore it likely represents the CoxIV signal sequence-cleaved form of the IscU fusion protein. The more rapidly migrating form may be a proteolytic product. The retarded gel mobility of E. coli IscU compared with yeast Isu1 may be explained by its lower pI (4.7 for the E. coli IscU versus 9.3 for the yeast Isu1), which is associated with decreased binding of SDS and slower migration in the gel [40]. We have observed a similarly aberrant slow migration of Yfh1 due to its many acidic residues and failure to bind SDS [41].
In the Δyfh1 [iscU-Met] mitochondria, IscU protein was markedly increased in abundance (Fig 5B, lanes 6 and 7). By contrast, in the Δyfh1 [IscU] mitochondria, the IscU protein level was comparable to that of the frataxin plus strains (Fig 5B, lanes 4 and 5, compare with lanes 2 and 3). The difference may be traced to defective Fe-S cluster assembly in the Δyfh1 [iscU-Met] cells. In cells with impaired Fe-S cluster assembly, Isu1 protein abundance was previously shown to be up-regulated due to increased iron-dependent transcription mediated by the Aft1/2 regulator, and decreased turnover mediated by the Pim1 protease [32]. Furthermore, the E. coli IscU proteins might be poorly recognized by the yeast Pim1, further slowing turnover and increasing abundance (Fig 5B, lane 1 versus lanes 2 and 3). Other mitochondrial proteins were also examined. Nfs1 protein levels were comparable in all cases, consistent with the lack of regulatory changes (Fig 5B). Yfh1 was detected in the cells consistent with the predicted genotypes, being present in YFH1 [ISU1], YFH1 [iscU], and YFH1 [iscU-Met] but absent in Δyfh1 [iscU] and Δyfh1 [iscU-Met] (Fig 5B). Cytochrome c, an indicator of cellular heme status, was present in the Δyfh1 [iscU] mitochondria expressing the authentic E. coli protein but completely undetectable in Δyfh1 [iscU-Met] mitochondria expressing the substituted form of the E. coli protein (Fig 5B, compare lanes 4 and 5 versus lanes 6 and 7).
Iron homeostasis was also evaluated. Two independent clones of Δyfh1 [iscU-Met] cells were tested because of the genetic instability and changeable phenotypes associated with this genotype. Both clones exhibited strongly increased cellular iron uptake compared with the unsubstituted control clones Δyfh1 [iscU] (Fig 5C, Cellular iron, bars 6 and 7). Iron accumulated in the post-mitochondrial supernatant and mitochondria of the Δyfh1 [iscU-Met] strains (Fig 5C, Post-mito supernatant and Mitochondrial iron, bars 6 and 7). The increase in mitochondrial iron levels was about 2–4 fold more than in the matched Δyfh1 [iscU] strains (Fig 5C, Mitochondrial iron, bars 4 and 5). Iron accumulated in both soluble and insoluble forms as assessed by centrifugation in the presence of Triton X-100, probably indicating ferric phosphate nanoparticle accumulation [25,26]. In all cases, the Δyfh1 [iscU-Met] mitochondria accumulated more insoluble iron than the Δyfh1 [iscU] mitochondria [Fig 5C, Mitochondrial iron distribution, bars 6 and 7 versus bars 4 and 5). In the frataxin-plus strains expressing E. coli proteins, YFH1 [iscU] and YFH1 [iscU-Met], iron homeostasis was mostly preserved (Fig 5C, all panels, bars 2 and 3). The most severe loss of iron homeostasis occurred in the Δyfh1 [iscU-Met] strain and correlated with the concurrent severe deficiency of Fe-S cluster proteins. The mechanism by which defective mitochondrial Fe-S cluster assembly perturbs iron homeostasis is still poorly defined. However, it is likely that important roles are played by loss-of-function of Fe-S cluster binding proteins such as Aft1/2 [42] and glutathione reductases [6].
Aconitase activity, measured by the in-gel assay of mitochondria, was present in Δyfh1 [iscU] mitochondria from two independent clones (Fig 5D, lanes 4, 5 and lanes 11, 12) and absent in Δyfh1 [iscU-Met] mitochondria from two independent clones (Fig 5D, lanes 6, 7 and lanes 13, 14). Aco1 protein was present in normal amounts in Δyfh1 [iscU] but markedly decreased in Δyfh1 [iscU-Met] mitochondria, consistent with increased turnover of the apoprotein (Fig 5B). By contrast, in frataxin plus mitochondria, aconitase protein and activity were present in all cases. Aconitase activity was greatest in the frataxin plus ISU1 expressing yeast mitochondria (Fig 5D, lanes 1 and 8), slightly less in the frataxin plus E. coli iscU expressing mitochondria (Fig 5D, lanes 2, 3 and lanes 9, 10), and slightly less again in the frataxin null iscU expressing mitochondria (Fig 5D, lanes 4, 5 and lanes 11, 12). Thus aconitase activity correlated well with other features of these strains, including growth, cytochrome c levels, and iron homeostasis.
Isu1/IscU entries in the public database RefSeq (6064 in total) were collected and aligned on the highly conserved 12 amino acid sequence LPPVK LH CSX LA, using the Muscle algorithm [43]. The sequences were then sorted according to the amino acid at position X, where X is Met in the yeast Isu1 and Ile in the Isu1 suppressor mutant (S2 Table).
Sorting on this position revealed highly interesting groupings. Firstly, we found that Isu1/IscU sequences with Met were present almost exclusively in eukaryotic species (302 of 307 entries, S2 Table). The converse was also true i.e. eukaryotic species had Isu1/IscU with Met present in almost all cases. The proteins with Met at position X were found in the most diverse branches of eukaryotes, including Excavata that lack classical mitochondria, Chromalveolata, various yeasts including Zygomycota, Basidiomycota, Ascomycota, land plants, photosynthetic single-celled organisms, various metazoans including worms, fish, flies, mice and humans (Fig 6). The only significant exceptions to the rule that Isu1/IscU with Met occurs in eukaryotes were several proteobacterial rickettsial species, including Holospora undulata, Neorickettsia risticii, Neorickettsia sennetsu, and Orientia tsutsugamushi. These organisms are intracellular parasites that are the closest living relatives of mitochondria. Interestingly, all these species have retained frataxin in their genomes [1], suggesting that IscU-Met and frataxin may have been co-inherited with the rest of the Fe-S cluster assembly machinery during the endosymbiotic event that gave rise to mitochondria (Fig 6) [1].
The lists of Isu1/IscU proteins using Cys, Ile, Leu, and Val at position X, i.e. those amino acid substitutions of yeast Isu1 conferring frataxin-bypass activity, included predominantly prokaryotic organisms. The Cys list had only 6 entries, all from bacteria, including several Clostridium species and an uncultured archeon (S2 Table). For the Ile list, almost all of the species were from prokaryotes (171 of 178 entries). The exceptions, i.e. eukaryotic species on this list, were interesting in that they often possessed more than one Fe-S cluster assembly scaffold gene. For example, the eukaryotic bumble bee, cucumber, armadillo and a single type of Mediterranean fly each carried Isu1 proteins with Ile at position 141, but the same organisms possessed another gene which used Met, as is typical for almost all eukaryotes. The Val column had almost exclusively prokaryotic or archaeal species (341 of 348 entries). In this column, the only eukaryotic species included Entamoeba histolytica and Giardia intestinalis, organisms that are known to have acquired Fe-S cluster assembly components from bacteria by gene transfer [44]. A single eukaryotic plant species, wheat or Aegilops tauschii, was found in this column. However, wheat also had another Isu gene using Met, and so it follows the theme of retaining more than one type of Isu, perhaps for adaptive advantages. Along the same lines, Bos mutus, a yak with a high altitude habitat, carried an Isu with Ile but also retained the more typical eukaryotic Isu with Met. The IscU proteins with leucine were found predominantly in prokaryotic species (181 of 199 entries). The outlier eukaryotic species on this list included selected apicomplexa (e.g. Plasmodium falciparum, P. vivax, P. yoelli), fungi (e.g. Cryptococcus and Aspergillus species) and mitosome containing microsporidia (e.g. Nematocida parisii).
The diversity of Isu1/IscU proteins is further increased by the expression of different splice forms in some species. Alternatively spliced forms of the scaffold protein genes were found in humans and armadillo. In humans, the X1 splice form inserts an Arg, and isoform X4 inserts a Lys at position X. The Lys containing isoform has been associated with destabilized protein and development of a human disease characterized by exercise intolerance and mitochondrial myopathy [45]. The substitutions of the Met 141 in yeast Isu1 with Lys and Arg were tested, and neither one was able to support scaffold activity or frataxin-bypass activity (Fig 1B). Nonetheless, it is still possible that these splice forms with Lys or Arg amino acid substitutions could serve a special function or afford an adaptive advantage under some special circumstances or in specific tissues.
In summary, Isu1/IscU amino acid sequences with Met at position X of the motif LPPVK LH CSX LA were predominantly found in eukaryotes but also occurred in several Rickettsia species. All contained frataxin in their genomes. Isu1/IscU proteins with the amino acids Cys, Ile, Val or Leu at position X were present primarily in prokaryotes, some with and some without frataxin in their genomes (Fig 6).
Frataxin was initially identified by reverse genetics after the Friedreich’s ataxia disease gene was cloned in 1996 [2]. Soon afterwards the protein was linked to iron metabolism by characterization of the striking iron homeostatic phenotype of the yeast deletion strain [23]. However, a more detailed understanding of frataxin’s function has been elusive. Recently, frataxin has been convincingly implicated in mitochondrial Fe-S cluster assembly [10]. Frataxin interacts with components of the mitochondrial Fe-S cluster machinery, including Nfs1, Isd11 and Isu1 [5]. It is vital for formation of Fe-S cluster intermediates on the Isu1 scaffold, a key step in the Fe-S cluster assembly process [3]. A suppressor Isu1 carrying the amino acid substitution Met to Ile at position 141 can correct or bypass most of the Δyfh1 phenotypes that result from lack of frataxin [28]. The change introduces the amino acid used by E. coli in the highly homologous IscU protein. The Δyfh1 yeast deletion strain is severely compromised and slow growing, whereas the E. coli strain deleted for the homologous frataxin protein is mildly compromised and grows normally. Here we have delved into the phenomenon of bypass of frataxin deletion, performing a number of genetic and biochemical experiments.
The suppressor ISU1-Ile bypassed the requirement for YFH1 when it was introduced on a plasmid. Similarly, the corresponding substitution in a plasmid carrying the paralogous ISU2-Ile conferred bypass activity. The effects were observed with chromosomal ISU1 and ISU2 genes remaining intact, thereby reflecting genetic dominance and suggesting a gain of function. It is therefore interesting to consider more specifically the nature of the gained function. One explanation could be that Nfs1/Isd11 exhibits only basal cysteine desulfurase activity, and that frataxin is needed to act as a positive effector, thereby inducing the optimal activated level of cysteine desulfurase [13,29]. The wild-type Isu1 has no stimulatory activity on its own and may even be inhibitory [13,33], but the suppressor Isu1-Ile is able to substitute for frataxin by stimulating the Nfs1/Isd11 cysteine desulfurase [29,30]. If both wild-type Isu1 and suppressor Isu1-Ile are present in the cell simultaneously, the stimulatory effect of the suppressor is the dominant effect, providing bypass activity. An implied consequence of this scenario is that the suppressor Isu1-Ile will require an active form of Nfs1 for it to be effective. Thus it makes sense that the hypomorphic allele of NFS1, nfs1-14 [46], with decreased cysteine desulfurase activity, would not support bypass. The results provide genetic support for the hypothesis that frataxin and the bypass Isu1 work to produce their effects on Fe-S cluster assembly, at least in part, by boosting the activity of the cysteine desulfurase.
The suppressor Isu1 was shown to stimulate persulfide formation on Nfs1, similar to frataxin [29,30]. Subsequently, sulfur for Fe-S cluster synthesis must be transferred to the Isu1 scaffold and assembled with iron to form the Fe-S cluster intermediate. It is generally assumed that this process occurs in a protein complex with other Isu1 molecules present [35,36]. Therefore, we wondered if the Nfs1 stimulatory effect of the suppressor Isu1 could promote Fe-S cluster formation in trans on a normal copy of Isu1 or Isu2. However, at least in a series of genetic experiments, this was not the case. The survey of all the possible M141 amino acid substitutions identified a set of best scaffolds, moderate scaffolds and poor scaffolds, based on complementing activity in the GAL1-ISU1/Δisu2 strain (Fig 1B). The same set of plasmids, scored for frataxin-bypassing capability, identified activity for M141 with Cys, Ile, Leu or Val changes, and these all fell into the top category for scaffold activity. Furthermore, if a second site substitution abolishing scaffold activity (e.g. changing a critical Cys to Ala) was introduced into the M141I-Isu1 protein sequence, and the doubly substituted Isu1 was introduced into a Δyfh1 shuffle strain with wild-type ISU1 and ISU2 present, bypass activity was abrogated (Fig 1C). Thus most likely the suppressor Isu1 does not productively interact with the wild-type Isu1 to mediate bypass, but instead it replaces the wild-type Isu1 in providing both bypass and scaffold functions.
Eukaryotic and prokaryotic Fe-S cluster machineries are highly conserved. Both yeast and E. coli utilize cysteine desulfurases, scaffold proteins and frataxin homologs. However major differences in the frataxin deletion phenotypes have been reported, with essentiality or severely deleterious phenotypes in the eukaryotic case [4], and normal growth and relatively mild phenotypes in E. coli [27]. Why the difference? One possibility is that E. coli possesses a redundant Fe-S cluster assembly system, the SUF system, which may compensate for lack of frataxin [47]. However this does not entirely account for the phenotypic differences, because the SUF system is not generally deployed under standard growth conditions. Significantly, the cysteine desulfurases show key differences. Most prokaryotic cysteine desulfurases such as IscS are constitutively active, although some regulatory changes in activity have been described [15]. The eukaryotic cysteine desulfurases, on the other hand, seem to be largely inactive in their basal state. Activation is required, and this activation involves frataxin. Purified Nfs1 is able to bind the substrate cysteine in its PLP containing substrate-binding site. However, binding is inefficient, and frataxin interaction increases exposure and utilization of substrate-binding sites of the enzyme [29]. The suppressor Isu1-Ile protein is able to generate a similar alteration of Nfs1, mimicking the effect of frataxin on Nfs1, and providing a plausible explanation for its frataxin-bypass activity. Nfs1 enzyme with the substrate bound must still undergo another activation step, mediated by Isd11. Isd11 triggers a conformational change and persulfide formation, thereby generating the intermediate for Fe-S cluster assembly [29]. Here we have seen that not only the Ile but also the Cys, Val, Leu substituted forms of Isu1 are able to stimulate persulfide formation on Nfs1 in the absence of frataxin. Thus these bypassing alleles of Isu1 activate Nfs1 independently of frataxin, rendering the mitochondria more prokaryote-like.
More than 17,000 amino acid changes of ISU1 were surveyed, but only a small subset of those conferred bypass activity. In all cases, the active changes altered the amino acid at position 141 of Isu1, introducing the amino acids Cys, Ile, Leu, or Val. The mutagenesis was not exhaustive, but nonetheless, the data support the very restricted nature of the changes that confer this activity. The biochemical features of these newly discovered bypass mutants (Isu1 with Cys, Leu, or Val substituted at position 141) were similar to the original one (Isu1 with Met replaced by Ile). The various mutant forms of Isu1 conferred improved growth in the absence of frataxin and more efficient Fe-S cluster assembly in mitochondria. The persulfide-forming activity was increased in mitochondria lacking frataxin, indicating enhanced cysteine desulfurase activity.
How does the Isu1 suppressor work? Isu1 is a central component of the Fe-S cluster assembly complex consisting of Nfs1/Isd11/Isu1/Yfh1. The components are highly conserved with their bacterial homologs, with the exception of Isd11. Structural information has been obtained only for the bacterial components [11,36]. For Isu1/IscU the structure includes alpha helices framing a platform of beta sheets, with three conserved cysteines oriented towards a binding pocket in the core and able to coordinate the Fe-S cluster intermediate (Fig 7). The amino acid motif LPPVK is found towards the beginning of a long C-terminal alpha helix [11,39]. Interestingly, the PVK motif (Fig 7, green in Fe-S scaffold) was shown to include the frataxin-binding site [33]. The suppressor residue Ile (Fig 7, red ball-and-stick) is predicted to lie on an exposed surface of this helix, opposite the Fe-S liganding Cys, which is found on the opposite interior face of the helix (Fig 7, blue ball-and-stick). Thus the Met to Ile amino acid change near to this frataxin-binding site on Isu1 might mimic the effects of frataxin binding. Substitutions of Cys, Val, or Leu would be predicted to produce similar changes.
Conformational changes of the Fe-S cluster assembly complex facilitating Fe-S cluster intermediate formation might ensue. Nfs1 might be altered in a way that exposes substrate-binding sites for interacting with cysteine [29]. (Fig 7, yellow balls for PLP in the binding site). The sulfur intermediate, converted to a persulfide and bound to the flexible loop of Nfs1 (Fig 7, magenta dotted line and adjacent residues), might be more readily transferred to the modified Isu1. The Cys 139, an Fe-S cluster ligand on Isu1, that is one helix turn away from the Ile suppressor residue, might be rendered more accessible for persulfide transfer (Fig 7). Iron delivery remains the least characterized step in Fe-S cluster formation. Frataxin might facilitate iron entry into the Fe-S cluster assembly complex by initiating a conformational change that promotes transfer of mitochondrial iron from a physiological ligand (still undefined) to Isu1. Alternatively, frataxin might play a more direct role in binding iron and delivering it to Isu1 as has been shown for Yfh1 [19]. The Isu1-Ile protein (or bacterial IscU) might accomplish a similar function by exposing iron-binding sites on Isu1 [19]. The iron delivery step in Fe-S cluster assembly will need to be better understood to support or to refute these possibilities.
Iron homeostasis was improved, correlating with the improvements in cytochrome c, a mitochondrial heme protein. In previously published work, cytochromes were virtually absent in Δyfh1 and restored by the ISU1-Ile allele as assessed by low temperature spectra of whole cells [28]. Other heme proteins such as cytochrome c peroxidase, Ccp1, were similarly affected [30]. Heme synthesis, measured by 55Fe incorporation into porphyrin in isolated mitochondria, was very low in Δyfh1 and recovered in the presence of the substituted ISU1 [28]. What is the mechanistic connection between a change in the Fe-S cluster assembly scaffold and these effects on heme synthesis? Heme synthesis occurs inside mitochondria and involves the insertion of iron into porphyrin by the enzyme ferrochelatase to make protoheme. Porphyrin and ferrochelatase activity are not lacking, and in fact, zinc protoporphyrin accumulates in the Δyfh1 mutant [25]. Instead iron is the likely source of the trouble. Iron accumulating in Δyfh1 mitochondria (and in other Fe-S cluster deficient cells) exhibits changes in its physical properties and solubility that could lead to loss of bioavailability [25]. The data suggest that Fe-S cluster synthesis acts upstream of heme synthesis in promoting iron bioavailability for heme synthesis, although many mechanistic details are still lacking.
Isu1/IscU amino acid sequences sorted according to the amino acid appearing in position 141 fall into clearly defined grouping, underscoring the importance of this residue for Isu1/IscU function. The amino acid Met appeared almost exclusively in eukaryotic organisms. The only significant exceptions were several prokaryotic Rickettsia species, which were classified as proteobacteria. In all of these organisms with IscU-Met, frataxin was also present in their genomes. We can imagine a scenario in which IscU-Met and frataxin originated together in proteobacterial ancestors of mitochondria such as Rickettsia, and from there gave rise to modern mitochondria via horizontal gene transfer during the endosymbiotic event (Fig 6). A key point is that the Met amino acid was found only in organisms with frataxin, as if Met serves to “lock in” frataxin by ensuring frataxin dependence of Fe-S cluster assembly. The co-dependence of these two elements, the IscU-Met and frataxin, appears to be quite profound, as they have been retained together throughout all the eukaryotic branches of the tree of life. In the prokaryotic world, on the other hand, various IscU variants were found, and various amino acids were found at position X adjacent to the PVK frataxin-binding motif of the IscU homologs. The amino acids Cys, Ile, Leu, or Val which conferred frataxin-bypass in biochemical experiments in yeast, were found in IscU proteins of species with or without frataxin homologs (Fig 6). Thus the evolutionary record suggests that these amino acids may be associated with relative frataxin independence of Fe-S cluster assembly. The advantages of retaining IscU-Met and frataxin versus IscU-Ile and no frataxin remain to be ascertained, as both arrangements are able to support efficient and regulated Fe-S cluster assembly.
Friedreich’s ataxia is a progressive degenerative disease affecting neurons such as dorsal root ganglia and cardiomyocytes, and certain other tissues. The disease is caused by frataxin deficiency in affected tissues and is associated with defective Fe-S cluster assembly [24,48]. The efficacy of frataxin-bypass in yeast can be viewed as a reprogramming of mitochondrial Fe-S cluster assembly to a more prokaryotic type, such that the cysteine desulfurase and other features of the assembly process become more frataxin-independent. The discovery of several substitutions, changes to Cys, Ile, Leu, or Val, that are able to confer frataxin-bypass suggests that there may be common structural features of these Isu proteins that could be mimicked by small molecules [49]. The genetic dominance of the suppressor activity in the Δyfh1 yeast also suggests that such an approach could be effective in a therapeutic setting in human mitochondria where frataxin is deficient and normal Isu1 is still present.
A set of plasmids was generated containing modified versions of the Isu1 coding sequence (between NdeI and XhoI), carried between the native Isu1 promoter (700 bp between EagI and NdeI) and terminator (200 bp between BamHI and SacI) on a centromere based plasmid, YCplac22. Residue 141 of the full length Isu1 precursor protein was changed from M to each of the other 19 amino acids, using QuikChange mutagenesis (Agilent Technologies Inc.). In addition, N123D and N123A mutants were generated. Cysteine mutants of the Isu1 coding sequence were constructed in which C69, C96 and C139 were each changed to A. The M141I change was then introduced into each of these cysteine mutants, creating plasmids C69A-M141I, C96A-M141I, and C139A-M141I. A cysteine swap mutant was created in which C139A was combined with M141C in the full length Isu1. The various mutant forms of Isu1 were tested for Yfh1-bypassing function and scaffold function by transforming into strains YFH1 shuffle and GAL1-ISU1/Δisu2, respectively (Table 1). For testing genetic dominance of the ISU1 bypass suppressor, the suppressor allele (YCplac22-ISU1-M141I) was introduced into strain 70–31 containing genomic copies of ISU1 and ISU2 (Table 1), and the covering YFH1 plasmid was removed by fluoroorotic acid (FOA) treatment. For testing ISU2 function, the native ISU2 sequence was amplified from plasmid pGP564-ISU2 including 700 bp 5’ and 200 bp 3’ of the coding sequence, and cloned between restriction sites HindIII and BamHI in YCplac22, making plasmid YCplac22-ISU2. The M133 was changed to Ile by site directed mutagenesis, creating YCplac22-ISU2-M133I. The ISU2 coding sequence was inserted into NdeI-XhoI sites in place of the ISU1 coding creating YCplac22-ISU2coding. The M133 was changed to Ile by site directed mutagenesis creating plasmid YCplac22-ISU2coding-M133I.
Strain GAL1-ISU1/Δisu2 was transformed with plasmid YCplac22-ISU1 in which M141 was changed to C, I, L or V and the chromosomal GAL1 promoter was turned off by shifting cells from galactose to glucose as the carbon source. The YFH1 gene was deleted by transforming with plasmid pRS405-gamma-yfh1 linearized with BamHI, and selecting for transformants on defined leucine drop-out medium in an argon-filled chamber. These low oxygen conditions were previously shown to mitigate the mutant phenotype and allow for stable propagation of the knockouts [30]. The knockouts were confirmed by PCR of the YFH1 locus. The strains were denoted Δyfh1 [ISU1] (115–26), Δyfh1 [ISU1-Cys] (116–54), Δyfh1 [ISU1-Ile] (115–28), Δyfh1 [ISU1-Leu] (116–53), and Δyfh1 [ISU1-Val] (116–51). Congenic wild-type strain YFH1 [ISU1] was included as a control.
The strains were thawed from -80°C vials and inoculated to CSM-Trp/2% raffinose defined medium agar plates kept in an argon-filled jar. Cells were inoculated from the plates into liquid medium of the same composition. In general, small cultures of 50 ml were grown in argon-filled bottles for two days without shaking, expanded to 100 ml cultures while shaking in air, and then diluted again into 1 L cultures supplemented with 10 μM ferrous ascorbate. The 1 L cultures were grown at 30°C for 16 h, and the total time exposed to air was approximately 24 h for the slow-growing strains.
Cellular and subcellular iron levels were determined by growing cells for at least four doublings in standard defined medium supplemented with 58 nM 55FeCl3, 10 μM unlabled ferric chloride and 100 μM ascorbic acid. Cells were washed free of unincorporated iron, and then they were ruptured by vortexing with glass beads in the presence of 50 mM Hepes/KOH, pH 7.5, 150 mM NaCl, 0.6 M sorbitol. Differential centrifugation was used to remove unbroken cells and to separate the remainder into mitochondrial and post-mitochondrial fractions [50]. The post-mitochondrial fraction included both cytoplasm and vacuoles, as no effort was made to separate these two cellular components. Mitochondria were shown to be mostly intact by evaluation of mitochondrial marker proteins, which remained with the mitochondrial fraction. Mitochondria were lysed for 10 min at room temperature in the presence of 0.1% Triton X-100 in hypotonic buffer (50 mM Hepes/KOH, pH 7.5, 150 mM NaCl). The supernatant (soluble) and pellet (insoluble) portions were separated by centrifugation at 20,000 x g for 30 min. Iron content was determined by scintillation counting for 55Fe, and protein content was measured by bicinchoninic acid assay (BCA, Pierce) [28]. For whole cells, iron content was reported as pmol iron per million cells, whereas for the cellular fractions iron content was reported as pmol iron per microgram protein. An in-gel activity assay for aconitase was performed. Briefly, mitochondria were lysed in buffer consisting of 50 mM Tris-HCl pH 8, 50 mM NaCl, 1% TX-100, 10% v/v glycerol, 2 mM Na-citrate and 15 U catalase. Samples were loaded on a native acrylamide gel containing 132 mM Tris base, 132 mM boric acid, 3.6 mM sodium citrate. The signal was developed in the gel by incubating in developing buffer containing 100 mM Tris-HCl, pH 8, 1 mM NADP, 2.5 mM cis-aconitic acid, 5 mM MgCl2, 1.5 mM methylthiazolyldiphenyl-tetrazolium bromide (MTT), 0.3 mM phenazine methosulfate, and 5 U/ml isocitrate dehydrogenase [51,52].
Persulfide formation on Nfs1 present in intact mitochondria was measured as described [16]. Isolated mitochondria were depleted for endogenous nucleotides and NADH by incubation for 10 min at 30°C in buffer (20 mM Hepes/KOH, pH 7.5, 0.6 M sorbitol). These mitochondria were incubated with 35S-cysteine (10 μCi) for 15 min at 30°C. Mitochondria were recovered, proteins were separated by non-reducing SDS gel, and the persulfide was viewed by radioautography. Fe-S cluster formation was measured as described [12]. Mitochondria were incubated with 35S-cysteine for 30 min in the presence of added ATP (4 mM), GTP (1 mM), NADH (5 mM) and iron (10 μM ferrous ascorbate). Mitochondria were recovered and the soluble proteins were released by freeze-thaw and sonication and separated on a native gel. The signal associated with newly formed [Fe-35S] aconitase was visualized by radioautography.
Pools of mutagenized linear fragments of 1443 bp including the coding region of the ISU1 gene were generated by mutagenic PCR in the presence of 1 mM MnCl2 and altered nucleotide concentrations (2 mM dATP, 2 mM dGTP, 10 mM dTTP, 10 mM dCTP). The primers used were: A (5’ TTTTTTCGGCCGTTCTTTTCTTTTTCTTGCACTACC 3’) and B (5’ TGATTTGAGCTCagcacgtccgtcccgctttcaccctgg 3’), and the templates were different YCplac22-ISU1 plasmids with M, Y, H or F at position 141 of the coding region. Each mutagenized pool was co-transformed with a gapped plasmid (pRS416-ISU1 digested with MscI and BamHI) into the YFH1 shuffle strain 109–9 (Δyfh1::TRP1 Δisu1::HIS3MX6 ISU2 cyh2 [pRS318-CYH2-LEU2-YFH1], Table 1). After selecting for transformants on uracil drop-out medium with glucose as the carbon source, the colonies were replicated to uracil drop-out, cycloheximide medium with glucose or raffinose as the carbon source. The large colonies appearing after several days on the raffinose plates were analyzed further. Colony PCR was used to check for absence of YFH1. Transformants still harboring YFH1 after counterselection were discarded. The remaining clones were expanded, and the plasmid-borne ISU1 alleles were rescued in E. coli and the DNA was sequenced.
The coding region of E. coli IscU was amplified from genomic E. coli DNA from strain DH5 alpha and inserted into the XbaI and XhoI sites of a YCplac22 derived plasmid. In this plasmid, the ISU1 promoter consisting of 700 bp, is followed by the first 22 amino acids of the cytochrome c oxidase subunit IV (CoxIV) mitochondrial signal sequence [53], and 200 bp of ISU1 terminator. The resulting plasmid was checked by DNA sequencing and confirmed to code for a CoxIV-IscU fusion protein with amino terminus as follows: MLSLRQSIRFFKPATRT^LCSSRHMAYSEKVID. The caret indicates the predicted signal sequence cleavage site by the mitochondrial processing peptidase, and the bolded letters indicate amino acids of E. coli IscU. The rest of the IscU sequence was confirmed to be the same as listed in Genbank accession No. AAJU02000016.1. The amino acid I108 of IscU (equivalent to M107 in signal-cleaved mature Isu1 or M141 in the full length precursor form of Isu1) was changed from I to M by QuikChange mutagenesis, generating a plasmid for expressing IscU-Met in mitochondria. The mitochondrial-targeted iscU plasmids were introduced into strain GAL1-ISU1/Δisu2, generating strains YFH1 [iscU] and YFH1 [iscU-Met]. YFH1 was deleted in these strains by transforming with pRS405-gamma-yfh1, linearized at BamHI, followed by selection on leucine drop-out medium in an argon-filled anaerobic jar. This created strains Δyfh1 [iscU] and Δyfh1 [iscU-Met] (Table 1). Deletion of YFH1 was verified by PCR. Strain GAL1-ISU1/Δisu2 transformed with YCplac22-ISU1 served as the control strain YFH1 [ISU1]. All strains were grown in an argon-filled chamber or in argon-bubbled medium, and cells were then exposed to air in defined raffinose medium for iron labeling and mitochondrial isolation.
Isu1/IscU related sequences were collected by using BLAST similarity to the Isu1 of Saccharomyces cerevisiae S288c from the RefSeq non-redundant database. All related entries in RefSeq (6064) were aligned using the Muscle program [43], according to the amino acid at position X (equivalent to Isu1 residue 141) of the amino acid motif LPPVK LH CSX LA. The lists were manually edited, and duplicates were removed, retaining one entry for each species. For each entry, the sequence was scored + if it contained 8 or more amino acids identical to the query, and 0 if contained 7 or less. The sequences were scored * for exceptions if they deviated from the rule that eukaryotic Isu proteins use only methionine at that amino acid position and prokaryotic Isu proteins do not use methionine at that position.
|
10.1371/journal.pcbi.1006883 | 3D computational models explain muscle activation patterns and energetic functions of internal structures in fish swimming | How muscles are used is a key to understanding the internal driving of fish swimming. However, the underlying mechanisms of some features of the muscle activation patterns and their differential appearance in different species are still obscure. In this study, we explain the muscle activation patterns by using 3D computational fluid dynamics models coupled to the motion of fish with prescribed deformation and examining the torque and power required along the fish body with two primary swimming modes. We find that the torque required by the hydrodynamic forces and body inertia exhibits a wave pattern that travels faster than the curvature wave in both anguilliform and carangiform swimmers, which can explain the traveling wave speeds of the muscle activations. Notably, intermittent negative power (i.e., power delivered by the fluid to the body) on the posterior part, along with a timely transfer of torque and energy by tendons, explains the decrease in the duration of muscle activation towards the tail. The torque contribution from the body elasticity further clarifies the wave speed increase or the reverse of the wave direction of the muscle activation on the posterior part of a carangiform swimmer. For anguilliform swimmers, the absence of the aforementioned changes in the muscle activation on the posterior part is consistent with our torque prediction and the absence of long tendons from experimental observations. These results provide novel insights into the functions of muscles and tendons as an integral part of the internal driving system, especially from an energy perspective, and they highlight the differences in the internal driving systems between the two primary swimming modes.
| For undulatory swimming, fish form posteriorly traveling waves of body bending by activating their muscles sequentially along the body. However, experimental observations have shown that the muscle activation wave does not simply match the bending wave. Researchers have previously computed the torque required for muscles along the body based on classic hydrodynamic theories and explained the higher wave speed of the muscle activation compared to the curvature wave. However, the origins of other features of the muscle activation pattern and their variation among different species are still obscure after decades of research. In this study, we use 3D computational fluid dynamics models to compute the spatiotemporal distributions of both the torque and power required for eel-like and mackerel-like swimming. By examining both the torque and power patterns and considering the energy transfer, storage, and release by tendons and body viscoelasticity, we can explain not only the features and variations in the muscle activation patterns as observed from fish experiments but also how tendons and body elasticity save energy. We provide a mechanical picture in which the body shape, body movement, muscles, tendons, and body elasticity of a mackerel (or similar) orchestrate to make swimming efficient.
| During the undulatory swimming of fish, a backward-traveling wave of body bending is formed to push against the water and generate propulsion. Muscle is the executor of the neural control and the source of mechanical power in fish swimming. Therefore, how muscles are used is a key question in understanding the control and mechanics of fish swimming and has been a focus multidisciplinary research over the past decades.
Experimentally, muscle activation during swimming is measured using electromyography (EMG) for various fish species ([1–5], for a review see [6]). During steady swimming, a common pattern emerges: the muscle elements are activated as a wave traveling posteriorly, but this EMG wave travels faster than the curvature wave. Consequently, the phase difference between the curvature and EMG waves varies along the body, which is known as “neuromechanical phase lags”. Nonetheless, the details of the muscle activation pattern vary among species. For anguilliform swimmers such as eels, the speed difference is not large, and the duration of the muscle activation on one side of the body is approximately half of the undulation period [3]. For carangiform swimmers such as carp, the propagation speed of the EMG onset is much higher than that of the curvature wave, whereas that of EMG termination is even higher, resulting in a decrease in duration towards the tail [4]. The EMG activity, along with the muscle contraction kinetics, the strain and the volume of the active muscle, can determine the absolute muscle power output along the body. With this approach, Rome et al. [5] showed that for scup, the power is generated mostly by the posterior part of the body.
To understand the muscle activation patterns and underlying mechanical principles of internal driving, researchers previously studied the internal torque and the corresponding power required. The sign of the torque has been used to predict which side of the muscle should be activated. Using theoretical models, namely resistive force theory [7], elongated body theory [8], and 3D waving plate theory [9, 10], previous studies obtained torque waves that travel faster than the curvature waves and qualitatively explained the neuromechanical phase lag. However, since positive and negative torques both occupy half of the period throughout the body, the decrease in the EMG duration in carangiform swimmers remains an obscure phenomenon.
Another approach used to understand the internal driving in the coupled system is to use neural control signals as an input and observe the kinematics emerging from the coupling of internal driving, the body, and the external fluid. Using resistive force theory and 2D computational fluid dynamics (CFD) with a prescribed uniform muscle activation, McMillen et al. [11] and Tytell et al. [12] studied lamprey-like swimmers and showed that the same muscle forces can generate body bending with different wavelengths, corresponding to varying magnitudes of the neuromechanical phase lags, depending on passive body properties such as stiffness. However, since the kinematics emerge from the coupling of many components, this type of approach may generate kinematics that do not match the experimental observations; therefore, the approach may cause difficulties for systematically studying the features of muscle activation and for explaining the differences in the muscle activation between species.
These previous modeling studies were all based on either theoretical models with strong assumptions or 2D CFD models, which cannot capture 3D flow around the top and bottom of the fish body and the jet left behind and 3D shapes for carangiform swimmers [13]. Therefore, while the qualitative explanations from the models are reasonable, the errors in these predictions are difficult to estimate.
To study the features of muscle activations among different species and elucidate the underlying mechanical principles, we use 3D CFD simulations to investigate the torque patterns and power output patterns for a typical anguilliform swimmer and a typical carangiform swimmer. By combining the simulation results with experimental observations, we aim to explain the features and their variations in the EMG patterns among fish with different swimming modes.
Treating water as an incompressible viscous fluid and the fish as moving bodies with prescribed deformations, we developed two-way coupled 3D models for the swimming of an eel and a mackerel (see Fig 1).
The carangiform body is modeled based on the actual anatomy of a mackerel, whereas the anguilliform body is created from a lamprey computed tomography (CT) scan (see [14] for details). Except for the caudal fin, the fins are neglected for the swimmers. The lengths of the fish bodies (L) are used as the unit length in the simulations. The bodies are meshed with triangular elements, and some sharp and small structures from the scan are removed to avoid instability in the CFD computation. After obtaining the surface data of the two fish, we reshaped the fish and remeshed the surface grid such that our code could accommodate the boundary between the fish and the fluid. The sharp and thin tail of the mackerel was modeled as a zero-thickness membranous structure. The number of surface mesh points was 3962 for the eel and 2127 for the mackerel (including 1962 for the mackerel’s body and 165 for the tail). The body mass (M) was computed by assuming a uniform distribution of density equal to the fluid density and was 1 in nondimensional units. M = 0.0019 for the eel and M = 0.0101 for the mackerel.
The kinematics for undulatory locomotion are generally in the form of a posteriorly traveling wave with the largest wave amplitude at the tail. To describe the deformation of the fish bodies, centerline curvatures κ are prescribed in the form of κ(s, t) = A(s) sin(ks − ωut), where s is the arc length measured along the fish axis from the tip of the fish head, A(s) is the amplitude envelope of curvature as a function of s, k is the wavenumber of the body undulations that corresponds to wavelength λ, and ωu is angular frequency. We use the undulation period as the unit of time; thus, ωu = 2π. The amplitude envelope A(s) for the anguilliform kinematics has the form A(s) = amaxes−1, where amax is the tail-beat amplitude. For carangiform kinematics, the amplitude envelope has the form A(s) = a0 + a1s + a2s2. The parameters for A(s) were adjusted to fit the envelope of the movement of real fish observed in experiments [8, 15]. The parameters used were amax = 11.41, and k = 2π/0.59 for the anguilliform swimmer and a0 = 1, a1 = −3.2, a2 = 5.6, and k = 2π/1.0 for the carangiform swimmer. To avoid generating spurious forces and torques in the interaction between the fish bodies and fluid, we added rotation and translation in the body frame of the swimmers to ensure that the movement of the bodies without external forces satisfied two conservation laws: linear momentum conservation and angular momentum conservation (see S1 Appendix for details). The resulting kinematics are shown in S1 Fig.
The in-house immersed boundary method code that is used is capable of simulating 3D incompressible, unsteady, and viscous flows in a domain with complex embedded objects including zero-thickness membranes and general 3D bodies [16, 17]. The flow is computed on a nonuniform Cartesian grid in x′y′z′ coordinates. The fluid domain has a size of 8.5 × 5 × 5, and a total of 620 × 400 × 400 ≈ 99 million points are used. The grid is locally refined near the body, with the finest spacing being 0.005 × 0.005 × 0.005. The fish models are placed in the center of the computational domain, and the body centerlines are in the z′ = 0 plane. A homogeneous Neumann boundary condition is used for the pressure at all boundaries. The flow speeds of the inlet flow at the front boundary is set as the swimming speed in the trial runs such that the model swimmers move only minimally in the computational domain. A zero-gradient boundary condition is used at all other boundaries. At the surface of the swimmers, nonslip boundary conditions are enforced. The time interval for the integration is 5 × 10−4.
Because the deformations of the bodies are prescribed, there are 6 degrees of freedom for the overall movement of the swimmers, the same as that of a rigid body. We computed the 3 degrees of freedom in the 2D plane of undulation from the fluid-structure interaction, namely, forward translation, lateral translation, and yaw motion. The velocity components of the swimmers are numerically integrated at the same time interval as the CFD based on Newton-Euler equations with forces and torques from the CFD. Those 3 degrees of freedoms related to the vertical direction are neglected but the force magnitude in the direction perpendicular to the plane of motion is on average less than 1/10 of the force magnitude in the plane of motion. Because the bodies of the swimmers are deforming, the governing equation for the angular degree of freedom is d(Iω)/dt = Ttot, where I is the moment of inertia, ω is the angular speed of the body, and Ttot is the total torque computed by integrating the contributions from the hydrodynamic forces on the surface of the swimmer. Since the deformation is prescribed, I and I ˙ are known. Therefore, ω can be obtained by numerically integrating ω ˙ = ( T tot - I ˙ ω ) / I while integrating other equations for the translational movement of the body and the flow of the fluid.
We set an initial swimming speed of 0.3 at the beginning of the simulation and waited at least two full cycles for the swimmer to reach steady swimming. All the data presented are collected from two periods. Because the swimming direction is not perfectly aligned with the -x′-axis of the computation grid, a new coordinate system is used such that the swimming direction is aligned with -x, y is the lateral direction, and the z-axis is the vertical direction. The Reynolds number is defined as Re = UL/νk, where U is the swimming speed, and νk = 1/15000 is the kinematic viscosity.
The force, internal torque, and power distributions along the fish body as a function of time are computed from the simulation. The force per unit length on the fish body, F, is calculated as follows: take an arc length Δs along the body centerline, and integrate all forces from every mesh point in Δs; then, divide the total force by the arc length Δs.
Considering the hydrodynamic forces, we compute the internal torque required to overcome the hydrodynamic forces and body inertia. The body elasticity and the other internal resistive forces are initially ignored and will be discussed later. The torque can be found by analyzing the force balance on either side of the body from the point of interest [18] and using the concept of inertial force. When the effect of acceleration on the torque of a segment is considered as inertial force (−mba), the effective external force can be considered F − mba, where mb is the body mass per unit length, a = v ˙ is the acceleration of the body segment, and the body is in static equilibrium. Then from the torque balance equation on either side of the body from the point of interest, we obtain the torque at the point of interest: T posterior ( s , t ) = - e z ∫ s 1 r × ( F - m b a ) d l or T anterior ( s , t ) = e z ∫ 0 s r × ( F - m b a ) d l, which is consistent with [9]. Although T = Tposterior = Tanterior theoretically, the relative error becomes significant at the ends where torques are small. To minimize the numerical error, we use a weighted average of the torques computed from both sides, namely, T = sTposterior + (1 − s)Tanterior.
The internal power by the torque and the power transferred to the fluid per unit length are computed as P T ( s , t ) = T κ ˙ [8] and PF (s, t) = −F ⋅ v, respectively, where κ ˙ is the time derivative of curvature, and v is the velocity of the body segment. The difference between the total power computed by integrating the internal power and the external power along the body is within the numerical error (< 3%).
We varied the kinematics (amplitude and wavelength) and the body shape (height and width) by 10% to examine the sensitivity of the results. We found that the force and torque patterns are qualitatively the same in these tests. Simulations with a smaller mesh size result in forces and torques within the numerical error. The detailed of the test parameters and the results are provided in S2 Appendix.
The free swimming speeds (U) are 0.285±0.004 and 0.245±0.005 in nondimensionalized units for the eel and the mackerel, respectively. The corresponding Strouhal numbers are 0.63 and 0.65. These values are consistent with previous numerical studies at similar Reynolds numbers (Re ≈4000) (e.g., [14]). For both fish, double row vortices are shed behind the tail, similar to previous numerical results (see Fig 1). The velocity field behind the mackerel clearly shows a backward flow, while a mean flow behind the eel in the fore-aft direction is not easily detected.
As expected from the input kinematics and body shapes, the forces are relatively uniformly distributed on the eel but concentrated on the tail of the mackerel (Figs 2 and 3A & 3C, S1 and S2 Videos). The fore-aft and lateral forces both show posteriorly traveling wave patterns similar to those of body bending, except at the head where the surface orientation rapidly changes. For the eel, the peaks in the force components near 0.7 body length correspond to an increase in the body height (in z direction) at that position. For the mackerel, the separation of the thrust and drag is clear: the tail generates most of the thrust, and the anterior part of the body generates drag at all times.
Because the phase of the force, especially the lateral force, is essential in determining the phase of the torque [18], we compare the phases of the lateral forces from the simulation with those of the velocity and the acceleration of the segments. In general, if the lateral force from the fluid on a segment is in phase with the negation of the segment velocity, it is a resistive-like force, and if the lateral force is in phase with the negation of the acceleration of the segment, it is a reactive-like force. We find that the phase of the observed lateral force on the body is closer to the phase of the negation of the acceleration except near the snout tips and the tail for the mackerel. In these regions, the phase of the lateral force is close to the negation of the velocity. In general, the forces on the fish are close to the predicted forces from elongated body theory, but discrepancies exist when the shape changes are rapid. Detailed discussions of the hydrodynamics underlying the force pattern are beyond the scope of this paper.
The torque required to overcome the hydrodynamic forces and body inertia in both species exhibits a traveling wave pattern moving posteriorly with a higher speed than the curvature wave (Fig 4). For the eel, the average speed of the torque wave (vT) is 1.41 in the nondimensionalized unit (body length/period). The traveling wave speed of the torque is even higher in the mackerel (vT = 2.11), exhibiting a nearly standing wave pattern. The torque wave speeds qualitatively match the observation that the EMG speed is much higher in carangiform swimmers [6]. The maximal value of the torque appears at approximately the middle of the body of the eel and slightly posterior to the middle point for the mackerel.
As shown in Fig 5, the power from the torque is mostly positive, indicating the energy output from the muscle, but negative values are observed on the posterior parts of both fish. For the eel, the power is nearly all negative for s > 0.7, similar to the case with a floppy body in a previous 2D study [12], whereas for the mackerel, the negative power is intermittent on the posterior part. The work over a cycle calculated by simply integrating the power is the minimal work needed, because the dissipation due to the internal resistance is not included; this method implies that the negative power transferred to the body is fully stored and recovered. The peak of this work per cycle is at the anterior part (≈0.4) for the eel and at a more posterior position for the mackerel (≈0.58), slightly posterior to the peak magnitude of the torque. We find that the work over a cycle is significantly negative on the posterior half of the eel body and slightly negative near the tail of the mackerel. If we assume that no energy-storing and energy-transmitting elements exist, then the work done by the muscles is the integration of only the positive power. We denote this quantity by W+. The differences between the two types of work per cycle are the greatest for the posterior part of the eel, indicating that power is lost if no spatial energy transfer is performed inside the body of the eel. The distribution of power transferred to the fluid from the body is relatively uniform on the eel but concentrated on the tail of the mackerel (cyan dashed lines in Fig 5B & 5D).
The mean total power Ptot averaged over a cycle is 2.2 × 10−4 (in nondimensionalized units) for the eel and 2.6 × 10−4 for the mackerel. If only the positive power is used, the power becomes P tot + = 8 . 6 × 10 - 4 and P tot + = 3 . 4 × 10 - 4, for the eel and the mackerel, respectively. The significant differences between Ptot and P tot + indicate a great potential to improve energetic efficiency through the spatiotemporal transfer of energy.
The torque pattern can be understood by applying the results obtained in a previous study [18]: the torque pattern in undulatory locomotion is determined mainly by the wavelength and phase of the lateral force relative to the lateral movement. The torque wave of the eel has a relatively low wave speed compared to that of the mackerel due to the short wavelength of the undulation. Because the phase of the force for the eel is overall close to the phase of the reactive force, the internal torque and power patterns are also similar to the patterns associated with pure reactive forces (Fig 6, left column). For the mackerel, the long wavelength of the curvature wave and the concentrated force on the tail result in nearly synchronized torques on the body. Because the force from the tail to the fluid is nearly in phase with the velocity, the rate of change of the curvature (κ ˙) and the torque are also nearly in phase. Consequently, the torque and power patterns are similar to the patterns associated with pure resistive forces (Fig 6, right column), and the internal power is nearly all positive.
Previous bending tests and experiences in the handling of fish indicate that the torque from the viscoelasticity of an eel body is significant but smaller than the torque generated by muscles [19]. For carangiform swimmers, since no muscles exist behind the peduncle region and the curvature is comparable (albeit greater) to the rest of the body, the torque from elasticity must be significant at least in the tail region. However, an accurate in vivo measurement of the body viscoelasticity distribution is not available. Therefore, we discuss the trend of the influences of the viscosity and elasticity individually when the elasticity or viscosity is small relative to the torque from hydrodynamics and the body inertia (Fig 7).
We assume that the magnitude of the torque from the body elasticity or viscosity is 40% of the torque at individual positions along the body, namely Te = 0.4〈T〉κ(s, t)/〈κ〉 or T v = 0 . 4 〈 T 〉 κ ˙ ( s , t ) / 〈 κ ˙ 〉, where “〈〉” means standard deviation over time. Then, the total torque that needs to be generated becomes T + Te or T + Tv. As shown in Fig 7, we find that the effect of elasticity on the torque is different along the body, separated by a position (s ≈ 0.5 for the mackerel and s ≈ 0.3 for the eel) where T and κ ˙ are in phase and the power is all positive. Anterior to that point, the torque magnitudes increase, and the torque wave speeds decrease; posterior to that point, the torque magnitude decreases, and the speed of the torque wave increases. For the mackerel, the torque wave can even reverse when the phase shift effect of the elasticity is strong. The reversal of the wave resembles the reversal of the wave of the offset of the EMG observed for carangiform swimmers [6]. As a result of the changes in the torque, the area of the negative power region in the posterior part of the body decreases, and W+ decreases (Fig 5D). This observation is consistent with the findings of previous studies that suitable elasticity can save and restore energy to improve efficiency (e.g., [20]). For the eel, the effect of the speed increase ends near s = 0.7 when the maximal curvature coincides with the minimal torque without elasticity. Therefore, the energy storage and release for the eel is in the middle part of the body (Fig 5B). Since the body viscoelasticity of the eel is weak and this effect is subtle, changes in the wave speed of the middle part of the torque wave or EMG are not obvious.
Since the body viscosity requires a torque that is in phase with the time derivative of the curvature, for both the eel and the mackerel, the resulting torques become more aligned with the time derivative of the curvature and hence have wave speeds closer to the speed of the curvature (Fig 8). Consequently, the negative power regions are reduced because the viscosity of the body always dissipates energy. The effects of the viscosity and elasticity are qualitatively the same as greater contributions, at least to a prefactor of 0.8 for Te and Tv (see S1 Appendix for details).
Although local elasticity can temporally transfer the energy flow into this region due to the fluid-structure interactions, the spatial transmission of such energy can only be achieved by other structures. In animals, coupled joint articulation by tendons over two or more joints is common and is an effective structure to save and transfer energy by connecting a joint with positive power and another joint with negative power [21, 22]. For carangiform swimmers, long tendons exist that span many vertebra [23]. Although the force transfer function of these tendons towards the tail has been experimentally confirmed, to our best knowledge, their function in saving energy has not been considered. We hypothesize that these long tendons are used to transfer energy from the posterior part to the middle part of the body when the negative power appears on the posterior part (Fig 5). This hypothesis can explain the observed decrease in the muscle activation duration among the carangiform swimmers, including some detailed features: the increase in the duration of the negative power from the middle of the body towards the tail matches the decrease in the EMG duration. The start of the positive power is aligned with the sign change of κ ˙ (the lines in Fig 4B), resulting in a low speed that is the same as that of the curvature wave. The end of the positive power is aligned with the sign change in the torque (the dashed lines in Fig 5B), resulting in a high speed that is the same as that of the torque wave. Such differences in wave speed qualitatively match the speed differences of the onset and offset of the EMG. Note that this hypothesis does not contradict the common view that force and energy are transmitted to the tail to interact with the fluid. Torque is still required when the power is negative on the posterior region and can be provided by the muscle in a more anterior position connected by the tendon. This hypothesis is also consistent with the observation that the EMG duration is nearly half of the undulation period on the whole body of anguilliform swimmers, which do not possess long tendons [23].
The energy transfer and savings by a tendon and the shortening of muscle activation can be further elucidated by a simplistic rope model (Fig 9). We take the positions s1 = 0.47 and s2 = 0.71 on the mackerel as an example. We assume that the designated muscles (muscle 1 and muscle 2 in Fig 9A) are attached to virtual struts with a height Hs. Based on experimental observations on the arrangement of muscles, tendons and vertebral segments in the posterior part of carangiform swimmers [24, 25], a pair of muscles on anterior position are connected to the posterior point by tendons (hereafter referred to as ‘tendon muscles’). A simple relationship between the muscle force Fm and the torque about the points on the centerline from a muscle Tm can be derived: Tm = HsFm. Correspondingly, the change in the muscle length and the curvature has the following relation: ΔLm = Lm − Lm0 = HsκΔs, where Lm0 is the muscle length at rest, and Δs is the arc length between the struts without bending. The power per unit length can be computed as P = F m L ˙ m / Δ s = T κ ˙. This relationship also holds for the tendon muscles. Because the height of the struts, the arc distance between the struts, and the resting length of the muscle do not affect the power, the exact values of these quantities are not important in this analysis.
We consider the case where the phases of T1 = T(s1), T2 = T(s2), κ ˙ 1 = κ ˙ ( s 1 ), and κ ˙ 2 = κ ˙ ( s 2 ) have the relation ϕ κ ˙ 1 ≈ ϕ T 1 > ϕ T 2 > ϕ κ ˙ 2 (Fig 9B). We assume a strategy in which the tendon muscles are only active when the torque required at these two points have the same sign and one of the tendon muscles generates a torque Td that is needed by both joints, namely, Td = sgn(T1) min(|T1|, |T2|). We first assume that any negative power of the muscles is wasted. Then, there are four stages (see Fig 9B & 9C):
In the example, because ϕ κ ˙ 1 < ϕ T 1, only stages I–III are present. The energy saved is 20% of the energy output from muscle 2 without the tendon and accounts for 60% of the total negative power that can potentially be saved. The remaining 40% of the energy is wasted in the tendon muscle due to the lengthening of the tendon-muscle system at the beginning of stage II (gray curves in Fig 9C). If the tendon is elastic, then the energy could be saved during the lengthening and be released during the shortening in the later stage II to further improve efficiency. The muscle activation at point 1 comes from the combination of muscle 1 and tendon muscle, which take a half period on one side of the body, and the muscle activation at point 2 only comes from muscle 2, which takes less than a half period (see the bars at the bottom of Fig 9B).
The low swimming speeds that we observed (compared with those of real animals) are likely due to the low Re used in our simulations. However, we argue that the results are qualitatively representative of real adult fish. First, a meta-analysis of previously reported fish swimming data indicates that the transition from the viscous regime to the turbulent regime occurs at a Re of several thousand [26]. Second, even the eel model in our study shows an inertia-dominated mode of swimming. Because the drag coefficient decreases with increasing Re in general, the speed of the simulated swimmer is expected to increase with increasing Re, and the contribution of the resistive force is expected to decrease for a real adult eel.
Using 3D numerical models, we provide the most accurate prediction of the torque and power required for hydrodynamic forces during the undulatory swimming of fish. By considering the torque and power transfer by tendons and the body viscoelasticity, we for the first time provide explanations for some long-standing questions in muscle activation patterns: the shortening of muscle activations in carangiform swimmers and reversal of the wave of the offset of EMG. Our study offers an integrative view of the function of the muscles as part of the mechanical system, highlights the differences in the internal driving of two primary swimming modes, and provides insights into the energy transfer and energy saving mechanisms by body elasticity and tendons in undulatory swimming. The numerical models developed and the mechanisms revealed in this study may guide the design of efficient bioinspired robots, especially soft robots with distributed driving systems and elastic bodies [27, 28].
|
10.1371/journal.ppat.1000940 | Epstein-Barr Virus-Encoded LMP2A Induces an Epithelial–Mesenchymal Transition and Increases the Number of Side Population Stem-like Cancer Cells in Nasopharyngeal Carcinoma | It has been recently reported that a side population of cells in nasopharyngeal carcinoma (NPC) displayed characteristics of stem-like cancer cells. However, the molecular mechanisms underlying the modulation of such stem-like cell populations in NPC remain unclear. Epstein-Barr virus was the first identified human tumor virus to be associated with various malignancies, most notably NPC. LMP2A, the Epstein-Barr virus encoded latent protein, has been reported to play roles in oncogenic processes. We report by immunostaining in our current study that LMP2A is overexpressed in 57.6% of the nasopharyngeal carcinoma tumors sampled and is mainly localized at the tumor invasive front. We found also in NPC cells that the exogenous expression of LMP2A greatly increases their invasive/migratory ability, induces epithelial–mesenchymal transition (EMT)-like cellular marker alterations, and stimulates stem cell side populations and the expression of stem cell markers. In addition, LMP2A enhances the transforming ability of cancer cells in both colony formation and soft agar assays, as well as the self-renewal ability of stem-like cancer cells in a spherical culture assay. Additionally, LMP2A increases the number of cancer initiating cells in a xenograft tumor formation assay. More importantly, the endogenous expression of LMP2A positively correlates with the expression of ABCG2 in NPC samples. Finally, we demonstrate that Akt inhibitor (V) greatly decreases the size of the stem cell side populations in LMP2A-expressing cells. Taken together, our data indicate that LMP2A induces EMT and stem-like cell self-renewal in NPC, suggesting a novel mechanism by which Epstein-Barr virus induces the initiation, metastasis and recurrence of NPC.
| Epstein-Barr virus (EBV) infects about 90% of people worldwide and persists benignly as a latent infection. However, EBV is associated with different types of human cancer. Nasopharyngeal carcinoma (NPC) is the most commonly known EBV associated cancer and expresses a well defined set of latent viral genes, including LMP2A, which has been detected in the majority of NPC samples. Several studies indicated this latent viral protein drove cellular invasion and metastasis. For this study, enforced LMP2A expressing NPC cell lines were generated. We show here that LMP2A induces an Epithelial–Mesenchymal Transition and increases the Stem-like Cancer Cells in NPC. Our results suggest that LMP2A supports tumor initiation and recurrence of the infected nasopharyngeal epithelial cells. For the first time we report a virus protein that functions in the initiation and progression of cancer by inducing the cancer stem-like cells. These findings permit a more detailed understanding of function and contribution to viral pathogenesis and provide a novel therapeutic target for NPC therapy.
| Nasopharyngeal carcinoma (NPC) is the most frequent head and neck tumor in Guangdong, South China, where the incidence peaks at 50 per 100,000, but is rare in the Western world (1 per 100,000) [1], [2]. NPC is a highly malignant cancer which often invades adjacent regions and metastasizes to regional lymph nodes and distant organs. Thirty to 60 percent of patients with NPC will eventually develop a distant metastasis. Although NPC tumors are sensitive to radiotherapy and chemotherapy, treatment failure is high due to local recurrence and distant metastases, which are the key contributors to NPC mortality [3]. However, the underlying cellular and molecular mechanisms of NPC metastasis and recurrence remain poorly understood.
The epithelial–mesenchymal transition (EMT) is characterized as a switch from a polarized, epithelial phenotype to a highly motile fibroblastoid or mesenchymal phenotype. EMT is critical to metazoan embryogenesis, chronic inflammation and fibrosis, and has been demonstrated to be a central mechanism in cancer invasiveness and metastasis [4]. Recently, Weinberg and colleagues reported that EMT generates cells with stem cell-like properties [5], which suggests that metastases are sometimes caused by cancer cells that acquire stem cell characteristics. Recent studies have also suggested that cancer stem cells (CSCs) represent a small proportion of the cells in a tumor mass and contribute to tumor initiation, metastasis and recurrence. It has been further reported that cancer stem cells are enriched in side population (SP) cells which can efflux the DNA binding dye, Hoechst 33342, from the cell membrane [6], [7], [8]. Most recently, Wang and colleagues have reported that SP cells in the human NPC cell line CNE2 display stem cell characteristics [9]. However, the molecular mechanisms underlying the regulation of SP cells in NPC remain unclear.
Epstein-Barr virus (EBV), which ubiquitously infects more than 90% of the world's population, was the first human tumor virus identified to be causally associated with various lymphoid and epithelium malignancies [10]. However, the underlying mechanism of how EBV contributes to cancer is still poorly understood. NPC, particularly the undifferentiated type, is the most commonly known EBV associated cancer [11] and three EBV latent proteins are expressed in these tumors [12], [13]. EBNA1, whose primary role is to enable replication of the viral episomal genome [14], is the most widely expressed protein in NPC. However, although both LMP1 and LMP2A are detectable in NPC samples, much of the recent research focus has been on LMP1 because of its known oncogenic properties in B cells [15], [16]. However, LMP2A has been detected in more than 95% of NPC samples at the mRNA level, and about 50% of these specimens at protein level, whereas LMP1 could be detected in only about 65% or 35% of NPC samples at mRNA or protein level, respectively [17], [18], [19], [20], [21]. In addition, the high LMP2A expression in NPC samples has been reported to correlate with a poor survival outcome, although this study was carried out using only a small cohort [22].
Functional studies indicate that LMP2A plays an important role in the maintenance of EBV latent infection in B cells but is dispensable for EBV-driven B-cell transformation [23]. In epithelial cells, LMP2A has been reported to have transforming properties i.e. to alter cell motility and inhibit cell differentiation [22], [24], [25], [26]. Activation of the PI3K/Akt, NF-κB, β-catenin, STAT and Syk Tyrosine Kinase pathways has been suggested to contribute to the various functions of LMP2A in epithelial cells and B cells [26], [27], [28], [29], [30]. ITGα6 is thought to be involved in the enhancement of cell migration mediated by LMP2A [22]. Most recently, LMP2A has been reported to induce promoter hypermethylation of the pten gene in gastric carcinoma [31]. In addition, some of the above functions and pathways modulated by LMP2A have been reported to play roles in regulating the proliferation and self-renewal properties of CSCs [32], [33], [34]. These findings thus raise the possibility that LMP2A may affect oncogenic processes by modulating the CSC population in NPC.
We report in our current study that the stable expression of LMP2A in NPC cells induces cell invasion and EMT-like molecular alterations. More importantly, the overexpression of LMP2A increases the size of the stem-like cell population and the number of tumor initial cells. Our data thus represent the first indication that LMP2A has an effect on stem cell-like populations and provides additional clues to elucidating the role of LMP2A in NPC progression.
To detect LMP2A protein expression, monoclonal antibodies (MoAbs) was raised against a glutathione-S-transferase-fused full-length LMP2A protein (Proteintech Group Inc.). After primary selection by ELISA, five clones were obtained from the Proteintech Group and further characterized by western blotting and immunofluorescence staining. To establish stably expressed LMP2A cell lines, CNE2 and SUNE1 cells were infected with virus expressing either LMP2A in the pBabe vector or with empty vector alone, followed by selection in puromycin. No differences in the efficiency of selection between vector and LMP2A-infected cells were observed. RT-PCR analysis showed that LMP2A mRNA was expressed in both of the LMP2A-infected cell lines (Figure 1A). The expression of LMP2A protein was detectable by immunoblotting with four different LMP2A MoAb clones. Representative results for clone 4A11B3A3 are shown in Figure 1A. In contrast to LMP2A-infected cells, there was no detectable LMP2A mRNA or corresponding proteins in the vector control cells. The membrane localization of LMP2A in the LMP2A-infected NPC cells was confirmed by specific detection with 4A11B3A3 using immunofluorescence staining (Figure 1B).
To further determine whether MoAb 4A11B3A3 could detect LMP2A protein in archival NPC patient's biopsies, we tested this antibody in paraffin-embedded nude mice xenograft samples. In accordance with our western blot results, by immunohistochemical analysis we found strong membrane staining of LMP2A in CNE2-LMP2A inoculated samples. No specific staining was observed in CNE2-vector inoculated samples or in IgG detected controls (Figure 1C). We then analyzed endogenous LMP2A expression in NPC patient biopsies with the same MoAb by immunohistochemical analysis and obtained similar results (Figure 1D). Hence, the specificity and sensitivity of this antibody for endogenous and exogenous LMP2A expression by immunoblotting, immunohistochemistry and immunofluorescence analysis was verified.
To further investigate the status of LMP2A expression in NPC biopsies, immunohistochemical analyses were carried out and revealed that 19 of 33 (57.6%) paraffin-embedded samples showed moderate (Figure 1D, right panel) to strong (Figure 1D, middle panel) staining of LMP2A in most of the tumor cells and in some scattered infiltrated lymphocytes. No positive staining was detected in adjacent noncancerous epithelial cells. As shown in Figure 1B and E, LMP2A is mainly expressed on the tumor cell membrane and preferentially located at the tumor invasive front. We then tested six archival relapse patient samples and found that were strongly positive for LMP2A expression. These data suggest that LMP2A is expressed in NPC samples at variable levels, that its localization at the invasive front is indicative of a potential role in promoting tumor invasion, and that the LMP2A protein levels may positively correlate with NPC recurrence.
It has been reported previously that LMP2A can promote the migratory/invasive properties of different epithelial cell types [35]. As determined by immunostaining, dissected tumor tissue samples from nude mice inoculated with CNE2-LMP2A cells showed a level of LMP2A that was comparable to that found in the NPC biopsies (Figure 1C and D). Hence, the established stable LMP2A expressing NPC cell line was found to contain physiological levels of LMP2A, and could thus be used in further studies of LMP2A function. Consistent with previous reports, the expression of LMP2A could enhance the migratory and invasive ability of NPC cells (data not show). Since the enhanced migratory/invasive ability of epithelial cells is often caused by EMT, we analyzed a panel of representative epithelial and mesenchymal markers by immunoblotting to determine whether this process occurs in LMP2A-expressing NPC cells. The results showed that the overexpression of LMP2A caused an EMT-like marker shift in the cells, including a dramatic downregulation of the epithelial markers E-cadherin and α-catenin, and upregulation of the mesenchymal markers fibronectin and the EMT-associated transcription factor snail, although the change of vimentin was moderate with about 2-fold increase in CNE2-LMP2A cells as analyzed by Quantity One software (Figure 2A). Immunofluorescence staining further revealed that the expression of E-cadherin and α-catenin, which shows membrane localization in control cells, dramatically decreased in LMP2A-expressing cells (Figure 2B, upper two panels). In contrast, the levels of fibronectin, vimentin and snail were strongly induced in LMP2A-expressing cells (Figure 2B, lower three panels). These results thus demonstrate that LMP2A induces EMT-like molecular alterations in NPC cells. However, similar to a previously reported observation in squamous epithelial cells [35], LMP2A did not induce any obvious morphological changes in NPC cells in monolayer cultures.
To exclude the potential effects of selection, we then examined the representative EMT markers after transient tranfection of LMP2A in NPC cells. As shown in Figure S1, EMT-like molecular alterations were induced by transient expression of LMP2A in both CNE2 and SUNE1 cells. To further investigate whether endogenous LMP2A contributes to the EMT phenomenon, we tested whether NPC cells lacking this endogenous expression demonstrated any EMT-like cellular marker reversal as compared with LMP2A-expressing cells. Following the knockdown of LMP2A in C666 cells (Figure S2A), we found by immunofluorescence staining that the expression of the epithelial marker E-cadherin was up-regulated, whereas the mesenchymal marker vimentin was down-regulated on the membranes of the cells (Figure S2B). These results indicate that LMP2A is necessary for the EMT-like marker shift in NPC cells.
It has been reported recently that EMT generates cells showing the properties of stem cells [5]. We thus determined whether stable expression of LMP2A could induce such stem cell-like phenotypes in NPC. Representative stem cell markers were thus analyzed by RT-PCR or western blot. As shown in Figure 3A (left panel), in comparison with the vector control, LMP2A expression up-regulates the stem cell markers ABCG2, Bmi-1, Nanog, and SOX2 at the transcriptional level. The increases in ABCG2, Bmi-1, SOX2 and Nanog were further confirmed at the protein level (Figure 3A, right panel). As expected, transient expression of LMP2A could also induce stem cell markers, as demonstrated by the increased expression of ABCG2 and Bmi-1 at both transcriptional and protein levels (Figure S3A).
Side populations (SPs) among NPC cells have been reported to exhibit cancer stem cell characteristics [9]. We wished therefore to determine whether the increased expression of stem cell markers we observed in LMP2A-expressing cells was caused by an increase in the size of the stem cell-like SPs. As shown in Figure 3B, the stable expression of LMP2A dramatically increases the size of the SP in the CNE2 (from 1.04% to 8.32%) and SUNE1 (from 3.38% to 13.72%) cell lines. Importantly, SPs were also increased in transient LMP2A expressing cells CNE2 (from 1.56% to 3.65%) and SUNE1 (from 3.91% to 11.37%) (Figure S3B). Interestingly however, we did not observe any SPs in either wild type or LMP2A knockdown C666 cells.
Previously, we have reported that the side population (SP) cells, isolated from CNE2 NPC cell line, exhibited cancer stem cell characteristics [36]. Thus, we sorted the SP fraction in CNE2-Vector, CNE2-LMP2A, SUNE1-Vector and SUNE1-LMP2A cells, respectively, and then performed colony formation assay. As shown in Figure S4, SP fraction from either LMP2A or vector control cells form larger and more colonies compared with the non-SP fraction, confirmed that the stem cell population is indeed within the SP fraction in NPC cell lines. Taken together, our results demonstrate that LMP2A could induce expression of stem cell markers and increase the stem cell population in NPC cells.
We next analyzed whether the increase in the sizes of the SPs in NPC is due to the enhanced self-renewal properties of the stem-like cells therein. LMP2A and control cells were cultured in suspension to generate spheres, the number and sizes of which reflect both the quantity and ability of cells to self-renew in vitro [36]. As shown in Figure 3C, LMP2A-expressing cells formed more and larger spheres than vector controls cells did in both NPC cell lines (CNE2, P = 0.04; SUNE1, P = 0.03). We conclude from this that LMP2A can indeed enhance stem cell self-renewal properties, and thereby increase the size of these populations.
To investigate whether LMP2A can enhance the transforming ability of NPC cells, we used both a colony formation and anchorage-independent growth assay in soft agar. We plated 200 NPC cells in triplicate wells of six-well plates for the colony formation assay. After 14 days of culture, LMP2A-expressing cells formed colonies that were significantly larger than those of the vector control cells (Figure 4A). There were also more LMP2A-expressing than vector control colonies. Statistical analysis showed significant differences in the number of colonies between the LMP2A-expressing and vector control cell lines (P<0.05; Figure 4A, right panel). In addition, the transforming ability of LMP2A expressing cells was also determined by soft agar assay. As shown in Figure 4B, LMP2A-expressing cells formed significantly more and larger colonies compared to the vector cells in soft agar assay.
As SPs are enriched for tumor initiating cells, we next assessed the effects of LMP2A upon the tumorigenicity of NPC cell lines in nude mice. As shown in Figure 5, when injected with 1×106 cells, the palpable tumors formed by LMP2A cells and control cells appeared at a similar time and grew at a comparable rate. As the injected cell number was reduced however (cell numbers at 1×105, 1×104 or 1×103), the growth rates of the LMP2A tumors were found to be higher than those of controls injected with the same cell numbers. The data in Figure 5B show that when injected with 1×105, 1×104 or 1×103 LMP2A-expressing NPC cells, 96% of the nude mice (27/28) developed tumors, whereas only 61% of these mice (17/28) did so when injected with the control cells. When 1×103 cells were injected, the control cells formed only small tumors in 5/10 mice after 20 days whereas LMP2A-expressing cells formed tumors in 10/10 mice. In addition, the first palpable tumor in the LMP2A groups injected with 1×103 cells appeared within 13 days, six days earlier than the control. Mice were sacrificed at 14, 17 or 20 days after injection, and the tumors were then weighed and photographed (Figure S5A and B). In all cases, the sizes of the tumors formed by the LMP2A-expressing NPC cells were larger than the vector control cells except in the 1×106 cell inoculation groups. This difference was most apparent in the 1×103 cell group (P = 0.004). Hence, LMP2A increases the number of tumor initiating cells in NPC.
To determine whether any correlation existed between LMP2A expression and the representative markers of EMT and stem cell in NPC biopsy samples, we obtained RNA from 15 inflammatory samples and 18 NPC samples and analyzed LMP2A, ABCG2, Bmi-1, E-cadherin (E-cad) and Fibronectin (FN1) expression using real-time RT-PCR. LMP2A, Bmi-1 and ABCG2 transcripts were found to be low or undetectable in the 15 inflammatory samples but extremely high in the NPC tumor tissue (Figure 6A). We also found that LMP2A expression positively correlates with ABCG2, Bmi-1 and Fibronectin, and negatively correlates with E-cadherin (Figure 6B).
In addition, we also detected LMP2A, Bmi-1, E-cadherin proteins in another 42 NPC biopsies. As shown in Figure 6C and Figure 6D, LMP2A correlated positively with Bmi-1, and negatively with E-cadherin.
As previously shown, the expression of LMP2A in B lymphocytes and HaCaT cells induces the activation of Akt in a PI3K-dependent manner [26], [28]. To investigate the Akt status in NPC cells in our current study, western blot analysis using an antibody that detects Thr308 phosphorylation of Akt was performed to detect activated Akt (Figure 7A). Phospho-Akt (Thr308) was found to be up-regulated at least 2.5 folds in LMP2A-expressing cells compared with control cells as analyzed by Quantity One. Phospho-GSK3β, a direct target of Akt GSK3β [37], was further found to be induced in LMP2A cells (Figure 7A). After treatment with Akt inhibitor (V) at 4 µM for 12hours, the phosphorylation of Akt was suppressed in both the LMP2A and vector control NPC cells (Figure 7B). It is noteworthy, however, that the SPs were dramatically reduced in NPC cell lines in the presence of Akt inhibitor (V), particularly in LMP2A-expressing cells. As shown in Figure 7C, the size of the SP decreased from 31% to 13.3% in CNE2-LMP2A cells and from 7.1% to 2.4% in SUNE1-LMP2A cells. Thus, the Akt pathway seems to play a role in the LMP2A-mediated increase of NPC SP cells, although this will need to be further confirmed using dominant negative mutants or shRNAs that target Akt in LMP2A-expressing cells.
Most importantly, these results were further confirmed by transient expression of LMP2A. Figure S6A, phospho-Akt (Thr308) and phospho-GSK3β were upregulated in LMP2A positive cells consistent with the above results. And after treatment with Akt inhibitor (V), we observed the similar results (Figure S6B). Moreover, as shown in Figure S6C, the size of the SP decreased from 13% to 7.06% in transient CNE2-LMP2A cells.
The results of our current study indicate a pivotal role for LMP2A in the progression of NPC via stem-like cancer cell induction. The novel functions of LMP2A in inducing EMT, increasing the size of the SP, enhancing the self-renewal properties of stem-like cancer cells, and increasing the number of cancer initiating cells, confirm the involvement of this protein in oncogenic processes through the modulation of the CSC population in NPC.
Many studies have recently focused on the function of LMP2A in B cells. Although LMP2A is not required for EBV transformation of human B cells [38], [39], [40], it can drive B-cell development and survival by mimicking B-cell receptor (BCR) signal transduction [23], [41], [42], [43], [44]. The profiling of genes that are involved in different biological signaling pathways in EBV-associated cancers, in comparison with those of LMP2A transgenic mice, indicates that LMP2A may play a key role in tumorigenesis [45], [46]. As reported, LMP2A can be detected in approximately half of Hodgkin's lymphomas [47], and routinely detectable in nasopharyngeal carcinoma [19], [20], [21], suggesting that it may play important roles in the induction of human cancers of EBV. Previous studies suggesting that LMP2A may be involved in cell proliferation have come from the Raab-Traub and Tsai laboratories, which have independently demonstrated that LMP2A has dramatic effects on epithelial cells, such as the ability to accelerate anchorage-independent growth, promote tumor growth in nude mice, inhibit epithelial cell differentiation, and activate cell motility [24], [26].
Consistent with previous findings [20], [21], [22], we observed LMP2A protein in 57.6% of the NPC specimens analyzed in our present study. More importantly, we found that LMP2A localized predominantly at the tumor invasive front, indicating its role in promoting tumor invasion and migration. Indeed, invasion and metastasis are major features of NPC. Many types of cancer cells derived from primary carcinoma appear to rely on the EMT program to facilitate most of the steps in the invasion-metastasis cascade, in which the down-regulation of E-cadherin is a key initial event [48]. However, the last step in this process, which is termed colonization, requires that the cells that migrate from the original tumor site possess the self-renewal capability to form the macroscopic metastases in addition to an EMT ability [5]. The significance of our present results is underscored by the fact we have demonstrated for the first time that the EBV latent membrane protein LMP2A can confer stem-like properties upon NPC cells whilst at the same time promote EMT. Our findings are in agreement with and support the previous observation of a direct link between EMT and the gain of epithelial stem cell properties [49], to our knowledge the first report of a connection of this nature. It will be important to perform a study in a larger cohort of samples in the future to further demonstrate the correlation between LMP2A and ABCG2 expression and the clinicopathological characteristics of NPC.
Acquisition of the EMT phenotype in epithelial tumor cells is believed to play critical roles in the increased invasiveness and metastatic potential of tumor cells, and this process is causative for the development of “cancer stem-like cell” characteristics. The evidence thus far suggests that tumors arise from a diseased stem cell derived from a progenitor cell population and that many cancers, not unlike normal organs, contain a small population of cells with a high proliferative capacity, self-renewing potential, multi-differentiation ability, and that are resistant to chemotherapy and radiotherapy. All of these properties are characteristic of normal adult stem cells and even embryonic stem cells. Thus, this subpopulation of cells is denoted CSCs or tumor stem cells [50], [51], [52].
Our current study demonstrates that the LMP2A induces a stem cell state, evidenced by an enhanced self-renewal and transformational capacity, and also increases the number of tumor initiating cells in vivo. This was further confirmed by the greatly increased SP size and higher expression at both the transcriptional and translational level of some stem cell markers, such as ABCG2, Nanog, Bmi-1 and SOX2. Compared with most non-tumorigenic cancer cells, SP cells have a strong ability to form tumors after transplantation. Because SP cells are also resistant to chemotherapy and radiotherapy, they may contribute to tumor relapse even after most non-tumorigenic cells are destroyed [53]. Recent studies indeed show that the transcription factors OCT4, SOX2, and Nanog have essential roles in early development and are required for the propagation of undifferentiated embryonic stem (ES) cells in culture [54], [55], [56], [57]. More importantly, the positive correlation of LMP2A and ABCG2 expression in NPC specimens provides a valuable clue to further elucidating the processes underlying clinical metastasis and recurrence in NPC.
It is tempting to speculate regarding the actual mechanisms by which LMP2A induces EMT and cancer stem-like characteristics. Previous studies showed that Syk inhibition could impair LMP2A-mediated cell migration; and the mutation of Tyr-74 and Tyr-85, the LMP2A ITAM, simultaneously blocked Syk activation and cell migration [30]. It also has been previously shown that the constitutive activation of the Ras/PI3-K/Akt pathway by LMP2A is a key element of LMP2A-mediated transformation, whereas the cell adhesion signaling and MAPK pathways are not activated in LMP2A tumors [25], [26]. Our current data also show that LMP2A activates Akt in NPC cells and that after treatment with Akt inhibitor (V) for 12 hours, the SP cell population decreased greatly in LMP2A-expressing cells. It is well known that NF-κB plays an important role in mediating the processes of EMT induced by different factors through the upregulation of the transcriptional repressor functions of ZEB1 and ZEB2, the zinc-finger E-box binding homeobox proteins essential in E-cadherin regulation [58], [59]. Moreover, the PI3-K/Akt/mTOR pathway could indirectly activate NF-κB activity by regulating glycogen synthase kinase (GSK)-3β phosphorylation [60]. The importance of PI3K/Akt signaling in the proliferation and maintenance of embryonic stem cell (ESC) self-renewal has previously been suggested [33], [34], and its involvement in the regulation of the representative pluripotency marker genes, Oct4, Sox2, and FoxD3, has also been reported in a recent study [61]. Further more, LMP2A induces expression of polycomb group protein Bmi-1, which has been recently reported to play an important role in progression of NPC by inducing EMT and maintenance of stem-like phenotype via PTEN/Akt/Snail signaling [62]. Taken together, these results suggest that the PI3K/Akt pathway, at least in part, contributes to the EMT process and also the SP stem-like cancer cell shift of NPC epithelial cells. However, the underlying mechanisms how LMP2A modulates the Akt activity and expression of Bmi-1 in NPC cells requires further investigation. Importantly, some effects induced by LMP2A on EMT are also induced by LMP1 [63]. It will be interesting to determine whether LMP1 and LMP2A collaborate in inducing EMT and stemness of NPC cells.
It must be noted that the lack of effective SP sorting of endogenous LMP2A knockdown cells limits our ability to determine whether LMP2A alone is sufficient to increase the size of the SP. In addition, the complex regulatory machineries associated with LMP2A during cancer stem cell induction have not been fully elucidated. Hence, elucidation of the underlying signaling network that regulates the LMP2A pathways by using various LMP2A mutants and different pathway inhibitors will provide important further insights into the exact role of this viral protein in the emergence of cancer stem cells.
We show for the first time herein that the EBV latent membrane protein LMP2A can induce EMT and increase the number of tumor initiating cells. Our data first indicated that LMP2A strongly up-regulates the cancer stem cell-like population in NPC, which may explain the onset of metastases and high rate of recurrence for these tumors. This raises the possibility that this viral protein plays a key role not only in EBV latency and persistence but also in the progression of NPC. Based on our novel findings, we believe that the pathologic diagnosis together with detection of LMP2A in tumor tissue will aid in predicting NPC progression, and that LMP2A can be considered to be a novel therapeutic target for this cancer.
All animal work was conducted under the institutional guidelines of Guangdong Province and approved by the Use Committee for Animal Care. Approval from the Sun Yat-sen University Institute Research Ethics Committee was obtained, and written informed consent was provided by each human subject.
Two poorly differentiated nasopharyngeal carcinoma cell lines (CNE2, SUNE1) were maintained in RPMI 1640 medium (Life Technologies, Carlsbad, CA) supplemented with 10% fetal bovine serum (FBS) in a humidified 5% CO2 incubator at 37°C. To generate stable cell lines, recombinant retroviruses expressing either vector pBabe or pBabe subcloned with LMP2A were generated as previously described [64] and used to infect CNE2 and SUNE1 cells [65]. Pooled CNE2 and SUNE1 cell populations expressing either pBabe or pBabe-LMP2A were selected with 0.5 µg/mL of puromycin (Sigma-Aldrich, St Louis, MO).
C666, the only well-known nasopharyngeal carcinoma cell line consistently carrying EBV, was chosen to perform the stable knockdown of LMP2A expression. Retroviral particles were generated and used to infect the target C666 cells as described previously [66]. The successful knockdown of LMP2A was verified by RT-PCR and immunofluorescence.
An LMP2A monoclonal antibody was obtained from Proteintech Group Inc. ABCG2 (Cat. 3380) and Nanog (Cat. 21603) antibodies were obtained from Abcam (Cambridge, UK). Antibodies raised against E-cadherin (Cat. 610181), α-catenin (Cat. 610193), fibronectin (Cat. 610077), and vimentin (Cat. 550513) were purchased from BD Biosciences (Franklin Lakes, NJ). Mouse anti-Bmi-1 (Upstate Biotechnology, Lake Placid, NY), and rabbit-anti-GSK-3β, p-GSK-3β, Akt (Cell Signaling, Beverly, MA) and p-Akt (Santa Cruz Biotechnology, CA.) primary antibodies, and FITC or rhodamine-conjugated goat anti-rabbit IgG or goat anti-mouse IgG (Jackson Laboratory, West Grove, PA) or Peroxidase-conjugated goat anti-rabbit IgG or goat anti-mouse IgG (Amersham Pharmacia Biotech, Piscataway, NJ) secondary antibodies were used for western blot or immunofluorescence analysis.
Freshly frozen biopsied tissues from a total of 18 NPC patients and 15 normal controls, and 81 paraffin-embedded NPC samples which had been histologically and clinically diagnosed were collected from the archives of the Department of Sample Resources, Cancer Center, Sun Yat-sen University (Guangzhou, China). Prior informed consent from the patients and approval from the Institute Research Ethics Committee was obtained.
Cells were analyzed by FACS when the cells had reached a logarithmic growth phase (24 hours after replating). Cells were digested with 0.25% trypsin (Sigma-Aldrich, St. Louis, MO), washed twice with calcium/magnesium-free PBS, resuspended in ice-cold RPMI 1640 culture (supplemented with 2% FBS) at a concentration of 1×106 cells/mL, and incubated at 37°C in a 5% CO2 incubator for 10 min. The DNA binding dye, Hoechst 33342 (Sigma-Aldrich, St. Louis, MO), was then added at a final concentration of 5 µg/mL and the samples were incubated for 90 min in the dark with periodic mixing. The cells were then washed twice with PBS, 1 µg/mL propidium iodide (Sigma-Aldrich) was added, and the cells were kept at 4°C in dark prior to sorting by a Moflo XDP (Beckman Coulter, Fullerton, CA). Because Hoechst 33342 extrudes from cells treated with verapamil (a calcium ion tunnel antagonist)-sensitive ABC transporters, a subset of the cells were incubated with 50 µmol/L verapamil for 30 min at 37°C before the addition of Hoechst 33342 to determine whether this would block the fluorescent efflux of SP cells in the CNE2 and SUNE1 populations.
Total RNA extracts from LMP2A-overexpressing cells and pBabe vector control cells were prepared using a Trizol reagent (Life Technologies, Grand Island, NY) according to the manufacturer's instructions. The RNA was then treated with DNase, and 2.5 µg aliquots were used for cDNA synthesis using random hexamers. The primers used for the amplification of the indicated genes are listed in Table S1.
The expression levels of LMP2A, ABCG2, BMI-1, E-cadherin and Fibronectin mRNA was determined by SYBR green real-time reverse transcription-PCR (RT-PCR). Total RNA from different human nasopharyngeal tissues were extracted using Trizol reagent (Invitrogen, Carlsbad, CA). Quantitative dertermination of RNA levels were performed in triplicate in three independent experiments. Real-time PCR and data collection were performed with an ABI PRISM 7900HT sequence detection system. The housekeeping gene GAPDH was used as an internal control to normalize the expression levels of different genes. The primers used for the amplification of the indicated genes are listed in Table S2.
Western blotting analysis was performed as previously described [67]. Where relevant, the blots were probed with antibodies as labeled in the figures, and the signals were detected using enhanced chemiluminescence (ECL) (Amersham Pharmacia Biotech, Piscataway, NJ). The membranes were stripped and probed with an anti-alpha tubulin mouse monoclonal antibody (Santa Cruz Biotechnology, Santa Cruz, CA) to confirm equal loading of the samples.
Immunofluorescence analysis was performed as described previously [67]. Cell lines were plated on culture slides (Costar, Cambridge, MA) and after 24 hours were rinsed with phosphate-buffered saline (PBS) and fixed in ice-cold methanol-acetone for 5 min at -20°C. The cells were then blocked for 30 min in 10% BSA (Sigma-Aldrich St. Louis, MO) in PBS and then incubated with primary monoclonal antibodies in PBS for 2 hours at room temperature. After three washes in PBS, the slides were incubated for 1 h in the dark with secondary goat anti-mouse, or goat anti-rabbit antibodies (Invitrogen, Carlsbad, CA). After three further washes, the slides were stained with 4-,6-diamidino-2-phenylindole (DAPI; Sigma-Aldrich St. Louis, MO) for 5 min to visualize the nuclei, and examined using an Olympus confocal imaging system (Olympus FV100).
Six-well plates were coated with a layer of 0.6% agar in medium supplemented with 20% fetal bovine serum. Cells were prepared in 0.3% agar and seeded in triplicate. The plates were then incubated at 37°C in a humid atmosphere of 5% CO2 for two weeks until colonies had formed. Each experiment was repeated at least three times. Colonies were photographed between 18–24 days (final magnification 20 X) under a phase contrast microscope, and colonies larger than 50 µm in diameter were counted under a light microscope.
Cells were counted, plated in triplicate at 200 cells for the pooled population or 100 sorted cells per well in six-well plates, and cultured with RPMI 1640 complete culture for 10 days. After most of the colonies had expanded to more than 50 cells, they were washed twice with PBS, fixed in methanol for 15 min, and dyed with crystal violet for 15 min at room temperature. After washing out the dye, the plates were photographed. To quantify the colonies objectively, the software Quantity One was used and colonies that lager than the averaging parameter of 3 or 1 and the minimum signal intensity of 1.0 were counted. At least three independent experiments were carried out for each assay.
Nude mice were purchased from the Shanghai Slac Laboratory Animal Co. Ltd and maintained in microisolator cages. All animals were used in accordance with institutional guidelines and the current experiments were approved by the Use Committee for Animal Care. Tumor cells were suspended in 200 µl RPMI 1640 complete culture with 25% Matrigel (BD Biosciences) and inoculated subcutaneously into the left flanks of 4- to 5-week-old nude mice. The mice were monitored daily for palpable tumor formation and tumors were measured using a Vernier caliper, and also weighed and photographed.
The Entrez Gene ID for genes and proteins mentioned in the text are 3783751 (LMP2A), 2597 (GAPDH), 10376 (α-Tubulin), 999 (E-cadherin), 2335 (Fibronectin), 7431(Vimentin), 1495 (α-Catenin), 9429 (ABCG2), 648 (Bmi-1), 79923 (Nanog), 6657 (SOX2), NG_012188 (Akt), NG_012922 (GSK-3β).
|
10.1371/journal.pgen.1007830 | Global transcriptional regulation of innate immunity by ATF-7 in C. elegans | The nematode Caenorhabditis elegans has emerged as a genetically tractable animal host in which to study evolutionarily conserved mechanisms of innate immune signaling. We previously showed that the PMK-1 p38 mitogen-activated protein kinase (MAPK) pathway regulates innate immunity of C. elegans through phosphorylation of the CREB/ATF bZIP transcription factor, ATF-7. Here, we have undertaken a genomic analysis of the transcriptional response of C. elegans to infection by Pseudomonas aeruginosa, combining genome-wide expression analysis by RNA-seq with ATF-7 chromatin immunoprecipitation followed by sequencing (ChIP-Seq). We observe that PMK-1-ATF-7 activity regulates a majority of all genes induced by pathogen infection, and observe ATF-7 occupancy in regulatory regions of pathogen-induced genes in a PMK-1-dependent manner. Moreover, functional analysis of a subset of these ATF-7-regulated pathogen-induced target genes supports a direct role for this transcriptional response in host defense. The genome-wide regulation through PMK-1– ATF-7 signaling reveals a striking level of control over the innate immune response to infection through a single transcriptional regulator.
| Innate immunity is the first line of defense against invading microbes across metazoans. Caenorhabditis elegans lacks adaptive immunity and is therefore particularly dependent on mounting an innate immune response against pathogens. A major component of this response is the conserved PMK-1/p38 MAPK signaling cascade, the activation of which results in phosphorylation of the bZIP transcription factor ATF-7. Signaling via PMK-1 and ATF-7 causes broad transcriptional changes including the induction of many genes that are predicted to have antimicrobial activity including C-type lectins and lysozymes. In this study, we show that ATF-7 directly regulates the majority of innate immune response genes upon pathogen infection of C. elegans, and demonstrate that many ATF-7 targets function to promote pathogen resistance.
| Convergent genetic studies of host defense of Drosophila melanogaster and mammalian innate immune signaling revealed a commonality in signaling pathways of innate immunity that has helped motivate the study of pathogen resistance mechanisms in genetically tractable host organisms such as Caenorhabditis elegans [1]. The simple C. elegans host has enabled the genetic dissection of integrative stress physiology orchestrating host defense of C. elegans[2–5]. Genetic analysis of resistance of C. elegans to infection by pathogenic Pseudomonas aeruginosa has defined an essential role for a conserved p38 mitogen-activated protein kinase pathway that acts on a CREB/ATF family bZIP transcription factor, ATF-7, in immune responses [6,7]. A complementary approach to characterizing the host response has been organismal transcriptome-wide characterization of genes induced upon infection by a number of different bacterial pathogens [8–15]. Putative effector genes encoding lysozymes and C-type lectin domain (CTLD)-containing proteins have been identified that have also served as useful markers of immune induction. Here, we report the genome-level characterization of the C. elegans response to P. aeruginosa that is mediated by ATF-7 activity downstream of PMK-1 activation, combining RNA-seq analysis of pathogen-induced gene expression with ChIP-seq analysis of ATF-7 binding, which suggests global regulation of the immune response of C. elegans through a single MAPK- transcription factor pathway.
We performed RNA-seq on wild-type (N2), pmk-1 mutant, or atf-7 mutant animals exposed to E. coli OP50 or P. aeruginosa PA14 to identify genes that are differentially regulated upon infection that also require PMK-1 or ATF-7 for induction (Fig 1A). We found that in wild-type animals, 890 genes were two-fold upregulated (adjusted p-value <0.05), and 803 genes were two-fold downregulated upon P. aeruginosa exposure, compared to animals exposed in parallel to E. coli (Fig 1B; S1 Table). Many of these upregulated genes have been previously implicated in the C. elegans immune response, including genes encoding C-type lectin domain (CTLD)-containing genes and lysozymes, corroborating prior microarray-based gene expression studies (Fig 1C) [8–10,12–14]. In contrast, gene ontology analysis of genes that are decreased in expression upon P. aeruginosa exposure shows enrichment for genes associated with homeostasis with significant ontology terms consistent with growth, development and reproduction (S1A Fig). Of note, many of the genes upregulated in response to P. aeruginosa exposure exhibit relatively low expression when animals are propagated on E. coli, whereas genes that are decreased in expression upon P. aeruginosa exposure are expressed at a higher basal level during normal growth conditions on E. coli (Fig 1D, S1B Fig). In parallel, we analyzed P. aeruginosa-mediated gene expression changes in pmk-1 and atf-7 mutants to identify the proportion of gene changes induced by P. aeruginosa exposure that required PMK-1 and/or ATF-7 for induction (S2 Fig). We classified genes that failed to significantly meet a two-fold expression change cutoff upon exposure to P. aeruginosa in the pmk-1 or atf-7 mutant background as being PMK-1- or ATF-7-dependent, respectively. We observed that 70% of genes significantly induced two-fold or greater by P. aeruginosa exposure were no longer fully induced upon loss of pmk-1, and that 53% of upregulated genes were no longer fully induced upon loss of atf-7 (Fig 1E, S1 Table). We also found that 41% of genes reduced two-fold or more by P. aeruginosa required PMK-1, and 50% required ATF-7 for reduction of expression (S2B Fig, S1 Table). These data suggest a high degree of involvement of PMK-1-ATF-7 signaling in the majority of changes in gene expression induced in response to infection by P. aeruginosa.
To evaluate the role of ATF-7 in the direct regulation of genes induced by P. aeruginosa infection, we performed chromatin immunoprecipitation followed by sequencing (ChIP-seq) of animals carrying a GFP-tag fused to the C-terminal end of the endogenous atf-7 locus. Using a GFP polyclonal antibody for immunoprecipitation, we generated ChIP binding profiles for animals in either the wild-type background (atf-7(qd328[atf-7::2xTY1::GFP])) or the pmk-1 mutant background (pmk-1(km25);atf-7(qd328[atf-7::2xTY1::GFP])) after a four hour exposure to either E. coli OP50 or P. aeruginosa PA14, for a total of four conditions, similar to the treatment described in Fig 1A. In all conditions analyzed, ATF-7 exhibited abundant association throughout the genome, with 8,962 total peaks identified as enriched by MACS2, which map to a +/- 1.5 kB region corresponding to 23.7% of genes and 24% of transcription start sites (TSSs) in the C. elegans genome (WS258) (S2 Table).
Analysis of the ATF-7 binding profile across all genes associated with enriched TSSs, as well as the subset altered in expression by P. aeruginosa in wild-type animals according to our RNA-seq data, revealed that ATF-7 is preferentially located at the promoter regions of genes that are increased in expression by P. aeruginosa, and that this enrichment for ATF-7 is lessened by pmk-1 loss (Fig 2B, S3A Fig). Strikingly, this dependence of ATF-7 promoter occupancy on PMK-1 exists on P. aeruginosa, but not on E. coli. This finding is consistent with our previously published model, whereby ATF-7 occupancy of its target promoters is not simply controlled by pathogen-dependent PMK-1 activity [7]. MEME analysis of the most enriched loci identified significant enrichment for the motif GACgTCA, which corresponds to the Jun D bZIP motif expected for ATF-7 (Fig 2A, S3B Fig). This motif is present in as many as 80% of the most highly enriched regions of the genome and its abundance is positively correlated with enrichment levels.
To identify the most likely immediate downstream targets of ATF-7, we set a peak threshold based on the fraction of peaks containing the bZIP motif after ranking ATF-7 peaks by enrichment (S3C Fig). This resulted in ~1500–4000 highly enriched locations per experiment. Overlap of the of the retained ATF-7 binding profile compared to the RNA-seq data from P. aeruginosa infection revealed consistent 2–4 fold enrichment of ATF-7 binding in genes that depend on ATF-7 for expression changes stimulated by P. aeruginosa exposure (S4A and S4B Fig). We further assessed the significance of these enrichments using a Gene Set Enrichment Analysis (GSEA), which showed that ChIP peaks were enriched for association with transcripts that are upregulated upon pathogen exposure in both E. coli and P. aeruginosa ChIP conditions (Fig 2C, S4C Fig). This association remains in the pmk-1 mutant, although at a weaker significance level (Fig 2D, S4D Fig). Moreover, we also evaluated ATF-7 binding at individual genomic loci induced by P. aeruginosa infection that were dependent on ATF-7 for full upregulation. Examinations of distinct genetic loci further support the conclusions drawn from the metagene analyses described above (Fig 3). These observations suggest a direct transcriptional regulatory role for ATF-7 in the induction of broad transcriptional changes upon immune challenge involving activation of p38/PMK-1 MAPK signaling in response to P. aeruginosa infection.
For functional validation of putative ATF-7-regulated immune response target genes, we focused on transcripts that were upregulated at least two-fold by P. aeruginosa exposure in an ATF-7-dependent manner and that were also bound by ATF-7 in any of our four ChIP-seq conditions. We chose to concentrate on genes upregulated in response to pathogen exposure (Fig 1C) in contrast to downregulated genes, which appear to reflect a reduction in general growth and metabolism as is reflected in our RNA-seq data (S1A Fig) [10]. Included among these putative ATF-7 targets were genes encoding antimicrobial effector molecules, such as CTLD-containing proteins and lysozymes (S3 Table). We determined whether RNAi-mediated knockdown of these genes resulted in enhanced susceptibility to killing by P. aeruginosa and observed that RNAi of 13 of 43 genes conferred enhanced sensitivity to killing by P. aeruginosa, without affecting survival on non-pathogenic E. coli (S3 Table, S4 Fig).
Our data suggest that ATF-7 is a direct regulator of immune effector genes that is regulated by PMK-1 p38 MAPK. We previously proposed a model in which PMK-1 phosphorylates ATF-7 in response to pathogen infection, switching the activity of ATF-7 from that of a transcriptional repressor to that of an activator, allowing the induction of immune response genes [7]. Our data here are consistent with this model, showing a strong dependence of pathogen-induced gene induction on PMK-1 and ATF-7, and a high degree of occupancy of regulatory regions of pathogen-induced genes by ATF-7 under basal and pathogen-induced conditions, with ATF-7 occupancy of pathogen-induced genes being strongly dependent on PMK-1. Moreover, our data reveal that PMK-1-ATF-7 signaling regulates over half of all pathogen-induced genes at the genome-wide level.
PMK-1 signaling has also been implicated in a number of non-infection contexts in C. elegans [2,16,17]. Interestingly, we observed that ATF-7 binds quite strongly to several key regulators of stress response pathways. We found that ATF-7 exhibits binding affinity to regulators of autophagy (lgg-1), the Unfolded Protein Response (xbp-1), and the oxidative stress response (skn-1), as well as several immunity regulators (hlh-30, zip-2, and interestingly, atf-7) (Fig 4). These observations suggest that initiation of other stress responses may be integrated with the immune response. For example, we have previously shown that immune response activation in developing larva is lethal without compensatory XBP-1 activity, establishing an essential role for XBP-1 during activation of innate immunity during infection of C. elegans [2]. Comparison of ATF-7 target genes identified by ChIP-seq with published target gene lists inferred from transcriptional profiling studies indeed suggests statistically significant overlap of ATF-7 regulation with SKN-1 (p<0.001, hypergeometric test with Bonferroni correction) and ZIP-2 (p<0.001, hypergeometric test with Bonferroni correction) (S6 Fig) [18]. However, caution should be taken when comparing these datasets, as they were collected by methods distinct from our intersection of ChIP-seq and RNA-seq results, and consequently additional experimental evidence is needed to corroborate these associations. We speculate that ATF-7 may function to activate anticipatory stress responses that can be activated in concert with innate immunity to promote host survival during microbial infection in a context-dependent manner. Our genomic and genetic findings in the simple, genetically tractable C. elegans host reveal a striking degree of global regulation of the organismal response to pathogenic bacteria through a single p38 MAPK-regulated transcriptional regulator. Our data support the idea that host defense, on a genome-wide and organism-wide level, is under the control of a limited number of stress-activated signaling pathways that regulate global regulators of gene transcription.
Strains used: N2, ZD386 (atf-7(qd22 qd130)), KU25 (pmk-1(km25)), ZD1807 (atf-7(qd328[atf-7::2xTY1::GFP])), ZD1976 (atf-7(qd328[atf-7::2xTY1::GFP]);pmk-1(km25)). C. elegans were maintained at 16°C on E. coli OP50 as described [19]. The atf-7(qd328) allele was generated by the CRISPR-Cas9 system as described [19,20] and verified by Sanger sequencing. GFP expression in ZD1807 (atf-7(qd328)) was verified by immunobloting, and pull-down was assessed by IP-IB. The atf-7(qd238) allele was confirmed to function as wild-type, as assayed by susceptibility to P. aeruginosa strain PA14 in a slow kill assay, and then crossed into the pmk-1(km25) mutant background.
Slow Kill Assay (SKA) plates were prepared as previously described [20]. P. aeruginosa strain PA14 or E. coli OP50 was grown overnight in Luria Broth (LB), seeded onto SKA media and then grown overnight at 37°C, followed by an additional day at room temperature as previously described [21]. Large populations of animals were synchronized by egg-preparation of gravid adult worms in bleach, followed by L1 arrest overnight in M9 buffer. L1 animals were dropped onto concentrated OP50 lawns seeded onto Nematode Growth Media (NGM) and raised to the L4 larval stage at 20°C (about 40 h). Upon reaching L4, worms were washed off growth plates with M9 and placed on SKA plates prepared as described above, seeded with either PA14 or OP50 and incubated at 25°C for four hours. At this time, worms were harvested by washing for downstream applications.
After three washes in M9 buffer, animal pellets were resuspended in an equal volume of PBS + complete ULTRA protease inhibitor tablets (Roche), flash frozen in liquid nitrogen, and stored at -80°C until chromatin immunoprecipitation (ChIP). ChIP was preformed as described [22,23] using Ab290, a ChIP-grade polyclonal GFP antibody (Abcam). Libraries were prepared using the SPRIworks Fragment Library System (Beckman Coulter) and single-end sequenced on an Illumina HiSeq2000 sequencer. Three biological replicates of at least 15,000 animals were prepared and sequenced for each condition, with the exception of only two replicates for atf-7(qd328) on PA14, as one of the samples failed to pass quality control.
ChIP-seq reads were aligned against the C. elegans WBPS9 assembly using bwa v. 0.7.12-r1039 [24] and the resulting bam files were sorted and indexed using samtools v. 1.3 [25]. Sorted bam files were pooled by strain and microbial treatment, and peaks were called using MACS2 (v. 2.1.1.20160309), as follows: callpeak on specific strain bam file (“-t” flag) against the N2_PA14 control sample bam file (“-c” flag) callpeak -c N2_PA14_control.sorted.bam -g ce—keep-dup all—call-summits—extsize 150 -p 1e-3—nomodel -B. Peak locations were intersected with regions +/-0.5kb around annotated TSS based on the WBPS9/WS258 annotation using bedtools intersect (v2.26.0) [26], and in cases of multiple peaks associated with a given TSS, peaks with maximal enrichment over N2 control were retained. For the purpose of motif identification, peaks were ranked by fold-enrichment over N2 control in descending order and the top 400 peaks were retained, regions +/- 200 bps around the summit were retrieved and sequences were obtained with bedtools getfasta. MEME-ChIP v. 4.12.0 [27] was used to call motifs using the following parameters: meme-chip -oc. -time 300 -order 1 -db db/JASPAR/JASPAR2018_CORE_nematodes_non-redundant.meme -meme-mod anr -meme-minw 5 -meme-maxw 30 -meme-nmotifs 8 -dreme-e 0.05 -centrimo-local -centrimo-score 5.0 -centrimo-ethresh 10.0. Presence of the top motifs under each peak called by macs2 was assessed using Mast v.5.0.1 [28] on the same +/- 200bp region around the summit of each peak. The number of peaks with one or more occurrences of the motif was tallied using a 200-peak window, and plotted across all peaks ranked either by log-fold enrichment over N2 or–log-transformed p-values. Inflection points in the motif density function were used to narrow down the number of peaks retained for downstream analyses.
After three washes in M9 buffer, TRIzol Reagent (Invitrogen) was added to worm pellets and flash frozen in liquid nitrogen. Following an additional round of freeze-thaw, RNA was isolated using the Direct-zol RNA MiniPrep kit (Zymo Research). Libraries were prepared using the Kapa mRNA Hyperprep kit and paired end reads were sequenced on the Illumina NextSeq500 sequencer. Three biological replicates of at least 1,000 animals were prepared and sequenced for each condition, with the exception of only two replicates for atf-7(qd22 qd130) on PA14, as one of the samples failed to pass quality control.
Reads were aligned against the C. elegans WBPS9 assembly/ WS258 annotation using STAR v. 2.5.3a [29] with the following flags: -runThreadN 16—runMode alignReads—outFilterType BySJout—outFilterMultimapNmax 20—alignSJoverhangMin 8—alignSJDBoverhangMin 1—outFilterMismatchNmax 999—alignIntronMin 10—alignIntronMax 1000000—alignMatesGapMax 1000000—outSAMtype BAM SortedByCoordinate—quantMode TranscriptomeSAM. with—genomeDir pointing to a low-memory footprint, 75nt-junction WBPS9/WS258 STAR suffix array. Gene expression was quantitated using RSEM v. 1.3.0 [30] with the following flags for all libraries: rsem-calculate-expression—calc-pme—alignments -p 8 against an annotation matching the STAR SA reference. Posterior mean estimates (pme) of counts and estimated “transcript-TPMs” were retrieved for genes and isoforms. Subsequently, counts of isoforms sharing a transcription start site (TSS) were summed, and differential-expression analysis was carried out using DESeq2 [31] in the R v3.4.0 statistical environment, building pairwise models of conditions to be compared (microbial exposures within each genotype). Sequencing library size factors were estimated for each library to account for differences in sequencing depth and complexity among libraries, as well as gene-specific count dispersion parameters (reflecting the relationship between the variance in a given gene’s counts and that gene’s mean expression across samples).
Differences in gene expression between conditions (expressed as log2-transformed fold-changes in expression levels) were estimated under a general linear model (GLM) framework fitted on the read counts. In this model, read counts of each gene in each sample were modeled under a negative binomial distribution, based on the fitted mean of the counts and aforementioned dispersion parameters. Differential expression significance was assessed using a Wald test on the fitted count data (all these steps were performed using the DESeq() function in DESeq2) [31]. P-values were adjusted for multiple-comparison testing using the Benjamini-Hochberg procedure [32].
Raw data presented in this manuscript have been deposited in NCBI’s Gene Expression Omnibus [33] and are accessible through GEO SuperSeries accession number GSE119294, which contains SubSeries GSE119292 (RNA-seq data, including count files) and SubSeries GSE119293 (ChIP-seq data, including wig files and peak calls).
Metagene analyses of gene expression and ATF-7 binding enrichment were generated by ngs.plot as described [34], using ChIP .bam files from each condition normalized to N2 control as input. Genes considered two-fold upregulated or downregulated are listed in the “N2_up” and “N2_down” tabs of S1 Table, respectively.
Correlations between ATF-7 binding and regulation of gene expression were interrogated using the gene set enrichment analysis (GSEA) framework [35]. Briefly, all transcription start sites (TSSs) associated with a protein-coding transcript were ranked based on differential expression results from DESeq2 (log2 fold-changes), which is a measure of the correlation between their expression and the host response to infectious agents. Biases in expression of ATF-7-bound TSSs were assessed using a walk down the list tallying a running-sum statistic, which increases each time a TSS is part of the list and decreases otherwise. The maximum of this metrics (i.e. where the distribution if furthest away from the background) is called the enrichment score (ES). Significance is estimated using random permutations of the TSSs to generate p-values gauging how often the observed ES can be seen in randomized gene sets, for each direction of the expression biases independently. Multiple-testing correction is addressed using a false-discovery rate calculation on permuted datasets.
Genes with adjusted p-values <0.05 were considered for Gene Ontology enrichment analysis using the DAVID online webtool, considering as a background the union of all genes with a non-zero baseMean value across any of the DE comparison, based on unique WormBase IDs.
PA14 plates were prepared as described as above. N2 animals were grown on NGM, supplemented with 25 ug/mL carbenicillin and 2mM isopropyl b-D-1 thiogalactopyranoside (IPTG), that was seeded with either the E. coli HT115 expressing plasmids targeting the gene of interest or the empty L4440 vector backbone for two generations prior to each experiment. Animal populations were synchronized by egg lay. At the L4 larval stage, approximately 30 worms were transferred to prepared SKA plates and incubated at 25°C. Animals were scored for killing twice daily until the majority of animals had died. Within each experiment, three plates were prepared and scored per RNAi treatment. All clones were obtained from the Ahringer [36] or Vidal [37] RNAi libraries and were verified by sequencing. For a list of all RNAi clones used, see S4 Table.
|
10.1371/journal.pgen.1007578 | The roles of SMYD4 in epigenetic regulation of cardiac development in zebrafish | SMYD4 belongs to a family of lysine methyltransferases. We analyzed the role of smyd4 in zebrafish development by generating a smyd4 mutant zebrafish line (smyd4L544Efs*1) using the CRISPR/Cas9 technology. The maternal and zygotic smyd4L544Efs*1 mutants demonstrated severe cardiac malformations, including defects in left-right patterning and looping and hypoplastic ventricles, suggesting that smyd4 was critical for heart development. Importantly, we identified two rare SMYD4 genetic variants in a 208-patient cohort with congenital heart defects. Both biochemical and functional analyses indicated that SMYD4(G345D) was pathogenic. Our data suggested that smyd4 functions as a histone methyltransferase and, by interacting with HDAC1, also serves as a potential modulator for histone acetylation. Transcriptome and bioinformatics analyses of smyd4L544Efs*1 and wild-type developing hearts suggested that smyd4 is a key epigenetic regulator involved in regulating endoplasmic reticulum-mediated protein processing and several important metabolic pathways in developing zebrafish hearts.
| SMYD4 belongs to a SET and MYND domain-containing lysine methyltransferase. In zebrafish, smyd4 is ubiquitously expressed in early embryos and becomes enriched in the developing heart at 48 hours post-fertilization (hpf). We generated a smyd4 mutant zebrafish line (smyd4L544Efs*1) using the CRISPR/Cas9 technology. The maternal and zygotic smyd4L544Efs*1 mutants demonstrated a strong defect in cardiomyocyte proliferation, which led to a severe cardiac malformation, including left-right looping defects and hypoplastic ventricles. More importantly, two rare genetic variants of SMYD4 were enriched in a 208-patient cohort with congenital heart defects. Both biochemical and functional analyses indicated that SMYD4(G345D) was highly pathogenic. Using mass spectrometric analysis, SMYD4 was shown to specifically interact with histone deacetylase 1 (HDAC1) via its MYND domain. Altered di- and tri-methylation of histone 3 lysine 4 (H3K4me2 and H3K4me3) and acetylation of histone 3 in smyd4L544Efs*1 mutants suggested that smyd4 plays an important role in epigenetic regulation. Transcriptome and pathway analyses demonstrated that the expression levels of 3,856 genes were significantly altered, which included cardiac contractile genes, key signaling pathways in cardiac development, the endoplasmic reticulum-mediated protein processing pathway, and several important metabolic pathways. Taken together, our data suggests that smyd4 is a key epigenetic regulator of cardiac development.
| Protein post-translational modifications (PTMs) are critical for the biological function of proteins. Histone modification is a common epigenetic mechanism that plays essential roles in the regulation of chromatin structure and gene expression. Different types of histone modifications, which are mediated by a series of specific enzymes, can either enhance or inhibit transcription to regulate specific cellular functions or signaling pathways. SET and MYND domain-containing proteins (SMYDs) belong to a unique family of histone lysine methyltransferases. This family is composed of five members, including SMYD1, SMYD2, SMYD3, SMDY4, and SMYD5. These proteins share a Su(var)3-9, an Enhancer-of-zeste and Trithorax (SET) domain with lysine-specific methyltransferase activity, a Myeloid, Nervy, and DEAF-1 (MYND) domain, and a tetratricopeptide repeat (TPR) domain, which are involved in protein-protein interactions [1–3]. Several biochemical studies and functional analyses showed that SMYDs 1–3 exhibit methyltransferase activities for both histone and non-histone proteins [3–6]. SMYD members are widely present in multiple cell types, including those of skeletal and cardiac muscles [7–9]. Genetic ablation of Smyd1 in mice led to defects in right ventricular development [8,10]. The knockdown of smyd1, smyd2, and smyd3 in zebrafish using morpholino technology also led to cardiac malformation and defects in skeletal and cardiac myofibrillogenesis [11–13]. Zebrafish smyd5 was recently reported to play important roles in hematopoiesis [14]. Our knowledge of the biological function of SMYD4 remains limited. A previous study found that dSMYD4 was crucial for muscle development in Drosophila [15]. However, the role of SMYD4 in vertebrate development and its underlying molecular mechanism have not yet been analyzed thoroughly.
Congenital heart diseases (CHDs) are the most common birth defects, with an annual incidence of approximately 1% of newborns worldwide [16]. Heart development involves complex genetic and epigenetic regulation [17–19]. Cardiac-specific ablation of HDAC1 and HDAC2 in the murine heart led to aberrant gene expression, contributing to defects in cardiac morphogenesis and contractility [20], suggesting the critical roles of histone modification and epigenetic regulation in cardiac development. The significant enrichment of mutations in genes encoding chromatin modifiers in patients with CHDs provided further evidence to support this notion [21,22].
In this study, we analyzed smyd4 expression in zebrafish embryos and generated a smyd4 loss-of-function mutation (smyd4L544Efs*1) using the CRISPR/Cas9 genome editing technology. We demonstrated that smyd4 is critical for cardiac malformation in zebrafish development. Two rare missense SMYD4 variants were identified in patients with CHDs. Both in vitro biochemical assays and in vivo functional analyses strongly suggested that the variant SMYD4(G345D) was pathogenic. Our data suggested that smyd4 functions as a histone methyltransferase and, by interacting with HDAC1, also serves as a potential modulator for histone acetylation. Transcriptome and bioinformatics analyses of the differentially expressed genes in developing hearts isolated from maternal and zygotic smyd4L544Efs*1 (MZsmyd4L544Efs*1) mutant and normal control embryos showed a significant enrichment of the key genes involved in cardiac development and contractile function, and the endoplasmic reticulum-mediated protein processing pathway and several important metabolic pathways. Taken together, our results suggest that smyd4, in association with hdac1, is an important epigenetic regulator of these critical pathways during zebrafish cardiac development.
To determine the smyd4 mRNA levels in early to late zebrafish embryos, we performed a qRT-PCR analysis on embryos at the single-cell (0.2 hours post-fertilization (hpf)), 64-cell (2 hpf), blastula (4 hpf), mid-gastrula (8 hpf), late-gastrula (10 hpf), early primordium (24 hpf), long pectoral (48 hpf), and protruding mouth stage (72 hpf) stages (Fig 1A). We found that smyd4 transcripts were highly present in the single-cell stage, suggesting that a large amount of smyd4 was present as maternal transcripts. Zygotic smyd4 transcription peaked at the blastula stage and was followed by downregulation in the gastrula stage, indicating the role of this transcript in zebrafish early development. The overall expression levels of smyd4 were significantly further downregulated at 48 hpf and followed by a quick reactivation of expression at 72 hpf. We then used whole-mount in situ hybridization to determine the spatio-temporal expression pattern of smyd4 in developing zebrafish embryos. As shown in Fig 1B–1E, smyd4 was expressed ubiquitously in mid- and late-gastrula embryos; the highest level of expression was found at the polster in late-gastrula embryos (Fig 1D and 1E). By 24 hpf, smyd4 transcripts were found to be enriched in the developing heart and blood vessel (Fig 1F and 1G) and became more restricted to the cardiovascular system at 48 hpf (Fig 1H–1M). Interestingly, the developing ventricle had a significantly higher level of smyd4 expression than the atrium (Fig 1J and 1K), suggesting its role in cardiac development. The reactivation of smyd4 expression at 72 hpf appeared not to be cardiac specific (Fig 2F).
To determine the biological function of smyd4, we generated smyd4-deficient zebrafish using the CRISPR/Cas9 technology. Three sgRNAs were designed to target different exons of smyd4 that contribute to the ZMYND, C-terminal TPR2, and SET domains. We were able to generate an 11-nt insertion in exon 6 that was targeted by sgRNA3 (Fig 2A). As shown in Fig 2B, the sgRNA3 targeted sequence is well conserved from zebrafish to humans. Sequencing analysis confirmed this 11-nt insertion, which led to a frameshift at amino acid 545 (smyd4L544Efs*1) (Fig 2C and 2D and S1 Fig). This frameshift mutation yielded a premature stop codon (TAA), as indicated by the asterisk in Fig 2C, giving rise to a truncated mutant protein that lacks the entire C-terminus containing the TPR2 domain (Fig 2D). qRT-PCR and in situ hybridization analyses demonstrated that the smyd4 mRNA transcript was completely abolished in MZsmyd4L544Efs*1 embryos (Fig 2E and 2F). The lack of the smyd4 mRNA transcript was likely due to the well-known effect of nonsense mutation-mediated degradation (NMD) [23], a process common to frameshift mutations. Thus, we concluded that smyd4L544Efs*1 was a loss-of-function mutation. To better analyze the cardiac defects in MZsmyd4L544Efs*1 mutants, we crossed the smyd4L544Efs*1 line with the cardiomyocyte-specific transgenic reporter line Tg(cmcl2:GFP). The entire work presented here was performed using this bi-genic line (smyd4L544Efs*1;cmcl2:GFP) with Tg(cmcl2:GFP) as a normal control, unless otherwise indicated in the text.
Using the heterozygous breeding scheme, we were able to create heterozygous and homozygous smyd4L544Efs*1 mutant and wild-type embryos. First, we analyzed the survival of smyd4L544Efs*1 homozygous mutants and found no early lethality in these mutants, as the total amount of surviving smyd4L544Efs*1 homozygous mutants reached 25% among all three genotypes, which matched the normal Mendelian ratio. These homozygous mutants appeared normal in terms of growth and reproductivity. However, when examining these mutants, we observed a slight increase in embryos with abnormal cardiac situs ambiguus in smyd4L544Efs*1 homozygous (6 out of 14, 43%) and heterozygous (17 out of 48, 35%) mutants compared to the number of wild-type embryos (2 out of 10, 20%) (S2 Fig). As situs ambiguus is highly relevant for defects in early patterning, this finding prompted us to speculate that the reduced but remaining maternal smyd4 transcripts in heterozygous females might contribute to the reduced severity of the resulting phenotype. Therefore, we established a homozygous breeding scheme to eliminate the effect of maternal smyd4 transcripts.
The MZsmyd4L544Efs*1 embryos from the homozygous breeding scheme displayed severe pericardial edema (Fig 3A and 3B) and/or congested blood flow in the ventral veins (Fig 3A and 3B), suggesting prominent cardiac defects or dysfunctions associated with the MZsmyd4L544Efs*1 embryos. The cardiac defects were analyzed carefully. Major defects included a significantly smaller ventricular size and anomalies of cardiac left-right asymmetric patterning or looping in approximately 60% of MZsmyd4L544Efs*1 embryos at 72 hpf (Fig 3C and 3G), such as situs ambiguus (straight heart tube) and situs inversus (D-loop heart tube) (Fig 3C). The 3D-reconstruction of confocal images of 96 hpf hearts further confirmed this looping defect (Fig 3F and S1 and S2 Movies), as well as hypoplastic ventricles with less trabeculated myocardial structures in the MZsmyd4L544Efs*1 hearts (Fig 3F and 3G). The number of cardiomyocytes on the largest section plane of the MZsmyd4L544Efs*1 mutant ventricle was found significantly reduced when compared to the comparable section plane of wild-type control hearts (S3 Fig and Fig 3H), suggesting a great reduction of total number of cardiomyocytes in MZsmyd4L544Efs*1 hearts. To determine whether reduced cell number was due to the decreased cellular proliferative activity, we used p-H3 and PCNA immune reactivities to respective antibodies as the indicators for the level of cell proliferative activities in mutant developing hearts. As shown in Fig 4A and S4 Fig, MZsmyd4L544Efs*1 mutant cardiomyocytes had a dramatically reduced level in cellular proliferation when compared to wild-type controls. In addition, we also evaluated the levels of apoptosis in mutant hearts and we found no evidence of increased level of apoptotic cell in MZsmyd4L544Efs*1 mutant hearts (S5 Fig).
As adults, these MZsmyd4L544Efs*1 mutants had abnormal gross cardiac morphologies and histologies, typically with less trabecular myocardia and sometimes with a dramatically increased thickness of the ventricular compact wall (Fig 4B and 4C). Despite these cardiac defects, we observed no apparent defects in skeletal muscle development in MZsmyd4L544Efs*1 mutants examined at 48 and 72 hpf (S6 Fig), suggesting that smyd4 is lesser important than smyd1 in vertebrate muscle development [11].
As shown in Fig 5A, SMYD4 was localized to both the nucleus and cytoplasm. To determine the biochemical function of SMYD4, flag-tagged SMYD4 (SMYD4flag) was first overexpressed in the HL-1 mouse cardiomyocyte cell line, followed by immunoprecipitation and mass spectrometric analysis to identify its interacting proteins. HDAC1 was identified as one of the major proteins to interact with SMYD4 (Fig 5B). This finding was further confirmed by Co-IP/western blotting analysis using a HEK293T cell line in which both SMYD4flag and HA-tagged HDAC1 (HDAC1HA) were co-overexpressed (Fig 5C). There are four functional domains in SMYD4. Two TPR domains are located at the N- and C- termini, an MYND domain can mediate interactions with partner proteins, and a SET domain functions as a methyltransferase. To determine the functional domain that was responsible for the SMYD4/HDAC1 interaction, as shown in Fig 5D and 5E, we generated several mutations in SMYD4 with different combinations of deletions in the SMYD4 protein and co-expressed these mutants with HDAC1 in HEK293T cells. We were able to use Co-IP/western blotting assays to demonstrate that the MYND domain was responsible for the interaction between SMYD4 and HDAC1.
To confirm the biochemical activities of SMYD4 as a methyltransferase and as a functional partner of HDAC1 in histone modification, we analyzed the changes in both histone methylation and acetylation modifications in the MZsmyd4L544Efs*1 embryos. As shown in Fig 5F and 5G, di- (me2) and tri-methylation (me3) at the lysine 4 site (K4) of histone 3 (H3) were significantly reduced, and mono-methylation of lysine 4 (H3K4me) was increased, while other lysine residues (i.e., K9 and K27) were not affected, suggesting that SMYD4 was specifically involved in H3K4 methylation. Interestingly, the acetylation of lysines 4, 9, 14, and 27 was dramatically reduced in the MZsmyd4L544Efs*1 mutants examined at 48 hpf (Fig 5F and 5G). This finding suggested that SMYD4 is a critical part of the HDAC1 functional complex and may function as an important negative regulator of HDAC1-mediated histone 3 modification and epigenetic regulation. Taken together, these data indicate that SMYD4 is a functional methyltransferase specific for H3K4 methylation and a functional partner of HDAC1 for regulating H3 acetylation.
We used Target Exome Sequencing (TES) and screened a cohort of 208 patients with CHDs for potential genetic variants of SMYD4. The patient information is summarized in S1 Table and S2 Table. Two rare missense variants (c.1034G>A, p. G345D and c.1736G>A, p. R579Q) were identified and confirmed by Sanger sequencing in two individual patients (Fig 6A and 6B). The variants were not recorded in the 1000G database or in our internal whole-exome sequencing database generated from more than 3,500 patients without CHDs. The frequencies of G345D and R579Q in the EXAC database are 1/121,404 and 3/120,896, respectively. G345D was identified in a patient who was diagnosed with DCRV/VSD, and R579Q was identified in a patient with TOF (S3 Table). Both sites were evolutionally conserved from zebrafish to humans (Fig 6C). Both variants (G345D and R579Q) were predicted to be highly pathogenic and harmful by SIFT, PolyPhen2 and MutationTaster. Based on 3D-computational structure prediction using the SWISS MODEL, as shown in Fig 6D and 6E, both mutations led to significant changes in protein structure compared with that of the wild-type SMYD4. To confirm the pathogenicity of the variant SMYD4(G345D), Co-IP/western blotting assays were performed, and we found that the biochemical interaction between SMYD4(G345D) and HDAC1 was greatly attenuated compared to the interaction between SMYD4 and HDAC1 in HL-1 cells (Fig 6F and 6G), consistent with the predicted alteration of protein structure and function in SMYD4(G345D).
To confirm SMYD4(G345D) as a pathogenic mutation, we performed gain-of-function transgenic overexpression experiments by injecting wild-type smyd4 and smyd4(G295D) mRNA into normal Tg(cmcl2:GFP) embryos (Fig 7A and 7C). Based on the amino acid sequences of zebrafish smyd4 and human SMYD4, we generated smyd4(G295D) mutant cDNA, which was equivalent to human SMYD4(G345D) (Fig 6C and S7 Fig). We found that the smyd4(G295D) mRNA caused significantly more embryos with severe heart defects (e.g., D-loop and tubular hearts) than wild-type smyd4 mRNA (Fig 7A and 7C), suggesting that SMYD4(G345D) was harmful to cardiac development.
To further confirm the effects of this mutation, we performed rescue experiments by analyzing and comparing the ability of mutant SMYD4(G345D) and wild-type SMYD4 to rescue the abnormal cardiac phenotypes of MZsmyd4L544Efs*1 mutants. We injected smyd4 wild-type and smyd4(G295D) mutant mRNAs into MZsmyd4L544Efs*1 single-cell embryos harvested from the homozygous breeding scheme described above. As shown in Fig 7B and 7C, wild-type smyd4 mRNA significantly reduced the number of embryos with malformed hearts compared to the number in the MZsmyd4L544Efs*1 mutant group, which indicates a partial but significant rescue phenotype for the wild-type smyd4 mRNA. In contrast, the smyd4(G295D) mRNA not only failed to rescue MZsmyd4L544Efs*1 but also significantly increased the number of embryos with malformed hearts and severe cardiac defects, such as tubular hearts (Fig 7B and 7C), further confirming that the smyd4(G295D) mutation is pathogenic. Taken together, our data implied that human SMYD4(G345D) was deleterious for heart development and a CHD-causing genetic variant.
Given the severe cardiac defects and abnormal histone modifications in the MZsmyd4L544Efs*1 embryos, we anticipated a large number of gene alterations in MZsmyd4L544Efs*1 mutant hearts. We performed RNA-seq analysis, comparing normal and MZsmyd4L544Efs*1 hearts harvested at 72 hpf. As shown in Fig 8, a total of 3,856 differentially expressed (DE) genes were identified in MZsmyd4L544Efs*1 hearts. Among those genes, 2,648 genes were upregulated, and 1,208 genes were downregulated (Fig 8A). Not surprisingly, some important genes that were highly relevant to cardiac development were altered, which included the genes involved in cardiac muscle contraction (upregulated: 10; downregulated: 22) and key cardiac signaling pathways, such as the canonical Wnt signaling pathway (upregulated: 36, downregulated: 10) and the Hedgehog signaling pathway (upregulated: 15, downregulated: 3) (Fig 8B). To investigate whether this altered transcriptional profile was associated with specific pathways or biological processes, we performed KEGG pathway analysis, which indicated that the upregulated genes were enriched in the endoplasmic reticulum-mediated protein processing pathway in the Biological Process domain (Fig 8C) and the downregulated genes were enriched in several metabolic pathways, including carbon metabolism and the glycolysis/gluconeogenesis pathway (Fig 8D). Gene ontology (GO) analysis of the upregulated genes also revealed major terms in cellular metabolic processes, in which 975 genes were involved. A GO term analysis of the downregulated genes revealed the enrichment of a large number of genes in the organonitrogen compound metabolic process (185 genes), the ATP metabolic process (49 genes) and the glycosyl compound metabolic process (55 genes) (Fig 8E). This finding provided an important hint that smyd4/hdac1-mediated epigenetic regulation likely occurred via the control of endoplasmic reticulum-mediated protein processing and several key metabolic pathways in the heart during zebrafish development.
SMYDs are a family of unique lysine-histone methyltransferases that contain the well-conserved SET and MYND domains, as well as two TPR domains. The biological functions of SMYDs are largely unknown. By generating smyd4 mutant zebrafish using the CRISPR/Cas9 technology, we have shown that smyd4 is indispensable for cardiac development. This finding is consistent with the observation that ubiquitously expressed smyd4 in early zebrafish embryos becomes more enriched to zebrafish developing hearts at 48 hpf during embryogenesis, despite the fact that gross expression levels are reduced dramatically at this stage, which is followed by a quick up-regulation of smyd4 expression in embryos at 72 hpf (Fig 1A). These are critical stages, in which the newly formed heart switches from the cardiac morphogenic pathway to cardiac maturation pathways to eventually form a normal functional heart.
There are several major cardiac defects in the MZsmyd4L544Efs*1 mutants, including the defects in left-right patterning and looping, and the hypoplastic ventricular walls in the developing hearts. Furthermore, the cell proliferative activities in MZsmyd4L544Efs*1 developing ventricles are significantly decreased, which likely leads to the hypoplastic ventricle. In adults, the thickening of the ventricular compact wall that is seen in some mutant hearts is likely a maladaptation of compromised cardiac function due to cardiac developmental defects.
Considering smyd4 shares similar protein structure and enzymatic activities with other SMYD family members, genetic and functional redundancy may occur in MZsmyd4L544Efs*1 mutants. It has been shown that smyd1b mutants display a similar pericardial edema and congestive blood circulation [24]. The morpholino knock-down of another SMYD family member smyd3 in zebrafish embryos also produces cardiac looping defects [13]. Both of these prior studies suggest potential genetic interactions and/or redundancies among these family members in zebrafish heart development. Some variations in the cardiac defects as well as certain degrees of varying severity of cardiac defects among the mutant embryos further support this notion. However, it is also clear that each member of SMYD family has its unique function. For example, myofibril disorganization in skeletal muscle are only seen in smyd1b mutants [24]. We have not observed any such a defect in MZsmyd4L544Efs*1 mutants, suggesting that the function of smyd4 is more relevant to cardiac development in zebrafish.
Our observations strongly suggest that smyd4 plays an important function in cardiac development. This conclusion is also supported by our human genetic study, which identified two rare SMYD4 variants in a 208-patient cohort with CHDs. The data obtained after the overexpression of mutant smyd4(G295D) and from rescue experiments in which smyd4(G295D) mRNA was injected into MZsmyd4L544Efs*1 mutants strongly indicates that the human variant SMYD4(G345D) is pathogenic, further suggesting SMYD4 as a genetic contributor to CHDs. This finding indicates the importance of including SMYD4 in CHD genetic screening panels in the future.
One of our key findings is that SMYD4 interacts with the major histone modification enzyme and epigenetic regulator HDAC1 via the well-conserved MYND domain. Our finding is consistent with the previous finding that dSmyd4 can interact with dHDAC1 in Drosophila muscle development [15]. The MYND domain is known for its role in protein-protein interactions and was previously shown to recruit the HDAC complex to regulate gene expression [25]. The mutation in SMYD4(G345D) is located at the edge of the MYND domain and between the MYND and SET domains. Our biochemical data show that SMYD4(G345D) has a dramatically reduced ability to interact with HDAC1. Although the TPR domains are still poorly understood, previous studies suggest that the C-terminal TPR2 domain is vital for methyltransferase activity and protein-protein interactions [3,26]. The TPR2 domain of SMYD2 is indispensable for its interaction with HSP90, which proved to be critical for titin filament organization [12,27]. Currently, we are in the process of generating smyd4-tpr2del mutant zebrafish similarly using the CRISPR/Cas9 technology to further investigate the role of TPR2 in smyd4 biological function.
SMYDs 1 and 3 catalyze mono-, di-, and trimethylation of H3K4 [11,28]. Similarly, SMYD2 can mono-methylate H3K4 (H3K4me) and di-methylate H3K26 (H3K26me2) and p53K37me [6,29]. In MZsmyd4L544Efs*1 mutants, H3K4me2 and H3K4me3 are reduced, suggesting that smyd4 is a specific methyltransferase for H3K4 methylation. Notably, H3K4ac, K9ac, K14ac, and K27ac were all abolished in the MZsmyd4L544Efs*1 mutants. This finding implies that the deficiency of smyd4 impacts the function of hdac1 (the gene homologous to both HDAC1 and HDAC2 in mammals). As previously demonstrated, cardiac-specific deletion of the mouse Hdac1 and Hdac2 genes evoked a strong heart failure phenotype [20], which is consistent with our finding. We are currently using biochemical approaches to determine whether smyd4 serves as a simple docking protein to provide chaperone functions for hdac1 or functions as a key modulating molecule for hdac1. Nevertheless, our data suggest that smyd4 plays a critical role in the epigenetic regulation of gene expression via its dual activities as a methyltransferase and negative regulator of hdac1.
RNA-seq analysis comparing wild-type and MZsmyd4L544Efs*1 mutant hearts demonstrates that the expression of over 3,000 genes is altered, which may reflect the potential function of smyd4’s broad epigenetic regulation of its target genes. In addition to genes related to cardiac muscle contraction and cardiac signaling pathways that are highly relevant to cardiogenesis (e.g., the canonical Wnt and Hedgehog signaling pathways) (Fig 8), our KEGG pathway and GO annotation analyses of altered genes revealed an overwhelming enrichment in several cellular metabolic pathways, including the endoplasmic reticulum-mediated protein processing pathway. This finding is very different from the publicly available RNA-seq database for zebrafish heart developmental defects [30–32]. This finding suggests that smyd4 has a unique and specific biological function in regulating cellular metabolism. However, as it is technically difficult to perform ChIP-seq on zebrafish embryonic hearts, we cannot currently determine which specific components of the pathways are primarily affected and which are affected secondarily. Our future study will switch to a mammalian system to determine the detailed molecular mechanism by which SMYD4 modulates cellular metabolism or signaling pathways via its important role in epigenetic regulation. We will re-evaluate whether these altered metabolic pathways play critical roles in cardiac development and the cardiac defects seen in smyd4 mutant embryos.
This study is the first characterization of SMYD4 in vertebrate development and physiological function. Taken together, our results demonstrate the critical role of smyd4 in embryonic development and heart formation. Our data suggest that smyd4 functions as a histone methyltransferase and, by interacting with HDAC1, also serves as a potential modulator for histone acetylation. In addition, our work has also provided genetic and functional evidences that rare SMYD4 variants likely contribute to CHDs.
All genetic studies were approved by the Ethics Committee of the Children’s Hospital of Fudan University, China. The approval number is: [2015]92). All patients provided written informed consent in accordance to the Declaration of Helsinki. The Research Ethics Committee of the Children’s Hospital of Fudan University, China, approved and monitored all zebrafish procedures following the guidelines and recommendations outlined by the Guide for the Care and Use of Laboratory Animals. The approval number is: [2015]92). For all experiments, wild-type zebrafish embryos of the Tu and transgenic Tg (cmcl2:GFP) (cardiac myosin light chain 2:GFP reporter) strains were used.
The expression of smyd4 was detected in zebrafish embryos from the 10 to 72 hpf stages using a smyd4-specific antisense probe. The template for the smyd4 probe was amplified from the cDNA of zebrafish at 24 hpf. The 508-bp fragment, which was obtained using specific primers (smyd4-probe-F: GAAGTGTGTGAAATGTGGAAAGCCTCTT and smyd4-probe-R: TTCACTCAGTTCCTGCAGTTCTTCACAG), was cloned into the pEasy-T vector (Promega, USA). After linearization of the plasmids, the antisense and sense probes were transcribed and labelled with digoxigenin in vitro. RNA in situ hybridization was performed as described previously [33]. Briefly, zebrafish embryos at different stages were collected and fixed in 4% paraformaldehyde at 4°C overnight. Embryos older than 24 hpf were digested by proteinase K at room temperature. Then, the embryos were pro-hybridized at 65°C for 4 hours and subsequently incubated with the antisense or sense probes overnight. An anti-digoxigenin antibody (Roche, USA) was used to bind the probes overnight at 4°C. Finally, the embryos were stained with NBT/BCIP (Vector, USA) and photographed in methylcellulose using a Leica M205C microscope.
Embryos of the wild-type Tu zebrafish strain were collected at 0.2, 2, 4, 8, 10, 24, 48 and 72 hpf. The total RNA was extracted using the TRIzol reagent (Invitrogen, USA) and converted to cDNA using the PrimeScript RT Reagent Kit (Takara Bio, Japan). The real-time qPCR reactions were performed with SYBR Premix Ex Taq (Takara Bio, Japan) using the Roche 480 plus system (Roche, USA). The real-time primers for the zebrafish are summarized in S4 Table.
smyd4 target sites were designed using the website http://zifit.partners.org/ZiFiT/CSquare9Nuclease.aspx. The provided sites were then screened in Ensemble. Three sites that specifically recognize the sequence of smyd4 in the zebrafish genome were chosen for the interruption of smyd4. sgRNA1 (GGAGTAATGAAGCACTGCTG), sgRNA2 (GGAGCTGATCTGCTGGCCAT), and sgRNA3 (GGAGCGTCAGCGCCTCCTGC) targeted exons 2, 4 and 6 of smyd4, respectively. We cloned these sites into the gRNA plasmid p-T7-gRNA, which was provided by Professor Li Qiang. gRNAs were transcribed in vitro using the MAXIscript T7 kit (Ambion, USA). Cas9 mRNA was transcribed from the pSP6-2Snls-spCas9 plasmid using the SP6 mMESSAGE mMACHINE Kit (Ambion, USA), and poly A tails were added using the poly A Tailing Kit (Ambion, USA). All gRNAs and the Cas9 mRNA were purified and dissolved in nuclease-free water before injection using the mirVana miRNA Isolation Kit (Ambion, USA) and the RNA Purification Kit (TIANGEN, China). gRNA and Cas9 mRNA were co-injected into the embryos at the single-cell stage. Twenty injected embryos were used to identify the efficiency, and the remaining embryos were raised to adulthood to obtain the mosaic founders. These mosaic fish were crossed with wild-type zebrafish to produce heterozygotes, which were genotyped using Sanger sequencing methods. To analyze the cardiac defects in MZsmyd4L544Efs*1 mutants, we crossed the smyd4L544Efs*1 line to the cardiomyocyte-specific transgenic reporter line Tg (cmcl2:GFP). All primers and target sites of smyd4 are summarized in S4 Table.
Embryos were collected at 48, 72, and 96 hpf for phenotype analysis. The embryos were fixed in methylcellulose and imaged using Leica M205C and Leica SP8 microscopes (Leica, Germany). To determine the number of cardiomyocytes in the developing ventricle, the embryonic hearts at 96 hpf were carefully collected and scanned using Leica SP8 confocal microscope. The z-step was set at 1μm. The images with largest section of ventricles were chosen for the analysis. DAPI and EGFP double positive cells were scored for cardiomyocyte. Adult fish at 6 months of age were photographed to record the body size and the developmental states of different organs, including the head, eyes, fins, and tails. These fish were dissected after anesthesia. The hearts of adult zebrafish were fixed and photographed in 4% paraformaldehyde. Serial sectioning with H&E staining was performed for these heart samples.
Embryos were collected at 48 hpf, fixed in 4% paraformaldehyde at 4°C overnight. For the proliferation assay, the embryos were digested using Collagenase, Type II (Life technologies, USA) at room temperature, and blocking was performed for one hour at room temperature, and followed by incubating with the primary antibody against p-H3(S10) (Abacam, USA) or PCNA (Genetex, USA) at 4°C overnight. Second antibody were from the series of Alexa Fluor (Life technologies, USA). For the apoptosis assay, In Situ Cell Death Detection kit, TMR red (Roche, USA) was used and all procedures were performed as the instruction manual described. Finally, embryo hearts were collected under a Leica M205C stereomicroscope and were imaged using a Leica SP8 microscope.
A plasmid containing wild-type human SMYD4(BC035077) was obtained from Abmgood (Abmgood, USA). Then, Flag-tagged wild-type SMYD4 and HA-tagged wild-type HDAC1 were cloned into the expression plasmid. The mutation (G345D) identified in patients was obtained using the KOD-Plus Mutagenesis Kit (Toyobo, Japan). All plasmids were confirmed via Sanger sequencing. The anti-SMYD4 (Proteintech, USA), anti-HDAC1 (Proteintech, USA), anti-Flag (Abmart, China), and anti-HA (Abmart, China) antibodies were used. Wild-type SMYD4 was overexpressed in HL-1 cells. After 48-h transfections, cell lysates were obtained in RIPA containing 1 mM PMSF and complete protease inhibitors (Roche, USA). Immunoprecipitation was performed using an anti-Flag affinity gel (Biotool, USA). SDS-PAGE was performed to resolve the eluates. After sliver staining, the proteins underwent mass spectrometry analysis. Protein-protein interactions were verified in HL-1 cells after transient overexpression of SMYD4. Co-immunoprecipitation was performed to confirm the interaction between SMYD4 and HDAC1 in HEK293T cells using anti-tag antibodies.
HEK293T cells were transiently co-transfected with pCDH-Flag-SMYD4 deletion mutants and pcDNA3-HA-HDAC1. Plasmids of SMYD4 were co-transfected into HEK293T cells, and cell extracts were prepared as described above. Immunoprecipitations were performed with anti-HA affinity beads (Biotool, USA). The beads were washed five times, and bound proteins were eluted in SDS-PAGE loading buffer and analyzed via western blotting.
Briefly, the immunofluorescence process is described as follows. The cells were fixed in 4% paraformaldehyde and then underwent cell permeation and blocking. The antibody used for immunofluorescence was anti-SMYD4 (Proteintech, USA). Primary antibodies were incubated overnight at 4°C. Secondary antibodies were from the Alexa Fluorescence Series (Life technologies, USA). Finally, the cells were imaged in Diamond anti-fade agent (Life technologies, USA) using a Leica SP8 microscope.
Tg(cmcl2:GFP) and MZsmyd4L544Efs*1 embryos were collected at 48 hpf. Protein was obtained in RIPA containing 1 mM PMSF and complete protease inhibitors (Roche, USA) after sonication. The histone modification antibodies used in this study include anti-H3 (CST, USA), anti-H3K4ac (Active Motif, USA), anti-H3K9ac (Active Motif, USA), anti-H3K14ac (Active Motif, USA), anti-H3K27ac (Active Motif, USA), anti-H3K4me1 (Active Motif, USA), anti-H3K4me2 (Active Motif, USA), anti-H3K4me3 (Abcam, USA), anti-H3K9me3 (Abcam, USA), and anti-H3K27me3 (Millipore, USA).
This study was approved by the research ethics committee of the Children’s Hospital of Fudan University in Shanghai, China (number: [2015]92). The diagnosis of CHD patients was based on echocardiography at the Children’s Hospital of Fudan University in Shanghai, China. All patients involved in this research had not been diagnosed with extra cardiac anomalies and did not have common chromosomal anomalies, such as the 22q11 microdeletion. Human cardiac tissue samples from CHD patients were obtained from the Biobank of the Children’s Hospital of Fudan University in Shanghai, China. Cardiac tissues were removed from the blocked right ventricular outflow tract during surgery. All tissue samples were maintained in RNAlater RNA Stabilization Solution (Thermo Scientific, USA) after surgery and stored at -80°C before use. RNA was extracted from the tissue samples using the Trizol reagent (Invitrogen, USA) and immediately underwent reverse transcription using the PrimeScript RT Reagent Kit (Takara, Japan). All detailed patient information is summarized in S2 Table.
The peripheral venous blood samples from 113 patients were prepared for DNA extraction using the Blood Extraction Kit (QIAGEN, Germany). The cardiac tissue samples from 95 patients were prepared for cDNA extraction. All of the regions covered by TES and all primers for smyd4 exon sequencing in cDNA are listed in S5 Table. Variant analysis was performed using the Mutation Surveyor software (Softgenetics, USA). All variants were screened in public databases, including the 1000 Genome database, the dbSNP database, and the ExAC database, and an internal database in the molecular diagnosis laboratory at the Children’s Hospital of Fudan University. A risk analysis of SNVs was performed using SIFT, Polyphen2, and Mutation Taster to predict the possible effects on protein function. A 3D structure analysis of the wild-type and mutant proteins was performed on the SWISS MODEL website (https://www.swissmodel.expasy.org/).
Based on its pathogenic prediction, the G345D SMYD4 variant was selected for mutation analysis. Homology analysis showed that the mutant G295D in smyd4 was equivalent to the mutant G345D in SMYD4. The template for smyd4 wild-type mRNA transcription was amplified from zebrafish cDNA at 24 hpf. The template of the mutant (G295D) in smyd4 transcription was obtained using the KOD-Plus Mutagenesis Kit (Toyobo, Japan). The wild-type and mutant mRNAs of smyd4 were transcribed using the mMessage mMachine T7 Ultra Kit (Ambion, USA), and poly A tails were added using the poly A Tailing Kit (Ambion, USA), according to the instruction manual. The mRNAs were resolved in nuclease-free water and finally quantified to 150 ng/μl for microinjection. Embryos of Tg(cmcl2:GFP) and MZsmyd4 L544Efs*1 were collected. A total of 3 nl of mRNA was microinjected into embryos at the single-cell stage. Fifty embryos from each group with cardiac GFP at 48 hpf were chosen for phenotype analysis. The embryos were fixed in 3% methylcellulose after anesthetization and then observed for cardiac morphology using a Leica M205C stereomicroscope.
Heart tissue was collected from 50 MZsmyd4L544Efs*1 or Tg(cmcl2:GFP) embryos. Cardiac-specific GFP helped us successfully obtain embryo heart tissue. Embryo hearts were collected at 72 hpf after anesthesia using a Leica M205C stereomicroscope. In-depth RNA sequencing was performed by the Novogene Experimental Department in China. The raw sequencing image data were examined via the Illumina analysis pipeline and aligned with the unmasked zebrafish reference genome. Differential expression analysis of the two groups was performed using the DESeq2 R package (1.10.1). The resulting P-values were adjusted using the Benjamini and Hochberg’s approach for controlling the false discovery rate. Genes with an adjusted P-value < 0.005 and |log2 (Fold change)|>1 found by DESeq2 were assigned as differentially expressed.
The Student’s t-test was used for all statistical analyses. A p-value of < 0.05 (2-sided) was regarded as statistically significant. All experiments were repeated three times. All data were analyzed with GraphPad Prism (version 5.0).
|
10.1371/journal.ppat.1006010 | Pyrimidine Salvage Enzymes Are Essential for De Novo Biosynthesis of Deoxypyrimidine Nucleotides in Trypanosoma brucei | The human pathogenic parasite Trypanosoma brucei possess both de novo and salvage routes for the biosynthesis of pyrimidine nucleotides. Consequently, they do not require salvageable pyrimidines for growth. Thymidine kinase (TK) catalyzes the formation of dTMP and dUMP and is one of several salvage enzymes that appear redundant to the de novo pathway. Surprisingly, we show through analysis of TK conditional null and RNAi cells that TK is essential for growth and for infectivity in a mouse model, and that a catalytically active enzyme is required for its function. Unlike humans, T. brucei and all other kinetoplastids lack dCMP deaminase (DCTD), which provides an alternative route to dUMP formation. Ectopic expression of human DCTD resulted in full rescue of the RNAi growth phenotype and allowed for selection of viable TK null cells. Metabolite profiling by LC-MS/MS revealed a buildup of deoxypyrimidine nucleosides in TK depleted cells. Knockout of cytidine deaminase (CDA), which converts deoxycytidine to deoxyuridine led to thymidine/deoxyuridine auxotrophy. These unexpected results suggested that T. brucei encodes an unidentified 5'-nucleotidase that converts deoxypyrimidine nucleotides to their corresponding nucleosides, leading to their dead-end buildup in TK depleted cells at the expense of dTTP pools. Bioinformatics analysis identified several potential candidate genes that could encode 5’-nucleotidase activity including an HD-domain protein that we show catalyzes dephosphorylation of deoxyribonucleotide 5’-monophosphates. We conclude that TK is essential for synthesis of thymine nucleotides regardless of whether the nucleoside precursors originate from the de novo pathway or through salvage. Reliance on TK in the absence of DCTD may be a shared vulnerability among trypanosomatids and may provide a unique opportunity to selectively target a diverse group of pathogenic single-celled eukaryotes with a single drug.
| Human pathogenic trypanosomatids are responsible for several life threatening diseases, together infecting 20 million people, while treatment is complicated by poor drug therapies. The unique biology of these organisms has led to the need for different drug therapies to be developed for each. Identification of enzymatic targets that could be used to develop a single drug capable of treating multiple parasites would be revolutionary. Herein we show that the trypanosomatid biosynthetic pathway used to synthesize key precursors for DNA biosynthesis is unexpectedly vulnerable. We find that seemingly redundant enzymes thymidine kinase and cytidine deaminase are required, not for their typical role in salvaging exogenous precursors, but are instead essential for de novo synthesis of thymine nucleotides. Trypanosomatids lack alternative routes to synthesize these nucleotides, which are found in other eukaryotic cells, while encoding a previously unknown activity that degrades them. For these reasons thymidine kinase is essential to support infection in mice and has strong potential as a new drug target. While our work focused on the causative agent of African sleeping sickness, the impact of our findings may extend to other pathogenic trypanosomatids and potentially to additional single-celled eukaryotic human pathogens.
| The parasitic trypanosomatids are vector-borne single-celled eukaryotic pathogens that cause significant disease and mortality in tropical and subtropical countries [1]. Human African trypanosomiasis (HAT), Leishmaniasis and Chagas disease affect 20 million people combined, but control is hampered by lack of good drugs, drug resistance and challenges in drug administration [2, 3]. New drugs for the treatment of all three diseases are badly needed. HAT, also known as sleeping sickness, is caused by Trypanosoma brucei, an extracellular parasite that replicates in the blood in early stages of infection. It crosses the blood-brain barrier in later stages leading to progressive neurological complications that disrupt the sleep/wake cycle and which eventually progress to coma and death [4]. The WHO estimates that yearly cases have dropped below 10,000 [5]; however, attempts to fully eliminate the disease face the difficulties of a substantial animal reservoir and problematic drug therapies that have high toxicity or are difficult to administer in a rural setting.
Pyrimidine and purine biosynthesis is essential in trypanosomatids to generate precursors needed for the biosynthesis of DNA, RNA and sugar nucleotides [6, 7]. Purines are obtained entirely by salvage routes through an array of interconnected and seemingly redundant pathways [8]. However, despite this redundancy, enzymes from the purine pathway including GMP synthase have been shown to be essential for pathogenicity in vivo [9]. In contrast, trypanosomatids are able to synthesize pyrimidines either through the de novo biosynthetic pathway or through salvage of preformed nucleosides and bases [7, 10, 11]. Genes have been identified for the complete de novo pyrimidine biosynthetic pathway, for several key salvage enzymes and for a number of interconversion enzymes [12](Fig 1). Genetic knockout studies have shown that loss of various de novo pyrimidine biosynthetic enzymes leads to pyrimidine auxotrophy that can be rescued by exogenous uracil [13–15]. These findings are consistent with reports that uracil transport is the primary route for pyrimidine salvage [10]. However knockout of UMP synthase lead to avirulence in mice suggesting that in vivo pyrimidine salvage may be insufficient to completely overcome loss of the de novo pathway [15]. These studies have shown that despite apparent redundancy, enzymes in both the pyrimidine and purine biosynthetic pathways can be essential, especially for in vivo virulence of T. brucei, which as an extracellular parasite lacks access to high intracellular concentrations of metabolites.
T. brucei lack several transporters and enzymes found in higher eukaryotes that may make them more vulnerable to disruption of the pyrimidine biosynthetic pathway. The primary pyrimidine transporter in T. brucei preferentially takes up uracil, whereas transport of uridine, 2’-deoxyuridine, thymidine and cytidine is either non-existent or inefficient requiring high nucleoside concentrations [10]. Trypanosomatids lack dCMP deaminase (DCTD), an important contributor to dTTP biosynthesis through deamination of dCMP to dUMP in many higher eukaryotes [17, 18]. Instead trypanosomatids were thought to rely on dUTPase to convert uracil nucleotides synthesized by the de novo pathway into the thymine nucleotide pools [19].
T. brucei encodes three pyrimidine salvage enzymes: uracil phosphoribosyltransferase (UPRT), thymidine kinase (TK), and uridine phosphorylase (UPP), and additionally a cytidine deaminase (CDA) that can convert deoxycytidine to deoxyuridine (Fig 1). T. brucei UPP was shown to prefer uridine and deoxyuridine as substrates, and the enzyme was reported not to be essential based on RNAi knockdown studies [20]. In contrast to mammalian cells, which encode both cytosolic TK1 and mitochondrial TK2 [21–23], trypanosomatids possess only TK1. TbTK is however a unique fusion of two TK domains that function as a pseudodimer. The N-terminal domain is catalytically inactive, while the C-terminal domain exhibits canonical TK activity [24]. Recent RNAi studies have suggested that TK is essential but a mechanistic understanding for why TK would be required has not emerged [25, 26].
Herein we report the results of genetic studies to determine the roles of pyrimidine salvage enzymes in pyrimidine nucleotide biosynthesis, growth and virulence in T. brucei. We found that knockdown of either TK or CDA led to cell death despite the finding that T. brucei does not require exogenous pyrimidines for growth. The TK conditional null mutant (c-null) could not be rescued by exogenous pyrimidines, whereas the CDA null was auxotrophic for thymidine or deoxyuridine. We were able to abolish the parasite’s reliance on TK by providing an alternative metabolic route to dUMP formation via DCTD. Metabolomic analysis showed that a significant cellular response to TK depletion was dead-end formation of pyrimidine nucleosides resulting in depletion of dTTP. These data taken together with the phenotype of the CDA null suggested the presence of an unidentified 5'-nucleotidase that promotes interconversion between the cytosine and thymine deoxynucleotide pools. Bioinformatics analysis suggested that T. brucei encodes a number of potential 5’-nucleotidases and we show that an HD-domain protein related to a well-characterized bacterial 5’-nucleotidase is able to catalyze this reaction. Thus our data show that TK is essential for the de novo biosynthesis of dTTP and that together with CDA and the 5’-nucelotidase these enzymes play a key role in maintaining balanced levels of the various deoxyribonucleotide pools. We propose that the absence of DCTD in trypanosomes, and other protists, may be a shared vulnerability, presenting a potential opportunity to develop a pan-trypanosomatid TK inhibitor to exploit this unique feature of the protozoan pyrimidine pathway.
To evaluate the essentiality of TK, we attempted to generate a TK null cell line in T. brucei BSF single marker (SM) cells. T. brucei is diploid, therefore minimally two alleles are present for each gene. Allelic replacement via homologous recombination was achieved by transfecting parasites with a PCR product containing a resistance marker flanked by the TK 5’ and 3’ UTRs. Removal of the first allele was successful; however we were unable to replace the last allele after repeated attempts, suggesting essentiality. To further evaluate this hypothesis, a TK conditional null (TK c-null) cell line was generated. The single allele knockout cells (SKO) were transfected with a vector conferring tetracycline (Tet) regulated expression of N-terminally FLAG tagged T. brucei TK (FLAG-TbTK). Expression of FLAG-TbTK was induced by addition of Tet and the remaining TK allele was successfully removed generating the TK c-null cell line. PCR amplification of the TK locus confirmed replacement by the two selectable markers (S1 Fig). To determine the effects caused by the loss of TK expression on cell growth, Tet was removed from the medium, which led to rapid growth arrest and near total cell death by day 3 (Fig 2A). Coincident with this growth arrest, TK transcript (by qPCR) and the TK protein (western blot) were depleted within 24 h after Tet removal, confirming good regulatory control of the ectopic TK copy (Fig 2A and 2B). Parasites reemerged several days later, likely due to the loss of Tet regulation. This has been reported to be a common phenomenon in T. brucei, likely due to mutations resulting in the loss of Tet regulation, e.g. [9, 27–29]. The TK c-null cells grew normally in pyrimidine-free medium (containing dialyzed fetal bovine serum) in the presence of Tet confirming that T. brucei is not auxotrophic for pyrimidines (Fig 2A). Upon subsequent removal of Tet, TK depletion led to cell death with a similar time course to medium containing non-dialyzed (normal) serum.
In vivo studies were performed to determine if TK was essential to support T. brucei infection in mice. In parallel to T. brucei SM infected mice, two groups (n = 3) of TK c-null infected mice were given either doxycycline (Dox) treated water or water only. As expected mice infected with SM cells in either condition (+/- Dox) had detectable levels of parasitemia by 72 h post infection, with fatalities occurring in all mice in both groups by day 6 (Fig 2C and S2 Fig). Similarly, all TK c-null infected mice treated with Dox to maintain expression of the Tet-regulated TK copy eventually died within the timeframe of the study. One mouse in this group showed a delayed time before succumbing to parasitemia suggesting some variability in TK expression levels in the TK c-null cells. Mice infected with the TK c-null strain treated only with water remained healthy and had no detectable levels of parasitemia past 30 days. Thus we conclude that TK is essential for T. brucei virulence and infectivity in vivo.
The finding that TK is essential in T. brucei is puzzling as no clear mechanistic role for TK in parasite fitness is apparent. T. brucei requires TK, which is a pyrimidine salvage enzyme; yet there is no requirement for salvageable pyrimidines for growth. To gain further mechanistic insight into this conundrum, we sought to address three possible explanations for the essentiality of TK: 1) parasites require an active TK enzyme, but it makes a novel product; 2) the TK protein, but not its catalytic activity is needed in some regulatory capacity; 3) parasites require formation of dUMP/dTMP by TK to balance pathway flux even under conditions where all pyrimidine precursors originate from the de novo pathway.
To provide additional mechanistic insight, an inducible RNAi cell line targeting TK mRNA was created so that we could easily introduce various rescue plasmids to address our mechanistic hypotheses. A Tet-regulated vector capable of producing a hairpin transcript targeting the 3’UTR of TK was generated and transfected into the TK SKO cell line. Induction of TK RNAi by addition of Tet led to a significant growth defect, although the growth defect was not as severe as observed for TK c-null cells (Fig 3A and 3B). TK mRNA expression was reduced to 20–25% of wild-type control levels by RNAi targeting the TK transcript (Fig 3C). However, the reduction in TK transcript levels was less in comparison to that observed in the TK c-null cells, explaining why the effect on cell growth was less pronounced.
To shed light on whether a novel TK product was being formed we transfected parasites with plasmids encoding rescue proteins from three different sources: AU1-tagged T. brucei TK, FLAG-tagged human HsTK, and Herpes simplex TK (HsvTK). HsTK has been shown to have more stringent substrate specificity than the T. brucei enzyme [24], whereas the viral HsvTK possesses broader substrate specificity than the human enzyme [30]. Rescue protein expression was also under control of Tet promoter. Thus, addition of Tet to these cells induces simultaneous knockdown of endogenous TK and expression of the tagged rescue protein. We found that the growth phenotype was reversed by expression of TK from all three species: T. brucei (TbTK)(Fig 3A), human (HsTK) (Fig 3B) and viral TK (HsvTK) (S3 Fig). Expression of TbTK and HsTK was confirmed by western blot (Fig 3A and 3B) and knockdown of endogenous TK was monitored by qPCR (Fig 3C). Viral TK expression was confirmed by the observance of ganciclovir sensitivity that was less apparent in cells expressing T. brucei TK (S3 Fig). Ganciclovir is a subversive substrate of HsvTK leading to premature chain termination of newly synthesized DNA [31]. The ability of both HsTK and HsvTK to rescue the TK RNAi growth phenotype shows that T. brucei TK is unlikely to catalyze a novel reaction, as the required activity is present in enzymes from other species that are known to have a range of substrate specificities.
To confirm that TbTK’s essential function is dependent on catalytic activity, mutations in the active site of both TbTK and HsTK were created. We targeted two conserved residues (S4 Fig) with described roles in the TK catalytic mechanism: T. brucei E286, which is reported to function as a proton acceptor [32] and human K32 which is an essential ATP binding residue [33]. Rescue plasmids were constructed as described above with the mutant TKs under the control of the Tet promoter and transfected into the TK RNAi line. In contrast to the wild type enzymes, neither the TbTK E286A nor HsTK K32I active site mutants were able to reverse the RNAi induced growth phenotype (Fig 3D and 3E). These data demonstrate that TK catalytic activity is required for its role in T. brucei cell survival.
While our data clearly show that T. brucei is not a pyrimidine auxotroph, we exploited the fact that the TK RNAi line retains partial TK activity (the knockdown is only 75–80% effective by RNAi (Fig 3C)) to assess whether we could use pyrimidine rescue to determine which TK product was needed for T. brucei growth. We found that high concentrations (significantly above physiological levels) of deoxyuridine (dUrd) resulted in the partial rescue of the RNAi growth phenotype (Fig 4A and S5A Fig) (5 mM rescued but 1 mM did not). However, similar levels of uridine (Urd) (Fig 4B and S5B Fig) or thymidine (dThd)(Fig 4C and S5C Fig) did not restore growth and dThd (0.15–1.0 mM) was in fact growth inhibitory to the TK RNAi +Tet induced cells but not to cells that expressed TK (-Tet), suggesting some type of feedback regulation. In contrast, the addition of dUrd or uracil to TK c-null cells, which are >99% depleted of TbTK, were unable to circumvent lethality of the TK knockout showing that TK activity is required for dUrd rescue (Fig 4C and S6 Fig). These data confirm that TK plays an essential role in maintaining dUMP pools.
The ability of dUrd to partially reverse the RNAi growth phenotype highlights an interesting feature in T. brucei pyrimidine metabolism. In most mammals, it has been suggested that a significant portion of dTTP is derived from dUMP produced by dCMP deaminase (DCTD)[17, 18]. Trypanosomatids lack this enzyme, restricting the number of metabolic routes dedicated to dUMP formation. We hypothesized that due to the lack of DCTD, trypanosomatids require TK to supplement dUMP pools. To test this hypothesis a Tet-regulated vector encoding human DCTD (HsDCTD) was transfected into the TK RNAi cell line to drive simultaneous expression of HsDCTD and knockdown of endogenous TbTK. The expression of FLAG-tagged HsDCTD completely rescued the TK RNAi-induced growth phenotype (Fig 5A). Expression of FLAG-tagged HsDCTD was confirmed by western blot and qPCR analysis confirmed that TK mRNA expression was simultaneously reduced (Fig 5B). To further demonstrate that HsDCTD can functionally replace TK, a TK null cell line was created in the background of FLAG-tagged HsDCTD Tet-regulated expression plasmid. Both TK alleles were replaced by selectable markers through homologous recombination in the presence of Tet to maintain expression of HsDCTD. PCR amplification of the region flanking the TK 5’ and 3’ UTRs confirmed replacement of TK with the selectable markers (S1 Fig). PCR analysis also confirmed that the TK gene was no longer detectable in genomic DNA from the TK null cells. Removal of Tet from this cell line led to depletion of FLAG-tagged HsDCTD and resulted in a severe growth phenotype by day 2 after Tet removal (Fig 5C). Cell growth of the HsDCTD TK null line was less severely impacted than the TK c-null line expressing Tet-regulated TbTK from the rescue plasmid, perhaps reflecting a higher residual expression level of HsDCTD (Fig 5C). By day 5 after Tet withdrawal cells began growing again coincident with re-expression of HsDCTD, again suggestive of emergence of cells that have mutations leading to loss of Tet regulation (Fig 5C). Thus, the data support the hypothesis that TK is essential in T. brucei and that its role is to contribute to the formation of dUMP in the absence of DCTD.
Analysis of ~130 soluble metabolites from TK c-null cells was performed by liquid chromatography-tandem mass spectrometry (LC-MS/MS) to determine the impact of TK depletion on metabolite pools (Fig 6 and S7–S9 Figs and S1 Appendix). Extracts were collected from TK c-null cells grown in medium containing normal non-dialyzed serum at 24 h ± Tet. An early time point was selected so that metabolite pools would be less affected by non-specific changes resulting from cell death later in the time course. Significant changes in the measured metabolite levels were mostly confined to pyrimidine nucleosides: there was a 15-30-fold accumulation of the TK substrates dUrd and dThd, a 3-fold accumulation of dCtd, and a 70-fold increase in thymine levels (Fig 6A). In contrast, the pyrimidine nucleotides CMP and UMP were not significantly changed by TK depletion, suggesting that the de novo pathway was able to maintain the uridine nucleotide pools. In further support, HPLC analysis of UDP-sugars was performed (Fig 6B). Nucleotide sugars are formed from UTP, thus their measurement provides a read-out of effects on intracellular UTP concentrations. The relative abundance of UDP-GlcNAc, UDP-Galactose, and UDP-Glucose were not significantly changed by TK depletion, confirming that UTP pools are not linked to TK activity. Thymidine nucleotide pools were not detected by the LC-MS/MS analysis so instead we quantitated dTTP levels using an enzymatic assay and found that dTTP levels were reduced to 30% of control levels 48 h after Tet removal, confirming TK is essential for synthesis of dTTP (Fig 6C).
To better understand the consequences of TK depletion in a pyrimidine-free medium the LC-MS/MS analysis was repeated for TK c-null cells grown in medium supplemented with dialyzed serum (Fig 6C and S7–S9 Figs and S1 Appendix). Similar to the results for cells grown in normal serum based medium, we observed a statistically significant buildup of TK substrates dUrd and dThd, and of the dUrd precursor dCyd, though the increases (3-6-fold) were less than observed for cells grown in normal serum based medium. We also observed changes in the levels of TCA intermediates including decreases in citrate and aconitate and an increase in α-ketoglutarate, and perturbations in other metabolites related to pyrimidine biosynthesis including decreased carbamoylphoshate (CP) and acetyl-ornithine (30 and 80%, respectively) and a 2-fold increase in homocysteine (Fig 6C and 6D). Not all of these latter changes reached statistical significance. Some modest effects were also observed on several purines: an increase in xanthine and/or hypoxanthine was observed in both medium conditions, which may suggest some type of cross-regulation between pyrimidine and purine biosynthetic pathways. Thymine was only observed in cells grown in medium supplemented with non-dialyzed serum, which could supply a source for thymine, confirming genomic analysis that suggests T. brucei does not encode thymine biosynthetic enzymes.
The finding that the deoxypyrimidine nucleoside pools buildup upon TK depletion for cells grown in pyrimidine-free (dialyzed serum based) medium suggests that an undiscovered biosynthetic route for formation of deoxypyrimidine nucleosides must be present in T. brucei. To investigate if deoxycytidine could be involved in formation of dUMP and dTMP we decided to characterize the effects of depleting cytidine deaminase (CDA) on parasite growth. Based on the current annotation of the pyrimidine biosynthetic pathway, CDA like TK should catalyze a redundant reaction, since the deoxynucleotide pools can be supplied from the de novo pathway. However, attempts to generate a CDA null were unsuccessful, and indeed the CDA null could only be obtained in the presence of 500 μM dThd (Fig 7 and S10 Fig). Upon removal of dThd, CDA null cells exhibited a severe growth defect (Fig 7A). Cells cultured for longer eventually died around day 7–10 after dThd removal. Additional pyrimidine rescue studies revealed that both dThd and dUrd were capable of rescuing growth (EC50 of 6–20 μM), whereas uracil (up to 250 μM) was not (Fig 7C–7E). The concentrations of dThd and dUrd required for rescue are significantly above reported human blood levels (range of 0.2–0.6 μM; http://www.hmdb.ca).
CDA null cells were grown ± dThd for 12 h in media supplemented with dialyzed FBS and the metabolite pools were analyzed by LC-MS/MS as described above. Overall the effects of CDA depletion on cellular metabolism were very similar to the effects observed after depletion of TK, confirming a link between the roles of the two enzymes (S11 Fig, S12 Fig and S1 Appendix). We observed a significant buildup (~7-fold) in the CDA substrate, dCtd, for the CDA null cells grown in the absence of Thd that was accompanied by changes in TCA intermediates and other metabolites that serve as precursors for de novo pyrimidine biosynthesis. These changes include increases in α-ketoglutarate (11-fold), glutamate (7-fold) and homocysteine (2-fold), and decreases in carbamoyl phosphate (2-fold). We also observed decreased levels of several amino acids, of polyamines, particularly N-acetyl putrescine and decreased levels of several purine mono-phosphate nucleotides. Neither dUrd nor dThd were detected in CDA null cells in the presence or absence of added dThd, suggesting they were either not formed (dUrd) or rapidly metabolized (dThd). In the absence of CDA, the buildup of dCtd suggests that the pyrimidine deoxynucleotides have been sequestered into a dead-end product that would be expected to lead to depletion of key nucleotides and to cell death as was observed for the TK null cells. These data support a role for CDA in the interconversion of deoxycytidine, deoxyuridine and thymidine pools, which is needed to balance these pools in a cell like T. brucei that expresses TK but not DCTD.
In order to identify the potential enzyme(s) responsible for converting deoxypyrimidine nucleotides into their corresponding nucleosides we undertook a bioinformatics analysis of the T. brucei genome. 5'-nucleotidase activities that catalyze formation of deoxyuridine, deoxycytidine, or thymidine from their respective mono-phosphate nucleotides are ascribed to two EC numbers (EC3.1.3.5 and EC3.1.3.89). Inspection of domain types performing these activities revealed nine protein families defined by PFAM, with all having homologous structure representatives in the PDB (Table 1). The families further merge into five different homologous fold types; including three different α/β sandwich folds (phosphoglycerate mutase-like, SurE-like, and HAD domain-related), one α+β four-layer sandwich fold (metallo-dependent phosphatases), and one all-α fold (HD-domain). The phosphoglycerate mutase-like representatives are limited to mammalian acid phosphatases, prostate (ACPP) enzymes that convert extracellular AMP to adenosine (i.e. ecto 5' nucleotidase activity) [34]. An additional identified enzyme (NT5E) from the metallo-dependent phosphatase fold group exhibits a similar ecto 5'-nucleotidase activity [35]. Examples of this fold group in both eukaryotes and bacteria contain signal peptides and are extracellular. In contrast, the remaining SurE, HAD domain-related, and HD-domain representatives appear cytosolic.
We found evidence for all of the 5'-nucleotidase homologous fold types in the T. brucei genome with the exception of the SurE fold class (S1 Table). We identified nine phosphoglycerate mutase-like sequences, 33 metallo-dependent phosphatase sequences, 16 HAD domain-related sequences, and one HD-domain sequence (S1 Table). The presence of numerous examples of potential 5'-nucleotidase domains, many of which are annotated as hypothetical proteins, suggests multiple possible proteins that T. brucei could use to form deoxyuridine, deoxycytidine, or thymidine de novo. However, three of the identified genes possess specific PFAM domains described as having 5’-nucleotidase activity (EC 3.1.3.5 or EC 3.1.3.89) and thus are the highest ranked candidates. One encodes a HD domain protein: hypothetical protein (Tb09.211.2190) and two encode HAD-like domains: a putative p-nitrophenylphosphatase (Tb927.8.7510) and a hypothetical protein (Tb09.211.1880). The T. brucei HD domain protein is a homolog of E. coli 5’-nucleotidase YfbR (S13 Fig) while the T. brucei HAD-domain proteins are related to enzymes shown to have 5’-nucleotidase activity in both yeast and E. coli. [36–38]. These enzymes have been reported to have broad substrate specificity functioning on all three pyrimidine deoxyribose monophosphates.
To provide support for our hypothesis that T. brucei encodes a 5’-nucleotidase we cloned, expressed and purified the recombinant T. brucei HD domain homolog (Tb09.211.2190) of bacterial 5’-nucleotidase YfbR (S14 Fig). We found that the T. brucei YfbR-like HD protein showed a metal dependent 5’-nucleotidase activity (Fig 8A). Similar to the bacterial enzyme it was most active in the presence of Co+2 (0.5 mM), but activities within 2-4-fold of levels observed for Co+2 were also obtained using Mn+2 (0.5 mM) and physiological levels of Mg+2 (10 mM). No activity was observed in the presence of Zn+2 or EDTA. The specific activity of the T. brucei HD domain 5’-nucleotidase was very similar to the reported activity of E. coli YfbR [38]. The T. brucei enzyme showed a broad substrate range functioning on both pyrimidine and purine deoxyribonucleoside and ribonucleoside 5’-monophosphates, though it was most active on the deoxypyrimidine nucleotides (dCMP, dUMP and dTMP)(Fig 8B). It showed no activity towards diphosphate nucleotides. The T. brucei enzyme was somewhat more promiscuous then E. coli YfbR, which was unable to catalyze hydrolysis of ribonucleoside 5-monophosphates [38]. Both T. cruzi and Leishmania encode homologs of the T. brucei HD-domain 5’-nucleotidase (S1 Table and S13 Fig) suggesting they both also will be able to convert 5’-deoxyribonucleotide monophosphates to their respective nucleosides.
To assess if TK essentiality was likely to extend to other pathogenic protozoa we utilized the KEGG pathway database to determine the distribution of TK and DCTD throughout eukaryotes (Fig 9). A striking disparity was observed within protists when compared to higher eukaryotes. The vast majority of higher eukaryotes possess both TK and DCTD, which may explain TK’s non-essential role in these organisms. In contrast, the kinetoplastids and a number of other protozoan human pathogens such as Giardia encode only TK, suggesting that TK may be essential in these organisms as well. We also note that several of the protists such as Entamoeba histolytica, which lack DCTD, instead encode dCTP deaminase, an enzyme found almost exclusively in bacteria. Similar to DCTD, the ability to deaminate dCTP to dUTP offers an alternative path from cytosine to thymine nucleotide pools and thus we would predict that TK would not be essential in these species. Interestingly, these organisms represent the only eukaryotic KEGG organisms that have dCTP deaminase.
T. brucei encodes a complete de novo pyrimidine biosynthetic pathway, as well as a number of pyrimidine salvage enzymes that were thought to be redundant based on the presence of the de novo pathway. Herein we describe the first comprehensive analysis of the role of the pyrimidine salvage enzymes in T. brucei and we show that while T. brucei is not auxotrophic for pyrimidines, both TK and CDA are essential for in vitro growth and TK is essential for infectivity in vivo as well. The finding that these enzymes are essential could not be explained by the current annotation of the pyrimidine pathway in T. brucei. Our mechanistic analysis of the TK and CDA null cell lines uncovered the existence of an interconversion network between the deoxypyrimidine nucleoside and nucleotide pools, including the presence of a previously unknown 5’-nucleotidase that converts deoxycytidine, deoxyuridine and thymidine nucleotides to their respective nucleosides. In the absence of TK or CDA to balance this 5’-nucleotidase activity, the metabolic cycle breaks down leading to dead-end buildup of deoxypyrimidine nucleosides and to cell death. The existence of this recycling pathway provides a mechanism for the parasite to interconvert and balance the relative levels of the deoxyuridine, deoxycytidine and thymidine pools whether they originate from the de novo pathway or through salvage. Our conclusions are supported by the following arguments.
Firstly, TK is essential for both in vitro growth and infectivity in a mouse model of T. brucei infection and for formation of dTTP despite the fact that T. brucei is not a pyrimidine auxotroph. Thus the essential role of TK is not to salvage externally acquired pyrimidine precursors. Our data clearly show that TK activity is required for its function and that it plays a key role in the synthesis of dUMP, even for cells grown in a pyrimidine-free environment. We found that the function of TK can be replaced by expression of human DCTD, which provides an alternative route to dUMP formation from dCMP in many higher eukaryotes [17, 18]. DCTD has been shown to be essential for cell cycle progression and formation of dTTP pools in eukaryotes that lack TK (e.g. Schizosaccharomyces pombe) [39]. These data suggest that DCTD and TK can have functionally redundant roles in contributing to dTTP pools, supporting our observation that TK is essential for formation of thymine nucleotides in T. brucei.
The next significant key to the puzzle came from analysis of metabolomic data from the TK c-null cell line. These data showed that even in the absence of external pyrimidines the TK substrates dUrd and dThd, as well as the dUrd precursor dCtd buildup, leading to a dead-end accumulation of these precursors away from the essential deoxynucleotide pools resulting in depletion of dTTP. In the absence of an exogenous supply of these nucleosides the current annotation of the T. brucei genome does not provide a mechanism for these nucleosides to be synthesized, suggesting the presence of a missing enzyme that catalyzes conversion of deoxynucleotides into deoxynucleosides. The findings that CDA null cells are auxotrophic for dThd or dUrd further support this hypothesis since based on redundancy in the pathway, CDA should not be essential under any conditions. Furthermore the CDA null data support the presence of an enzymatic link between the deoxycytidine-containing nucleotide pools and dCtd/dUrd, since either dUrd or Thd are required for growth of CDA null cells. These data are consistent with previous published untargeted metabolomics data showing that isotope-labeled glucose was incorporated into both dUrd and dThd, and thus that T. brucei was capable of synthesizing these nucleosides de novo [40]. Lastly, the inability of uracil to rescue the growth deficit of the CDA null cells shows that uridine phosphorylase is not able to efficiently convert uracil to dUrd, eliminating the only known potential source for dUrd biosynthesis in T. brucei. Uridine phosphorylase was previously suggested to be the source of dUrd, based on the isotope-labeled glucose study [40], but our result is instead consistent with previous reports that 5-fluorouracil and 5-fluoro-Urd are not substrates for this enzyme [10].
Thus taken together, our data lead to the conclusion that T. brucei encodes an unidentified 5'-nucleotidase that converts dCMP and dTMP to dCtd and dThd, respectively. We identified a number of potential candidate genes in T. brucei that could encode this activity, including a homolog of the E. coli HD protein YfbR and two strong candidates from the HAD-domain related family. Notably we showed that the T. brucei YfbR homolog encodes a metal dependent HD domain 5’-nucleotidase with broad substrate specificity functioning on all three pyrimidine deoxy-mononucleotides. Whether or not the T. brucei YfbR homolog is the only 5’-nucleotidase in T. brucei, or whether it is even the dominant enzyme with this capability remain open questions. Mammalian cells encode at least seven 5’-nucleotidases with overlapping specificities [41–43] and E. coli encodes minimally three, one each from the HD, HAD and SurE superfamilies [38]. Thus it is likely that other candidate T. brucei genes identified in our bioinformatics analysis will also display activity. In mammalian cells the 5’-nucleotidases have been shown to be required for regulation of cellular dNTP levels and to provide a mechanism to maintain balanced ratios between the pools, which is essential for high fidelity DNA synthesis [43]. Like the T. brucei HD-domain 5’-nucleotidase, all described nucleotidases from the various families exhibit broad substrate specificity. The broad specificity allows these enzymes to function in a ubiquitous capacity for interconversion of the nucleotide pools.
The finding of 5’-nucleotidase activity in T. brucei leads directly to the essentiality of both TK and CDA, as in their absence the dead-end buildup of pyrimidine nucleosides leads to depletion of pyrimidine deoxynucleotides and to cell death. Within this context, the ability of DCTD to rescue the TK null cell line suggests that DCTD is able to effectively compete with the 5'-nucleotidase for the dCMP pools, converting sufficient amounts to dUMP where it can be efficiently shunted to dTMP even in the absence of TK. The existence of the metabolic cycle involving TK, CDA and 5’-nucleotidase provides the cell with a mechanism to interconvert between the deoxyuridine, deoxycytidine and deoxythymidine pools allowing presumably for better regulation and balance of their relative levels. While dUMP can also be formed from UDP, this pathway is apparently not sufficient to keep up with dUMP needs in the face of the dead-end accumulation of the TK substrates in the absence of TK. However this pathway remains an important additional source of dUMP as null mutants of dUTPase have been reported to be thymidine auxotrophs [19].
Our metabolomic analysis also uncovered some additional insights into T. brucei metabolism and regulation. In the presence of an outside source of pyrimidines (non-dialyzed serum), the pyrimidine nucleosides dUrd, dCyd, dThd and thymine accumulated in the TK c-null cells to higher levels than for cells grown in pyrimidine-free medium (dialyzed serum). These data confirm that in the absence of TK there is dead-end accumulation of these nucleosides but they also suggest that uridine phosphorylase is not a significant drainage point for these pools. Thus T. brucei uridine phosphorylase primarily catalyzes conversion of Urd to uracil, while it is not capable of synthesizing dUrd (as described above), or using it efficiently as a substrate. This hypothesis is consistent with previous reports that the recombinant T. brucei enzyme is 10-fold more active on Urd than dUrd [20]. Our metabolomic data also suggest that one response of T. brucei to TK depletion is increased nucleoside transport despite the fact that the upregulated transport was unable to relieve the growth block. A similar accumulation in dUrd in the presence of normal serum was previously reported for T. brucei BSF treated with thymidylate synthase inhibitors [10] suggesting this is a common response to starvation of thymine nucleotides.
Finally, we noted that the levels of TCA intermediates were significantly perturbed in both the TK c-null and CDA null cells, including significant increases in α-ketoglutarate and homocysteine upon loss of TK expression or removal of thymidine from the CDA null cells. α-ketoglutarate is formed in the transamination reaction that generates L-Asp, which in turn is required for the first step in de novo pyrimidine biosynthesis, while homocysteine leads to formation of methionine then 5,10-methylene tetrahydrofolate, needed to convert dUMP to dTMP. Taken together with an observed decrease in carbamoyl phosphate, another precursor of the de novo pathway, the data suggest the cells may attempt to compensate for the loss of TK by increasing flux through the de novo pathway. Finally we also observed a significant decrease in acetyl-ornithine/acetyl-putrescine. It is not immediately apparent how these metabolites are synthesized, but their presence in T. brucei has been previously noted [40]. It is also not immediately clear what role they may play in pyrimidine biosynthesis, but both the synthesis and degradation of acetyl-ornithine can be catalyzed by aminotransferases, and in the case of its degradation this pathway links back to glutamate pools, and thus potentially to pyrimidine biosynthesis. The specific aminotransferases that catalyze these reactions are not annotated in the T. brucei genome, but aminotransferases have been reported to have broad and redundant substrate specificities in E. coli [44].
While a key aspect of our work was to elucidate the role of TK and CDA in linking the de novo pathway to synthesis of the deoxynucleotide pools, we have also validated TK as a drug target in T. brucei by showing that it is essential both in vitro and in vivo. The finding that the TK c-null cells cannot be rescued by exogenous pyrimidines shows it would not be possible for even an intracellular parasite to get around the block. Our work additionally showed that CDA is essential for in vitro growth of blood form T. brucei. While we did not determine if CDA is required for virulence in vivo, it remains a possibility provided that blood thymidine levels are below those required for rescue. Pyrimidine deoxynucleosides levels in human blood are reported to be ~10-fold below the EC50 that we measured for efficient rescue of CDA null growth. Furthermore, T. brucei has low affinity and/or poor efficiency transporters for deoxynucleosides [10], suggesting that CDA may be essential for infection in humans. However, additional studies will be needed to address this question conclusively.
The presence of multiple pathways to synthesize dUMP appears to be an important shared characteristic amongst many eukaryotic cells, with the data suggesting that some organisms require either TK or DCTD to link de novo biosynthesis to the thymine deoxynucleotides. Interestingly, our bioinformatics analysis shows that other single-celled eukaryotic pathogens, including all three disease-causing trypanosomatids, encode TK but lack DCTD. These data suggest that TK may be essential in these other pathogens and may potentially provide a path forward to develop drugs that have pan-activity against a range of human pathogens. However the essentiality in other organisms would be dependent on the presence of the 5’-nucleotidase activity and likely also on limited catabolism of dUrd back to uracil by uridine phosphorylase. In support, a homolog of the T. brucei HD domain 5’nucleotidase is present in Leishmania and T. cruzi. Furthermore Leishmania major TK null cell line showed severely reduced growth rates [45]. In contrast, deletion of TK from Cryptosporidium parvum was not lethal, which is predicted by the presence of both TK and DCTD [46]. The finding that human cells contain both TK and DCTD, and that TK is not essential in human cells [47] supports the potential for selectively targeting TK from the eukaryotic pathogens that lack DCTD. Thus in conclusion, the unexpected finding that TK is essential in T. brucei and its mechanistic role in supporting de novo pyrimidine biosynthesis has uncovered a unique opportunity for the potential development of a pan-trypanosomatid therapy.
All animal care and experimental procedures were in accordance with the office of Laboratory Animal welfare (OLAW) guidelines provided by the National Institute of Health (NIH) USA (Assurance Number D16-00296) and the USDA (Registration Number 74-R-0072) as specified by the University of Texas Southwestern Medical Center Animal Care and Use Committee (IACUC) guidelines. Animal experiments were approved by the Ethical Review Committee at the University of Southwestern Medical Center and performed under the IACUC-2012-0021 protocol. All mice were housed in an animal facility with barrier SPF conditions.
T. brucei gene sequences were obtained from TriTrypDB and the gene accession numbers are as follows: TK (Tb927.10.880), CDA (Tb927.9.3000), HD-domain 5’-nucleotidase (Tb09.211.2190), TERT (Tb927.11.10190). The HsTK1 (P04183) and HsDCTD (P32321) amino acid sequences were obtained from UniProt, and HsvTK was derived from Addgene plasmid #48356.
Experiments were performed using T. brucei BSF SM cells genetically manipulated to express T7 RNA polymerase and the Tet repressor (TetR) [29]. Cells were grown in HMI-19 medium supplemented with 10% fetal bovine serum (FBS) at 37°C in 5% CO2. HMI-19 is a modified medium that we previously reported [9]. It was designed to contain more physiologically relevant purine and pyrimidine levels and it is supplemented with only 10 μM hypoxanthine and no added thymidine, except that present in FBS. To obtain completely pyrimidine-free conditions, normal FBS was replaced with dialyzed FBS in media where indicated. All cells were maintained in exponential growth (105−106 cells/mL). SM cells were maintained in G418 (2.5 μg/mL) to retain the T7 polymerase and TetR. TK and CDA RNAi and knockout lines were cultured in the appropriate antibiotic depending on the transfected plasmid at the following concentrations unless otherwise stated: 2.5 μg/mL G418 (Life Technologies), 2.5 μg/mL blasticidin (InvivoGen), 2.5 μg/mL phleomycin (InvivoGen), 1–2 μg/mL hygromycin (Sigma), 0.1 μg/mL puromycin (Sigma), and 1 μg/mL Tet (RPI). For all nucleoside supplementation experiments 100 mM stocks of sterile filtered deoxyuridine (Sigma), thymidine (Sigma), uridine (Sigma), and uracil (Sigma) were added to cultures at concentrations indicated. All c-null lines were supplemented with 1 μg/mL Tet daily to maintain steady expression of Tet-regulated proteins. For pyrimidine and Tet free conditions, cells were washed (3 x 20 mL) with the appropriate media prior to beginning the growth experiments. For evaluation of growth rates, cells were washed and replated in media containing no antibiotics at a density of 20,000 cells/mL and diluted over the course of the study to maintain exponential growth. Cell density was determined using a hemocytometer (Bright-Line) with a lower limit of detection of 104cells/ml. Two technical replicates were averaged for each counted sample. Total cell numbers were calculated by multiplying cell density by the dilution factor and volume [48].
For each transfection, parasites (107) were suspended in Human T Cell Nucleofector Buffer (Lonzo)(100 μL) containing NotI linearized vector (5 μg) or purified PCR product (1 μg) as described [49]. All transfected DNA was confirmed by sequencing prior to transfection. Negative controls cells transfected with buffer only were prepared alongside samples to optimize selection conditions. Cells were transfected using protocol X-001 on the Amaxa Nucleofector (Lonza) and then transferred to media (25 mL) and allowed to recover 8 h prior to addition of selection antibiotics. Two dilutions (1:20 and 1:40) of culture, containing selection antibiotics, were plated in 24-well plates at 2 mL/well. Negative control plates were monitored throughout the experiment to ensure selection was achieved. After several days, wells containing a cell density of about 106 cells/ml were selected for generation of clonal lines by limiting dilutions.
The T. brucei TK and human (Hs) DCTD expressing TK c-null cell lines were generated utilizing the fusion PCR method [49, 50]. Cloning primers are shown in S2 Table. The first TK allele was replaced by the HYG resistance gene by PCR fusion of TK 5’ and 3’ UTRs to HYG. The HYG resistance gene was derived from the pLew90 vector [29] (a gift from George A.M. Cross). UTRs were amplified from genomic DNA isolated from SM cells. To generate the TK single allele knockout (SKO) line the purified PCR product was transfected into SM cells and hygromycin resistant cells were selected in medium containing G418 and hygromycin. The TK SKO line was then used to generate the remaining cell lines. A Tet-regulated vector containing either FLAG-tagged TbTK or FLAG-tagged HsDCTD was cloned as follows. The T. brucei TK or the human DCTD genes were PCR amplified from T. brucei SM genomic DNA or from human cDNA synthesized from RNA extracted from a human breast adenocarcinoma cell line (MDA-MB-231), respectively. For both constructs, the forward direction PCR primer contained a flanking 5’ HindIII restriction site and an N-terminal FLAG tag; the reverse direction primer contained a flanking 3’ BamHI restriction site. The restriction digested PCR products were ligated into the pLew100v5-phleo vector (a gift from George A.M. Cross). The pLew100v5-phleo vector was linearized by the NotI restriction enzyme to facilitate integration into the rRNA spacer region. Linearized vector (5 μg) was transfected into the TK SKO cell line and selected for resistance to phleomycin. The resulting clones were screened to identify those with the tightest level of Tet regulation of the ectopically expressed protein. Finally, the remaining TK allele was replaced by a PAC resistance gene synthesized by GenScript. The PAC TK UTR fusion product was generated as described above and transfected into TK SKO cells expressing either T. brucei TK or HsDCTD grown in Tet containing medium for 2 days prior to transfection. TK c-null cells were selected and maintained in G418, phleomycin, hygromycin, puromycin, and Tet (added daily). PCR primers flanking the 5’ and 3’UTRs were used to confirm that the TK gene had been replaced by the selectable markers.
RNAit (http://trypanofan.bioc.cam.ac.uk/software/RNAit.html) was used to identify a suitable 566 bp region located in the TK 3’UTR. The 3’UTR was targeted to allow compatibility with TbTK rescue plasmids (described below), that utilize instead the ALD 3’UTR. Genomic DNA isolated from SM cells was used as template for PCR amplification of the target region and TA cloned into the Gateway vector pCR8/GW/TOPO (Life Technologies). The Tet inducible stem loop was created by addition of Gateway LR Clonase to a reaction containing both the Gateway vector (100 ng) and pTrypRNAiGate vector (100 ng) [51]. TK SKO (hyg) cells were transfected with the vector and integration into the rRNA spacer region was selected using phleomycin. For studies of the effects of TK knockdown, Tet was added daily to induce formation of the hairpin leading to knockdown of TK mRNA. Cells were grown in the absence of other antibiotics for these studies.
The TbTK rescue construct under control of the Tet promoter was generated using the same approach described above for the c-null cell line except that an N-terminal AU1-tag was included instead of a FLAG-tag to allow detection of the expressed protein. The gene encoding the HsTK open reading frame was synthesized by GenScript and cloned into the pUC57 vector, which was used for subsequent PCR amplification to generate the FLAG-tagged HsTK rescue construct. HsvTK was amplified from the pHJ17 Hyg-TK-loxP vector (Addgene). To generate the TbTK E286A and HsTK K32I mutants, both wild-type genes were subcloned into the pCR2.1-TOPO TA vector (Invitrogen). Complimentary PCR primers containing the desired point mutation were synthesized. Phusion polymerase (NEB) was used to amplify the entire vector according the to the following parameters: initial denaturation at 95°C for 30 s followed by 18 cycles of denaturation at 95°C for 30 s, annealing at 68°C for 1 min, and amplification at 72°C for 5 min. Each reaction (50 μL) was treated with DpnI (NEB)(1 μL) overnight at 37°C followed by transformation into T10 cells and selection with ampicillin (100 μg/ml). Clones were sequenced using M13 primers. All constructs contained flanking 5’ HindIII and 3’ BamHI restriction sites that permitted ligation into the pLew100v5-bsd vector[49]. The vectors were linearized with NotI and transfected into TK RNAi cells as described above.
SM cells were transfected with the HYG resistance gene flanked by the CDA 5’ and 3’ UTRs, generated by fusion PCR as described above, to generate the SKO in medium containing G418 and hygromycin. For the remaining allele, a fusion PCR product containing the PAC resistance gene was transfected into the SKO cells. Null cells were selected in growth medium containing G418, hygromycin, puromycin, and Thd (500 μM). PCR primers flanking the CDA 5’ and 3’ UTRs were used to confirm replacement of the CDA alleles.
T. brucei TK expressing TK c-null cells were grown with or without Tet for 24 h and CDA null cells were grown with or without thymidine (0.5 mM) for 12 h. Cells (108) were harvested by centrifugation (3500 RPM, 5 min) and then washed in cold PBS (50 mL). Washed pellet was resuspended in 1 mL pre-chilled (-80°C) 80% methanol and incubated on ice for 10 min. The cell extract was centrifuged (16,000 x g, 4°C, 20 min) to remove insoluble debris and 0.9 mL of supernatant was dried using a vacuum centrifuge. Samples were stored at -80°C prior to analysis. For pyrimidine-free studies, a starter culture was washed, as described above, and grown in pyrimidine-free medium for 48 h prior to the start of the experiment. Targeted metabolite profiling by LC-MS/MS was performed as previously described allowing for detection of ~ 130 standard metabolites [52]. While this method allowed for quantitation of many key nucleosides and bases, the deoxynucleotides were not profiled as they are not part of the trained set of the facility. In order to attempt to identify these metabolites we isolated a larger cell number (5×108cells) and again used targeted LC-MS/MS for detection as described [53]. Levels of deoxynucleotides in wild-type control SM cells were barely detectable so null lines were not analyzed.
Because we were unable to quantitate dTTP by LC-MS/MS approaches we employed a previously reported enzymatic assay that monitors Klenow DNA polymerase catalyzed incorporation of 3H-dATP into synthetic oligonucleotides in a reaction that is proportional to the amount of dNTP [54, 55]. Through use of a standard curve the targeted dNTP concentration in the sample was determined. The oligo template that was used for the assay was as reported [54]. To prepare cell extracts, TK c-null cells were grown with or without Tet for 24 or 48 h in standard HMI-19 media supplemented with normal FBS and were harvested by centrifugation (3500 RPM, 5 min) and washed once with PBS. The pellets were resuspended in 60% methanol (250 μL) and incubated at -20°C overnight. Cell extracts were placed in a boiling water bath for 5 min, centrifuged (16,000xg, 20 min, 4°C), and then the soluble fraction was dried by vacuum centrifugation. Dried extracts were dissolved in 100 μL sample buffer (40 mM Tris-HCl pH 7.4, 10 mM MgCl2). Each reaction (100 μL) contained 40 mM Tris-HCl (pH 7.4), 10 mM MgCl2, 5 mM DTT, 0.25 μM oligonucleotide template, 1.5 μg RNase A, 0.25 μM 3H-dATP (ARC- 17.2 Ci/mmol), 0.3 units Klenow Fragment (NEB), and cell extract (10 μL) or dTTP standard. Reactions were incubated at 37°C for 1 hr before spotting (85 μL of reaction) onto DE81 paper disks (23 mm-GE Healthcare), which were then air dried. Disks were washed (3 x 10 min) with 25 mL 5% Na2HPO4, rinsed once with water (25 mL) and absolute ethanol (15 mL). Dried disks were placed in scintillation liquid and radioactivity was measured by scintillation counting. A standard curve was generated with 0–4 pmol dTTP (New England BioLabs).
T. brucei cells (2x107) were pelleted by centrifugation (3500 RPM, 5 min), washed in cold PBS once and then resuspended in 1 mL pre-chilled methanol. Samples were freeze-thawed in liquid nitrogen, placed on ice for 10 min, and centrifuged (16,000 x g for 10 min at 4°C) to remove insoluble debris. A portion of the supernatant (0.8 mL) was vacuum dried and the residue resuspended in 40 mM sodium phosphate, pH 7.4. Samples were syringe filtered (Millex GV 0.22 μM–Millipore) to remove fine particles. High-Performance Anion Exchange Chromatography (HPAEC) analysis was performed as published [56] but with a modified elution gradient that was optimized for the separation of UDP-GalNAc, UDP-GlcNAc, UDP-Gal and UDP-Glc. Briefly, chromatography was performed on a Dionex ICS3000 HPAEC system with a CarboPac PA1 analytical column (4 mm Å~ 250 mm) and guard column (4 mm Å~ 50 mm). The method was performed with eluents 1 mM NaOH (E1) and 1 M NaOAc, 1 mM NaOH (E2) as follows: 0 min—20% E2, 10 min—45% E2, 25 min—45% E2, 35 min—45% E2, 40 min—100% E2, 50 min—100% E2, 55 min—20% E2, 65 min—20% E2. Equal amounts of sample (approximately 50 Mio cells) were analyzed within each experiment and the signal divided by cell count. Synthetic standards were acquired from Promega.
DNAzol (Molecular Research Center) was used to isolate genomic DNA from T. brucei cells. Typically, 5x107 cells were collected for DNA extraction using guidelines recommended by the manufacturer. Total RNA was extracted from samples (3x107cells) using TRIzol (Invitrogen), following the manufacturer’s protocol.
As described above, total RNA was isolated from samples and treated with DNaseI (Invitrogen) to eliminate gDNA contamination. A cDNA reverse transcription kit (Applied Biosystems) was used to synthesize cDNA for downstream analysis. Relative mRNA abundance was quantified using iTaq SYBR Green Supermix with ROX (Bio-Rad) utilizing a standard curve for each set of primers per experiment. For all experiments, TERT was used as a reference gene [57]. Data was collected on the CFX96 (Bio-Rad) and analysis was performed using the Pfaffl method [58].
Mice (C57BL/6J) were purchased from Wakeland Laboratory (UT Southwestern) and were group-housed in filter-top cages. The animal facility has standard laboratory conditions: 21 to 22°C ambient temperature and a 12 h light/12 h dark cycle. Chow and water were available ad libitum. Both doxycycline (Dox) water and water only (controls), were supplemented with 0.1% saccharin to ensure animals drank the Dox supplemented water. Mice from each group were introduced to the study drinking water 2 days prior to infection. Water bottles were protected from light and replaced every 2–3 days. Mice drank approximately 12.5 mL of water daily. Mice (8 weeks old, n = 6) (12 in total, 3 per study arm) were infected intraperitoneally with 103 SM or TK c-null parasites ± Dox. Prior to the infection TK c-null parasites were propagated in +Tet conditions to ensure parasite viability at the start of the study. Mice were monitored for parasitemia starting three days post-infection by collecting 1 μL of blood from the tail in a 1:150 dilution of medium and counted using a hemocytometer as described [59]. Mice were monitored for 30 days post infection.
Cells (5x107) were harvested by centrifugation (3500 RPM, 5 min), washed twice in cold PBS (10 mM Na2HPO4, 1.8 mM KH2PO4, 137 mM NaCl, 2.7 mM KCl, pH 7.4), and resuspended (50 uL) in trypanosome lysis buffer (50 mM Hepes, pH 8.0, 100 mM NaCl, 5 mM, β-mercaptoethanol, 2 mM PMSF, 1 mg/mL leupeptin, 2 mg/mL antipain, 10 mg/mL benzamidine, 1mg/mL pepstatin, 1 mg/mL chymostatin). Cells were lysed by 3 freeze/thaw cycles. Insoluble debris was pelleted by centrifugation (16,000 x g, 20 min, 4°C) and the soluble fraction was collected. The Bio-Rad Protein Assay reagent was used to determine protein concentration with bovine serum albumin (BSA) used to generate a standard curve. Total protein (20 μg) was resolved by 12% SDS-PAGE and transferred to a PVDF membrane using the Mini Trans-Blot Cell (BioRad). Membranes were blocked by 5% non-fat dry milk in Tris-buffered saline (TBS) (20 mM Tris (pH 7.6), 150 mM NaCl) for 1 h then incubated with a primary antibody in 5% milk and TBS-T (TBS + 0.1% Tween-20) overnight at 4°C. The following dilutions were used for each primary antibody: αFLAG 1:1000 (Rabbit polyclonal-Thermo Fisher), αAU1 1:1000 (Mouse monoclonal – Covance), αHsTK 1:500 (Mouse monoclonal—Thermo Fisher), αTbBiP 1:100,000 (Covance). For detection of the primary antibody, the membrane was incubated in a 1:10,000 dilution of a Protein A-HRP conjugate (Abcam) in TBS-T (5% milk) for 1 hour at room temperature. The membrane was washed 5 x 5min with TBS-T and incubated in SuperSignal West Pico Chemiluminescent Substrate (Thermo Scientific) for 5 minutes. The membrane was visualized by the ImageQuant LAS 4000 (GE).
Graphs were generated in GraphPad Prism version 7.0a for Mac, GraphPad Software, San Diego California USA (www.graphpad.com), and statistical analysis was performed as indicated in the figure legends.
The KEGG (Kyoto Encyclopedia of Genes and Genomes[60, 61]) pyrimidine metabolic pathway highlights two enzymatic reactions (EC 3.1.3.5 and EC 3.1.3.89) that perform the 5'-(deoxy)nucleotidase activity required to produce pyrimidine deoxynucleosides in other organisms. We searched all reviewed UniProtKB [62] entries with these two described enzyme activities, identifying 561 genes with EC 3.1.3.5 activity and 70 genes with EC 3.1.3.89 activity. We sorted the identified genes according to their assigned PFAM domains [63], keeping nine representative sequences from each unique PFAM, which correspond to five different fold groups (Table 1). For stringent identification of T. brucei protein sequences with potential 5' nucleotidase activity, the representative sequences were used as queries to search the NCBI NR database using PSI-BLAST [64] (5 iterations, E-value cutoff 0.001), storing the resulting position-specific scoring matrix as a checkpoint file for re-initiating BLAST against a database of protein sequences from the T. brucei genome (E-value cutoff 1). Identified T. brucei protein sequences were assigned PFAM domains using batch CD search [65], keeping those sequences with PFAM domains described as possessing 5’-nucleotidase activity. To identify all potential T. brucei protein sequences with domains related to those described as having 5'-nucleotidase activity, we queried the T. brucei genome using RPS-BLAST (E-value cutoff 0.006) with a library of sequence profiles downloaded from the conserved domain database (CDD ID in S1 Table) corresponding to each described 5'-nucleotidase PFAM in Table 1. Identified sequences were crosschecked for the presence of the query domain using batch CD search [65] or HHPRED [66, 67], reporting the positive hits using the initial RPS-BLAST E-values in S1 Table. Additional methods details are provided in S1 Text.
The DNA sequence for the T. brucei 5’-nucleotidase (HD-fold, YfbR-like) was obtained from TriTryDB (Tb09.211.2190). The E. coli codon optimized gene was synthesized by GenScript. PCR was used to generate flanking BsaI and XbaI restriction sites that allowed for cloning into the pE-SUMO(KAN) vector (LifeSensors, Malvern, PA) and expression as a N-terminal His6-SUMO fusion protein. The nucleotidase pE-SUMO vector was transformed into BL21 cells. Cells were cultured in 2L LB-KAN (50 μg/mL) media at 37°C until OD600 0.7, then cooled to 16°C and induced by 500 μM IPTG (Isopropyl β-D-1-thiogalactopyranoside) for 16 h. Cells were collected by centrifugation and suspended in buffer A (500 mM NaCl, 50 mM HEPES pH 7.5, 5 mM imidazole, 5% glycerol, 5 mM 2-mercaptoethanol) supplemented with 2 mM PMSF (phenylmethane sulfonyl fluoride) and a protease inhibitor cocktail (1 mg/mL leupeptin, 2 mg/mL antipain, 10 mg/mL benzamidine, 1 mg/mL pepstatin, 1 mg/mL chymostatin). Cells were lysed by cell disruptor and the cell debris removed by centrifugation. Supernatant was applied to a HisTrap HP column (GE Healthcare) and washed with buffer A. The His6-SUMO nucleotidase fusion was eluted using a gradient of 5–45% buffer B (500 mM NaCl, 50 mM HEPES pH 7.5, 5% glycerol, 5 mM 2-mercaptoethanol, 500 mM imidazole). Fractions were analyzed by SDS-PAGE and those containing the T. brucei HD domain 5’-nucelotidase were pooled, concentrated (10 kDa MWCO Millipore) and dialyzed against buffer A. The His6-Sumo tag was removed by overnight incubation at 4°C with His6-tagged Ubiquitin-like-specific protease 1 (ULP1) as described [68]. After incubation the mixture was applied to a second HisTrap HP column and cleaved 5’-nucelotidase was collected in the flow-through, while His6-tagged ULP-1 and impurities were retained on the resin. The flow through was concentrated and protein purity was assessed to be >98% by SDS-PAGE analysis (S14 Fig). This tagless protein was used to collect the data shown in Fig 8. To provide an alternative purification method, His6-Sumo tagged 5’-nucleotidase eluted from the first Ni+2 column was further purified by Gel Filtration chromatography on a SuperDex 200 Prep Grade column (GE Healthcare) using buffer A. The His6-Sumo tagged 5’-nucleotidase showed similar activity (within 2-fold) of the untagged version. Protein concentrations were determined by Bradford Assay (Bio-Rad).
T. brucei 5’-nucleotidase activity was assessed using an endpoint assay for released inorganic phosphate (Pi) using Malachite Green as the detection reagent as described [38, 69]. Briefly, each reaction (160 μL) contained 50 mM HEPES pH 7.5, 0.5 mM CoCl2, 1 mM substrate, and enzyme. The reactions were incubated for 10–30 min (to confirm linearity) at 37°C and then 50 uL was treated with 100 mM EDTA. Malachite green reagent (150 μL) was added to each quenched reaction and incubated at room temperature for 5 min prior to measurement at 650 nm. A range of substrate (0.1–1.0 mM), metal (0.5–10 mM) and enzyme (25–100 nM) concentrations were tested to confirm linear dependence on enzyme concentration and to confirm that the reaction rate versus substrate curve was a saturable process. The production of Pi was measured at 650 nm. Absorbance was converted to μmoles of Pi using a standard curve generated using phosphate standard (Cayman Chemical) ranging from 0–100 μM. For the substrate and metal ion specificity studies shown in Fig 8, assays were run with 100 nM enzyme for 20 min using a substrate concentration of 1 mM. Metal ion concentrations are indicated on the graph. All data were collected in triplicate.
|
10.1371/journal.pgen.1002486 | Precocious Metamorphosis in the Juvenile Hormone–Deficient Mutant of the Silkworm, Bombyx mori | Insect molting and metamorphosis are intricately governed by two hormones, ecdysteroids and juvenile hormones (JHs). JHs prevent precocious metamorphosis and allow the larva to undergo multiple rounds of molting until it attains the proper size for metamorphosis. In the silkworm, Bombyx mori, several “moltinism” mutations have been identified that exhibit variations in the number of larval molts; however, none of them have been characterized molecularly. Here we report the identification and characterization of the gene responsible for the dimolting (mod) mutant that undergoes precocious metamorphosis with fewer larval–larval molts. We show that the mod mutation results in complete loss of JHs in the larval hemolymph and that the mutant phenotype can be rescued by topical application of a JH analog. We performed positional cloning of mod and found a null mutation in the cytochrome P450 gene CYP15C1 in the mod allele. We also demonstrated that CYP15C1 is specifically expressed in the corpus allatum, an endocrine organ that synthesizes and secretes JHs. Furthermore, a biochemical experiment showed that CYP15C1 epoxidizes farnesoic acid to JH acid in a highly stereospecific manner. Precocious metamorphosis of mod larvae was rescued when the wild-type allele of CYP15C1 was expressed in transgenic mod larvae using the GAL4/UAS system. Our data therefore reveal that CYP15C1 is the gene responsible for the mod mutation and is essential for JH biosynthesis. Remarkably, precocious larval–pupal transition in mod larvae does not occur in the first or second instar, suggesting that authentic epoxidized JHs are not essential in very young larvae of B. mori. Our identification of a JH–deficient mutant in this model insect will lead to a greater understanding of the molecular basis of the hormonal control of development and metamorphosis.
| The number of larval instars in insects varies greatly across insect taxa and can even vary at the intraspecific level. However, little is known about how the number of larval instars is fixed in each species or modified by the environment. The silkworm, Bombyx mori, provides a unique bioresource for investigating this question, as there are several “moltinism” strains that exhibit variations in the number of larval molts. The present study describes the first positional cloning of a moltinism gene. We performed genetic and biochemical analyses on the dimolting (mod) mutant, which shows precocious metamorphosis with fewer larval–larval molts. We found that mod is a juvenile hormone (JH)–deficient mutant that is unable to synthesize JH, a hormone that prevents precocious metamorphosis and allows the larvae to undergo multiple rounds of larval–larval molts. This JH–deficient mutation is the first described to date in any insect species and, therefore, the mod strain will serve as a useful model for elucidating the molecular mechanism of JH action. Remarkably, precocious larval–pupal transition in mod larvae does not occur in the first or second instar, suggesting that morphostatic action of JH is not necessary for young larvae of B. mori.
| The number of larval instars in insects varies greatly across insect taxa, and can even vary at the intraspecific level [1], [2], [3]. In general, phylogenetically higher insects tend to have fewer larval instars (three to eight) compared to species from basal lineages, such as Ephemeroptera, Odonata and Plecoptera (more than ten) [1], [2], [3]. In many species, the number of larval instars is affected by genetic and environmental factors, such as temperature, nutritional conditions, photoperiod, humidity, injuries, and sex [1], [2]. The variation in the number of larval instars in the insect lifecycle is generally considered to be an adaptive response to diverse environmental conditions in order to ensure the attainment of a threshold-size for metamorphosis [1], [2], [3], [4].
The silkworm Bombyx mori, a classic model organism for endocrinology, has been reared by humans for thousands of years, and more than 1,000 strains are currently maintained [5], [6], [7]. Among these, several “moltinism” strains have been identified that exhibit variations in the number of larval instars [6], [7]. Silkworms typically have five larval instars, but the moltinism strains vary between three and seven [6], [7]. For example, precocious larval-pupal metamorphosis is observed in the mod (dimolting, chromosome 11–27.4 cM), rt (recessive trimolting, 7–9.0) and M3 (Moltinism, 6–24.1) strains, while extra larval molting is observed in the M5 (Moltinism, 6–24.1) strain [6], [7]. To date, however, none of these loci has been characterized at the molecular level. Given the availability of whole genome data and post-genomic tools in B. mori [8], [9], [10], these strains offer a valuable resource for elucidating the molecular mechanism that underlies plasticity in the number of larval instars.
Here we report the identification and characterization of the gene responsible for the mod mutation that causes precocious larval-pupal metamorphosis in the third or fourth instar [11]. Most mod larvae form larval-pupal intermediates, but some individuals can become miniature moths with normal fertility. Thus, the mod mutant strain can be maintained as homozygous stocks [6], [11], [12]. We demonstrate that the mod locus encodes CYP15C1, a cytochrome P450 involved in the biosynthesis of juvenile hormones (JHs), whose “status quo” action allows the progression of multiple larval-larval molting until the larva attains the required size for metamorphosis [13], [14], [15]. CYP15C1 is specifically expressed in the corpus allatum (CA), an endocrine organ that produces and secretes JHs. Enzymological analysis revealed that CYP15C1 converts farnesoic acid (FA) to JH acid (JHA) in a highly stereospecific manner. We further demonstrated that CYP15C1 plays an indispensable role in JH biosynthesis, and its molecular defect results in the loss of JHs in the hemolymph, thereby causing precocious metamorphosis in the mod strain. Remarkably, precocious larval-pupal transitions in mod larvae always occur after the larval third instar, but not in the first or second instar. Our data provide further evidence supporting the hypothesis that authentic (epoxidized) JHs are essential for the classic “status quo” molting in late larval stages (third and fourth instar), but not in early larval stages (first and second instar) of B. mori [16].
Larvae of standard B. mori strains undergo molting four times and thus have five larval instars; these larvae are conventionally termed “tetramolter” in silkworm genetics. The spontaneous mutant mod was identified in a standard strain [11] and mod larvae undergo precocious metamorphosis in the third (dimolter) or fourth instar (trimolter). First, we obtained a detailed developmental profile of larvae from two batches of the mod strain. All mod larvae underwent precocious metamorphosis in the fourth instar and no individuals reached the fifth instar (Figure 1A and 1B). We plotted the timing of the onset of spinning in the mod larvae (Figure 1C and 1D). Consistent with previous reports [11], [12], we found that spinning occurred at two clearly distinguishable timings: (1) from 56 to 80 h and (2) from 112 to 144 h after the third molt: these larvae were termed early- and late-maturing trimolters, respectively. This segregation in the timing of the onset of spinning was not observed in the standard strain p50T (Figure 1C) or other moltinism strains [11], and thus is a unique characteristic of the mod strain. Importantly, development in almost all early-maturing trimolters was arrested and they remained as larval-pupal intermediates (93.4%, 85/91 larvae); only 3 of the 91 larvae (3.3%) successfully survived to adulthood (Figure 1B). In contrast, the late-maturing trimolters did not show such severe developmental impairment and 88.5% (77/87) became miniature adults (Figure 1B). In the larval-pupal intermediates, we usually observed prothetelic phenotypes such as a mixed pupal cuticle on the exoskeleton of animals having overall a larval appearance (Figure 1A), suggesting that hormonal switching of molting and metamorphosis may be aberrant in the mod strain. Notably, despite their small body size, reproduction in mod moths seemed normal, and their eggs hatched without apparent abnormalities.
In the silkworm, premature metamorphosis can be induced by the loss of or low levels of JH signaling, which can occur due to the surgical removal of the CA [17] or to overexpression of the JH-degrading enzyme [16]. We therefore hypothesized that precocious metamorphosis in the mod strain was caused by the prevention of JH biosynthesis or signaling. To examine this hypothesis, we first determined whether the mod phenotype could be rescued by treatment with methoprene, a JH analogue. We topically applied several doses of methoprene to newly-molted third or fourth instar mod larvae and found that a fourth larval molting was induced by the treatment (Figure 1E). Fifth instar larvae that had undergone fourth larval ecdysis grew normally, began to spin after ∼6 days, and eventually metamorphosed to pupae and adults that were normal and fertile. This result suggests that JH reception and subsequent JH signaling is normal in the mod strain. Therefore, we next compared the JH titers in the hemolymph of third instar larvae of mod and p50T strains at 24 h after molting to the third instar. JHs were extracted from the hemolymph and their methoxyhydrin derivatives were analyzed by liquid chromatography-mass spectrometry (LC-MS). We detected JH I and JH II in the hemolymph of p50T, whereas the JH titer in the hemolymph of the mod strain was below the detectable level (Figure 1F). These results indicate that the mod strain is a JH-deficient mutant in which complete (or almost complete) loss of JH caused precocious metamorphosis.
To identify the gene responsible for the mod locus, we performed positional cloning using backcross 1 progeny (BC1) obtained from crossing females of the mod strain (t011 strain, see http://www.shigen.nig.ac.jp/silkwormbase/index.jsp) with F1 heterozygote males of mod and p50T strains (see Figure S1). We mapped the mod locus within ∼400 kb region on the scaffold Bm_scaf16 (chromosome 11) [8] using 792 BC1 individuals. Twenty-five genes were predicted to be present within this region. Among them, we focused on BGIBMGA011708, a gene encodes a cytochrome P450 monooxygenase. Based on sequence homology and phylogenetic analysis (Figure 2B), the gene was designated as CYP15C1. We found that CYP15C1 shares high homology with the CYP15A1 of the cockroach Diploptera punctata, which is involved in JH biosynthesis in CA of the cockroach [18]. Given that the mod phenotype is a result of the loss of the JH titer (Figure 1F), we speculated that the mod phenotype is due to the loss of function of CYP15C1. To examine this possibility, we first determined the nucleotide sequence of the full-length CYP15C1 cDNA from p50T and mod strains. We identified a 68-bp deletion in the mod allele that introduces a premature stop codon in the coding region of CYP15C1 (Figure 2C–2E). This deletion seemed to produce a functionally null mutation in CYP15C1, since a heme-binding motif, which is essential for enzymatic activities in P450s [19], was eliminated in the mod allele (Figure 2D). This result indicates that CYP15C1 is a strong candidate for the mod locus. Therefore, we further characterized CYP15C1 and its gene product.
The strict regulation of JH biosynthesis in CA is critical for the successful development and reproduction of insects [14], [15], [20]. We next examined the spatial expression pattern of CYP15C1 mRNA. We examined 12 tissues at four different developmental stages and found that CYP15C1 mRNA was highly specific to the corpus cardiacum (CC)-CA complex (Figure 3A). A whole mount in situ hybridization experiment in the brain (Br)-CC-CA complex (Figure 3B and Figure S2) showed that the signal for CYP15C1 was strictly limited to CA, where JH is synthesized, and could not be detected in the brain or CC. These results showed a close spatial correlation between CYP15C1 expression and JH biosynthesis.
Next, we carried out a detailed analysis of the temporal expression pattern of CYP15C1 in the CC-CA complex and compared it to that of the gene for JHA methyltransferase (JHAMT), a key enzyme that acts in the final step of the JH biosynthetic pathway in CA [21]. CYP15C1 mRNA was constitutively expressed in CA from the first instar larval to adult stages (Figure 3D), even when JH is not synthesized (Figure 3C) [20]; no apparent differences in levels of CYP15C1 mRNA were observed between males and females during pupal and adult stages (Figure 3D). In contrast, the temporal expression pattern of JHAMT correlates well with the JH synthetic activity of CA (Figure 3D and Figure S2). JHAMT transcript completely disappeared by day 4 of the fifth instar when CA ceased production of JH (see Figure 3C). It reappeared from the mid-pupal stage and increased to a very high level in the female CA. This was consistent with the temporal profile of JH biosynthesis activity in CA as this occurs only in females during the pupal and adult stages [20]. Taken together, our results strongly indicate that CYP15C1 is involved in JH biosynthesis in CA, but does not appear to act as a rate-limiting factor for JH biosynthesis.
The cockroach CYP15A1, the ortholog of B. mori CYP15C1, catalyzes the epoxidation of (2E,6E)-methyl farnesoate (MF) to JH III [18]. Although biochemical studies predicted the presence of FA epoxidase in the CA of the lepidopteran insect Manduca sexta [22], [23], the corresponding gene has not been identified to date. Therefore we examined the enzymatic activity of B. mori CYP15C1 against two plausible substrates, FA and MF. First, we employed a transient expression system using Drosophila S2 cells. When S2 cells expressing CYP15C1 were incubated with medium containing FA, a major HPLC peak was generated that had the same retention time (15.1 min) as standard JH III acid (JHA III) (Figure 4A, middle). This peak did not appear when S2 cells expressing GFP were used (Figure 4A, bottom). The ESI-MS spectrum of this peak gave an [M-H]− at m/z 251, consistent with the C15H24O3 formula of JHA III, confirming that CYP15C1 catalyzes the conversion of FA to JHA III. The enzymatic properties of CYP15C1 was further examined in a stable Sf9 cell line (Sf9/BmCYP15C1) that constitutively expresses CYP15C1. When the Sf9/BmCYP15C1 cells were cultured in medium containing FA, significant levels of JHA III were detected; in contrast, JHA III production was difficult to detect when original Sf9 cells were used (Table S2, Exp.1). When Sf9/BmCYP15C1 cells were cultured in medium containing MF, JH III generation was detected at low levels. However, a similar level of JH III production was also detected in the original Sf9 cells when they were cultured in the same medium (Table S2, Exp.1). These results suggest that JH III production observed in Sf9/BmCYP15C1 was might be due to the presence of endogenous P450 epoxidases in Sf9 cells, which have been reported previously to have lower substrate specificity and stereospecificity [18], [24]. The addition of the JH esterase inhibitor 3-octylthio-1,1,1- trifluropropan-2-one (OTFP) did not increase production of JHA III (Table S2, Exp. 2), indicating that the degradation of JH III by intrinsic JH esterases in the cells was negligible. Therefore, we were able to estimate the conversion ratio of FA and MF to JH III by CYP15C1. This showed that CYP15C1 exhibited at least 18-fold higher activity for FA than MF (Table S2, Exp. 1), a result that is consistent with previous biochemical studies on lepidopteran FA epoxidase in CA.
To further examine the stereospecificity of CYP15C1, the JHA III generated by Sf9/CYP15C1 was chemically methylated and analyzed by a Chiral-HPLC. The methylated product had a major (R)-JH III and a minor (S)-JH III peak (R∶S = 97∶3) (Figure 4B). These results show that B. mori CYP15C1 encodes a functional P450 epoxidase that preferentially converts FA to JHA III rather than MF to JH III, and does so in a highly (R)-enantioselective manner (Figure 4C).
To obtain direct evidence that CYP15C1 is responsible for the mod mutation, we performed transgenic rescue experiments using the GAL4/UAS system [25]. We generated transgenic silkworm lines carrying the UAS-CYP15C1 transgene with the eye-specific 3xP3-EGFP marker [26]. The UAS-CYP15C1 transgene was driven using a silkworm enhancer trap line ET14 in which GAL4 was strongly expressed in CA (Figure 5A), although weak expression was also detected in peripheral tissues including fat bodies and the midgut [9], [27]. As these lines were generated using the standard Shiro-C (w-1; +mod) strain, we changed the genetic background to w-1/w-1; mod/mod by crossing to the mod strain. The resultant w-1; mod; ET14/+ females were then crossed with w-1; mod; UAS-CYP15C1/+ males to determine whether the mod phenotype could be rescued by CYP15C1 overexpression. We used two independent UAS-CYP15C1 lines with ET-14 (Figure 5B). In both UAS-CYP15C1 lines, CYP15C1 overexpression efficiently prevented precocious metamorphosis and 97.1% of the larvae (34/35 in total) underwent the fourth larval molt to become fifth instar larvae (Figure 5B and 5C). Only one larva (1/35) became a late-maturing trimolter, but neither dimolters nor early-maturing trimolters appeared. This result was in contrast to what was observed in control larvae or larvae carrying either the GAL4 or UAS construct alone: approximately half of the larvae became dimolters and the remainder became trimolters, while no larvae became tetramolters. We also measured the JH titer in the hemolymph (Figure 5D). As expected, the JH titers in control, ET14, and UAS larvae were below the detectable limit. In contrast, we were able to detect JH I and JH II in the hemolymph of mod larvae carrying both ET14 and UAS-CYP15C1 constructs. Taken together, these results provide direct evidence that CYP15C1 is responsible for the mod mutation and is essential for JH biosynthesis.
In this study, we identified and characterized the gene responsible for the mod locus that causes precocious larval-pupal metamorphosis in B. mori. The data we present here have two important implications. First, we provide direct genetic evidence for the significance of P450 epoxidase in the late step of the JH biosynthetic pathway, whose expression is essential for normal growth and metamorphosis. Second, we show that the mod strain is a JH-deficient mutant strain carrying a null allele of CYP15C1, in which developmental abnormalities are mostly limited to larval-pupal transitions and are not observed before the second larval molt.
JH III is the most common JH in many insect orders, although its ethyl-branched homologs (JH I and II) are the major JHs in the order Lepidoptera [22], [28]. Biochemical studies have shown that in the late steps of JH biosynthesis in many insect species, including cockroaches and locusts, FA is first methylated to MF and then epoxidized to JH III in CA [22]. However, the final two steps of JH biosynthesis are reversed in Lepidoptera: ethyl-branched homologs of FA (homo-FAs) are first epoxidized and the resultant JHAs (i.e., JHA I and II) are then methylated to the authentic JHs (i.e., JH I and II) [22]. This study showed that B. mori CYP15C1 epoxidizes FA to JHA III in a highly stereospecific manner. CYP15C1 might also epoxidize MF to JH III, but in a far less efficient manner (Table S2). Given that B. mori JHAMT can methylate both FA and JHAs with similar efficiencies [21], our data clearly demonstrate the major JH biosynthetic pathway in B. mori: homo-FAs are first epoxidized to JHAs by CYP15C1, and then methylated to JHs by JHAMT (Figure 4C and Figure 6A). Interestingly, D. punctata CYP15A1 does not convert FA to JHA III [18]. Thus, the difference in specificity of CYP15 to the substrates FA and MF may determine the order of the final steps of JH biosynthesis in insects.
The expression of most early JH biosynthetic enzyme genes and JHAMT in B. mori is limited to the CA and shows dynamic developmental fluctuations [20], [21], [29]. In particular, the temporal expression profile of JHAMT correlates well with JH biosynthetic activity in B. mori [20], [21], [30], [31] and in the Eri silkworm Samia cynthia ricini [32], indicating that JHAMT is a key regulatory gene whose transcriptional control is critical for the regulation of JH biosynthesis in Lepidoptera. Here, we found that expression of CYP15C1 was also limited in CA but in a different pattern to other JH biosynthesis genes in that it was constitutively expressed from larval to adult stages. This result suggests that the transcriptional regulation of CYP15C1 is less important than JHAMT for the temporal regulation of JH production in B. mori. CA of the silkworm ceases JH biosynthesis by day 3 of the last (fifth) instar [20]; however, it is speculated that CA synthesizes and secretes JHAs during the following prepupal period. Our data indicate that this endocrine switch can be explained by constitutive CYP15C1 expression and the shut-off of JHAMT expression in CA (Figure 6A). During the larval-pupal transition, homo-FAs are constantly converted to JHA I and II by CYP15C1, and the resultant JHAs are secreted from the gland without further conversion because of the absence of JHAMT.
CYP15 P450 family members are found in both hemimetabolous and holometabolous insects [33]. In a similar manner as CYP15C1 expression in B. mori, CA-specific CYP15 expression has also been observed in two cockroach species, D. punctata and Blattella germanica [18], [34], in the locust Schistocerca gregaria [35], and in the mosquito Aedes aegypti [36], suggesting a conserved function in JH biosynthesis. However, the enzymatic properties of CYP15 products, with the exception of those of D. punctata [18] and B. mori (this study), have not been studied and the physiological role of CYP15s in the development of other insects remains unknown. By characterizing the CYP15C1-null mutant silkworm, we have demonstrated here that CYP15C1 plays an essential role in JH biosynthesis and for the maintenance of the proper timing of metamorphosis.
Accumulating data have suggested that CYP15 genes are evolutionarily diversified in terms of their gene regulation and nature. For example, unlike B. mori CYP15C1, A. aegypti CYP15 shows developmentally and dynamically regulated changes of expression, which appear to correlate well with the JH synthetic activity in the CA [36]. In addition, CYP15 is not present in the genome of D. melanogaster, but a P450 gene (Cyp6g2) is expressed in CA in a highly tissue-specific manner [37]. More extensive research on the transcriptional controls and enzymatic properties of JH epoxidases across a broader range of insect taxa will shed light on the roles of these enzymes.
Our results consistently indicate that the mod strain is a JH-deficient mutant that is unable to synthesize JHs in CA. One unique characteristic of the precocious pupation in the mod strain is the variation in the timing of the onset of spinning (Figure 1). The feeding period in early-maturing trimolters was unusually short (50 h after molting) compared with that observed in surgical allatectomy of newly molted fourth instar larvae. In the latter larvae, the feeding period was comparable in length to that of the late-maturing trimolters [e.g. ∼130 h [17]] and no timing segregation was observed [17]. In addition, most of the early-maturing trimolters displayed a larval-pupal intermediate phenotype and eventually died, unlike allatectomized larvae, most of which successfully developed into small but normal pupae [17]. One explanation for this phenomenon is that the early-maturing trimolters were destined to undergo larval molting to the fifth instar on day 2, while the late-maturing trimolters were destined for pupation after a prolonged fourth instar, similar to allatectomized larvae [17] (Figure 1D). Molting in early-maturing trimolters on day 2 usually resulted in the formation of larval-pupal intermediates. One possible explanation for this mixed phenotype is that metamorphosis in the mod strain is induced in the presence of homo-MFs (unepoxidized JH I and II), presumed products instead of epoxidized JH I and II in CA of the mod strain (see Figure 6B). MF is known as a crustacean JH and has recently been reported to have JH activity in D. melanogaster [38], [39]. Therefore, MF and its homologs might have JH-like activity but not able to fully substitute for authentic (epoxidized) JHs in the physiology of the silkworm. Alternatively, other P450 epoxidases in B. mori that have low substrate specificity and stereospecificity, like CYP9E1 [18] and CYP6A1 [24] in other insects, might substitute for the absence of CYP15C1 in peripheral tissues of mod larvae, and such locally-synthesized JHs may prevent precocious metamorphosis in the first and second instar larvae carrying the mod mutation. Further studies are needed to elucidate the mechanism for this unique characteristic of the mod strain.
We found that the precocious phenotype was more severe in the w-1; mod strain compared to that in t011, a genetic stock of the mod strain. We rarely observed dimolter larvae in the t011 stock (Figure 1B). However, in the original manuscript in 1956, it was reported that 28–92% of mod larvae became dimolters [11]. This difference might have developed as a consequence of unintended artificial selection during stock maintenance that favored broods producing trimolters in higher proportions, as it is difficult to obtain sufficient number of eggs using dimolter moths [11], [12]. Thus, we speculate that the present t011 stock may be genetically fixed to produce mostly trimolters, and that this attribute can be varied by outcrossing to other strains.
In the silkworm, premature metamorphosis can be induced by surgical removal of JH-producing CA (allatectomy) [17], by application of an imidazole-based insect growth regulator KK-42 [40] or an anti-juvenile hormone agent KF-13S [41], [42], or by continuous overexpression of the JH-degrading enzyme, JH esterase [16]. In any case, however, premature pupation is not induced in larvae younger than the third instar. In agreement with these studies, we did not observe precocious pupation in first or second instar mod larvae, nor did we observe apparent developmental abnormalities during these early instars. Therefore, our data support the hypothesis that there are two physiological phases in the life of silkworm larvae [16]: the JH-independent phase (first and second instar) in which JH does not have a morphogenetic function; and, the JH-dependent phase (third instar and thereafter) in which the morphostatic action of JH is required to prolong the larval stage until the attainment of the appropriate body size for metamorphosis. Given that most generally the minimum number of the larval instar in insects is three [1], [2], our data further imply that insect larvae need to experience at least one [e.g., L2 pupae in D. melanogaster [43]] or two (e.g., B. mori) larval-larval molts and/or require a certain length of time of postembryonic development in order to acquire competence for metamorphosis.
The silkworm is a classic model organism that has been used for pioneering studies in genetics, physiology, and biochemistry [5]. The availability of whole genome data [8], post-genomic tools [10], and unique mutant resources [6], together with the classic “status quo” responses to JHs in this insect [14], [15], [17], makes the silkworm well-suited for study of hormonal control of growth and development. Indeed, these advantages have greatly contributed to the identification of essential components in the biosynthesis of ecdysteroids, the insect molting hormones [44]. Moreover, recent success in targeted gene disruption using a zinc-finger nuclease [45] increases the utility of this model organism. We are hopeful that our present study will encourage further studies on other “moltinism” strains in the silkworm, and consequently pave the way for a greater understanding of physiological control, developmental plasticity, and evolutionary history of the number of larval molting in insects, which may reflect adaptive strategies of insects to diverse environmental conditions. It is also noteworthy that the late step of the JH biosynthetic pathway is insect-specific and is therefore a potential target for biorational insecticides [46].
Silkworms were reared on an artificial diet or mulberry leaves at 25–27°C under standard conditions as described previously [47]. The silkworm strain t011 (mod/mod) was obtained from Kyushu University [6]. The Spodoptera frugiperda Sf9 and Drosophila melanogaster S2 cells were maintained as described previously [48]. To determine the developmental profile of mod, larvae from two batches of t011 were individually reared in plastic dishes, and their developmental stages were recorded at ∼8-h intervals.
The JH analog, methoprene (a kind gift from S. Sakurai) was applied to newly molted third or fourth instar larvae (∼8–12 h after molting). Methoprene was diluted with acetone and the selected doses (0.01–10 µg/larva) were topically applied to the dorsum using a 10-µl Hamilton microsyringe. The same volume of acetone was applied as a control.
Positional cloning of the mod locus was performed as described previously [49]. Codominant PCR markers and p50T-specific PCR markers were generated for each position of the scaffold Bm_scaf16 (chromosome 11) [9], and used for genetic analysis (Figure 2A and Figure S1). Homozygotes of the mod locus were collected from the BC1 population [t011×(p50×t011)] based on the phenotype of precocious pupation.
Total RNAs were collected from CA of day 0 fifth instar larvae of p50T and Kinshu×Showa strains and used for 5′- and 3′-rapid amplification of cDNA ends (RACE) using the GeneRacer Kit (Invitrogen). PCR was performed using the primers listed in Table S1. The PCR products were subcloned and sequenced as described previously [47]. The obtained cDNA sequence was deposited in the GeneBank (accession number: AB124839).
qRT-PCR was performed essentially as described previously [21]. The primers used for the quantification of the CYP15C1 transcript are listed in Table S1.
In situ hybridization was performed essentially as described previously [50]. A CYP15C1 cDNA fragment (∼1.1 kb) was amplified by PCR listed in Table S1 and subcloned into a pDrive plasmid vector (QIAGEN).
(2E,6E)-farnesoic acid (FA) and (2E,6E)-methyl farnesoate (MF) were purchased from Echelon Research Laboratories (Salt Lake City) and racemic JH III from Sigma. JH III acid was prepared from the racemic JH III as described previously [21]. (R)-JH III was a kind gift from W.G. Goodman.
CYP15C1 overexpression in S2 cells was achieved using a GAL4/UAS system [51]. To generate a vector for expressing CYP15C1 under the control of the UAS promoter (UAS-CYP15C1-HA), a cDNA fragment coding the entire CYP15C1 ORF was ligated into the pUAST vector. UAS-GFP.RN3 [52] was used as a negative control. UAS-CYP15C1-HA or UAS-GFP.RN3 was transfected with the Actin5C-GAL4 construct (a gift from Yasushi Hiromi, National Institute of Genetics, Japan). Forty-eight hours after transfection of S2 cells in a 60-mm dish, the old medium was replaced with 2 ml of fresh medium. S2 cells were detached from the bottom of the dish by pipetting, and 1 ml of the cell suspension was transferred to a siliconized glass test tube. FA or MF (100 µM at final concentration) was then added to the tube. After incubation at 25°C for 16 h, 500 µl of medium was collected and mixed with 500 µl of acetonitrile. Samples were centrifuged for 10 min at 15,000 rpm, followed by incubation at 25°C for 10 min. After filtration using a 0.2 µm filter, 10–20 µl of each sample was subjected to HPLC analysis as described below.
A cDNA with the full ORF of CYP15C1 cDNA was subcloned into the pIZT/V5-His vector (Invitrogen). The plasmid was transfected into Sf9 cells with Cellfectin reagent (Invitrogen), then cells transiently expressing CYP15C1 were selected successively with Zeocin according to the manufacture's instruction and a cell line (Sf9/CYP15C1) stably expressing CYP15C1 was established. Sf9/CYP15C1 cells were placed in a glass tube (12×75 mm) coated with PEG20,000 and cultured in SF900-II SFM medium containing FA or MF (10 µg/ml) for either 2 or 6 h at 26°C. In some experiments, the specific JH esterase-specific inhibitor OTFP (6 µM) was added to the medium to prevent possible degradation of the generated JH III by intrinsic JH esterase present in the cells. After incubation, an equal volume of CH3CN was added to the medium, vortexed vigorously and centrifuged for 4,800 rpm for 10 min to remove cell debris. The supernatant was directly subjected to an HPLC analysis as described below for JH III acid or JH III, which were expected to be generated from FA and MF, respectively.
JH III was analyzed by reversed-phase HPLC as described previously [21]. JH III acid was analyzed by reversed-phase HPLC (column, Shiseido ODS UG80, 150 mm×3.0 mm ID; solvent, CH3CN-20 mM CH3COONH4, pH 5.5, 35∶65, flow rate, 0.5 ml/min; detection, UV 219 nm). ESI-MS spectrum of JH III acid was obtained by TSQ system (Thermo Quest Finnigan, USA).
The stereospecificity of the epoxide group of JH III acid formed by CYP15C1 was analyzed as follows under semi-dark conditions. Sf9/CYP15 cells were cultured in medium containing 10 µg/ml FA for 48 hrs. An equal volume of CH3CN was added to the medium (2 ml), vortexed vigorously and centrifuged at 4,800 rpm for 10 min. One ml of 1 M CH3COONH, (pH 5.5) was added to the supernatant and extracted with 5 ml of CH2CH2; this step was performed 5 times. The extract was dehydrated with anhydrous Na2SO4 and evaporated to dryness in vacuo at 40°C, then the residue was dissolved in 200 µl of CH2Cl2, 50 µl of MeOH and 100 µl of TMS-diazomethane were then added and the solution was incubated at room temperature for 30 min. The reaction was dried with an N2 gas stream, the residue dissolved in 100 µl of hexane, and subjected to a normal-phase HPLC (column, Shiseido SG80, 250×4.6 mm ID; solvent, hexane-EtOH, 99∶1; flow rate, 0.5 ml/min; detection, UV 211 nm). The peak corresponding to JH III (r.t. = 9.8 min) was collected. The stereospecificity of the epoxide group of the JH III was analyzed by a chiral-HPLC (column, Chiralapack IA, 250×4.6 mm ID, DAICEL; solvent, hexane-EtOH, 99∶1; flow rate, 0.5 ml/min; detection UV 219 nm) as described previously [31].
Ten microliters of deuterium-substituted JH III (d3-JH III) [53] in toluene (67.1 pg/ml) was transferred to a clean glass tube to which 0.5 ml of methanol was added. The hemolymph sample (100 µl) was then added and mixed vigorously, and 1.5 ml of 2% NaCl was added to the JH sample. JH was extracted by partition with 0.5 ml hexane; this step was performed three times. The combined solvent containing JH (1.5 ml) was evaporated under a stream of nitrogen. One hundred microliters of methanol and 2 µl trifluoroacetic acid were added to the crude JH extract and mixture heated at 60°C for 30 min. After removal of the methanol, methoxyhydrin derivatives of JH (JH-MHs) were purified using a Pasture pipette packed with 1.0 g of aluminum oxide (activity grade III, ICN Ecochrom) prewashed with hexane. After loading the extract and washing with 2 ml of 30% ether in hexane, JH-MHs were eluted with 2 ml of 50% ethyl acetate in hexane and then dried under a stream of nitrogen. The residue was dissolved in 25 ml of 80% acetonitrile containing 5 µM sodium acetate.
The HP1100MSD system (Agilent) was equipped with a 150×3 mm C18 reversed phase column (UG80, Shiseido) protected by a guard column with 70% acetonitrile containing 5 µM sodium acetate at a flow rate of 0.4 ml/min. For MS analysis, electrospray ionization in the positive mode was used under the conditions of drying gas temperature at 320°C with 10 l/min flow rate, ionization voltage of 70 V. Under these conditions, selected ion masses for each JH-MH were monitored as [M+Na]+, i.e., m/z 321, 324, 335, and 349 for JH III, d3-JH III, JH II, and JH I, respectively.
Overexpression of CYP15C1 was performed in transgenic silkworms using the GAL4/UAS system as described previously [25], [27], [54]. A coding sequence of CYP15C1 was introduced into a silkworm UAS vector carrying the marker gene 3xP3-EGFP. B. mori transformants were established using standard protocols [10]. To overexpress CYP15C1 on the mod/mod background, established UAS lines and an enhancer trap line ET14 [27] were crossed with the t011 strain, and the resultant F1 animals were sib mated to obtain the F2 generation. In the F2 generation, we collected animals showing premature pupation with white eyes (i.e., mod/mod; w-1/w-1) and confirmed the presence of the fluorescent marker gene using a fluorescent microscope (SZX12, Olympus). The established w-1; mod lines carrying UAS-CYP15C1 or ET14 were crossed, and their offspring were examined to determine whether precocious metamorphosis was blocked by CYP15C1 overexpression.
|
10.1371/journal.pcbi.1002309 | Tuning the Mammalian Circadian Clock: Robust Synergy of Two Loops | The circadian clock is accountable for the regulation of internal rhythms in most living organisms. It allows the anticipation of environmental changes during the day and a better adaptation of physiological processes. In mammals the main clock is located in the suprachiasmatic nucleus (SCN) and synchronizes secondary clocks throughout the body. Its molecular constituents form an intracellular network which dictates circadian time and regulates clock-controlled genes. These clock-controlled genes are involved in crucial biological processes including metabolism and cell cycle regulation. Its malfunction can lead to disruption of biological rhythms and cause severe damage to the organism. The detailed mechanisms that govern the circadian system are not yet completely understood. Mathematical models can be of great help to exploit the mechanism of the circadian circuitry. We built a mathematical model for the core clock system using available data on phases and amplitudes of clock components obtained from an extensive literature search. This model was used to answer complex questions for example: how does the degradation rate of Per affect the period of the system and what is the role of the ROR/Bmal/REV-ERB (RBR) loop? Our findings indicate that an increase in the RNA degradation rate of the clock gene Period (Per) can contribute to increase or decrease of the period - a consequence of a non-monotonic effect of Per transcript stability on the circadian period identified by our model. Furthermore, we provide theoretical evidence for a potential role of the RBR loop as an independent oscillator. We carried out overexpression experiments on members of the RBR loop which lead to loss of oscillations consistent with our predictions. These findings challenge the role of the RBR loop as a merely auxiliary loop and might change our view of the clock molecular circuitry and of the function of the nuclear receptors (REV-ERB and ROR) as a putative driving force of molecular oscillations.
| Most organisms have evolved an internal clock which allows them to anticipate and react to the light/dark daily rhythm and is able to generate oscillation with a circa 24 hour rhythm. A molecular network involving feedback loops is responsible for the rhythm generation. A large number of clock-controlled genes pass on time messages and control several biological processes. In spite of its medical importance (role in cancer, sleep disorders, diabetes and others) the mechanism of action of the circadian clock and the role of its constituent's feedback loops remains partially unknown. Using a mathematical model, we were able to bring insight in open circadian biology questions. Firstly, increasing the mRNA degradation rate of Per can contribute to increase or decrease of the period which might explain contradictory experimental findings. Secondly, our data points to a more relevant role of the ROR/Bmal/REV-ERB loop. In particular, that this loop can be an oscillator on its own. We provide experimental evidence that overexpression of members of the ROR/Bmal/REV-ERB lead to loss of Bmal reporter mRNA oscillations. The fact that REV-ERB and ROR are nuclear receptors and therefore important regulators in many cellular processes might have important implications for molecular biology and medicine.
| Circadian rhythms can be found in most organisms, from bacteria to humans and are a fundamental property of living cells [1]. These endogenous rhythms provide a way to anticipate external cues and to adapt molecular and behavioural processes to specific day-times with the advantage of temporally separating incompatible metabolic processes [2].
At the core of the system is the circadian clock, a complex network of genes able to generate stable oscillations with a period of circa 24 hours. The circadian clock has been studied in detail in various organisms such as cyanobacteria [3], Neurospora [4], Arabidopsis [5], Drosophila [6] and mammals [7]. In mammals the main oscillator resides within the suprachiasmatic nucleus (SCN) and is directly entrained by light via the retinohypothalamic tract [8]. This central pacemaker in the SCN is formed by a set of roughly 20.000 neurons which produce rhythmic outputs and orchestrate local clocks in the brain and peripheral clocks throughout the body. Peripheral clocks in the liver, heart, kidney and skin are implicated in the regulation of local transcriptional activity. These can be synchronized also by external cues such as temperature and feeding schedules [9], [10].
Circadian clocks are evolutionarily conserved [6], [11] and designed to maintain an overall optimal organism activity. The internal pacemaker is responsible for the regulation of several biological processes at the cellular level. Such processes include sleep-awake cycles, memory consolidation [12], [13], metabolism of glucose, lipids and drugs [14], [15], bone formation [16], hormone regulation, immunity [17], the timing of cell division cycle and the physiological rhythms such as heart rate, blood pressure and body temperature [18]. Malfunctions of the circadian clock have been reported to be involved in many diseases and disorders such as susceptibility to cancer [19], familial sleep disorders (FASPS) [20], bipolar disorder, sleep problems in the elderly, seasonal affective disorders (SAD) [21], [22], diabetes [23] and obesity [24].
The daily regulation of molecular processes has severe consequences on therapy optimization and timing of drug intake, with the potential of minimizing toxicity and increasing treatment efficacy in complex diseases such as cancer [25]. Therefore, many efforts have been made to identify and understand the molecular circuitry of the clock and its role in disease and therapy [26], [27].
The mammalian molecular clock network is constituted by at least two large interconnected feedback loops which are able to generate approximately 24 hour rhythms [28], [29]. The heterodimer complex, CLOCK/BMAL, formed by the product of the genes circadian locomotor output cycles kaput (Clock) and brain and muscle aryl hydrocarbon receptor nuclear translocator like – Arntl (Bmal) represents the central node in the network and the transcription initiator of the feedback loops.
CLOCK/BMAL binds to E-box cis-elements in the promoter regions of target genes Period homolog 1, 2 and 3 genes (Per1, Per2, Per3), Cryptochrome genes (Cry1, Cry2), retinoic acid-related orphan receptor (Rora, Rorb, Rorc) and Rev-Erb nuclear orphan receptor (Rev-Erbα, Rev-Erbβ) to activate their transcription [7], [30].
The negative PER/CRY (PC) feedback loop is commonly seen as the primary generator of the circadian rhythm [31]. Transcription of Pers and Crys is initiated during the circadian day. Aided by post translational modifications, PER and CRY proteins enter the nucleus, probably as a multimeric complex (PER/CRY) [32], and inhibit CLOCK/BMAL-mediated transcription after a certain delay [31]. The PER/CRY complex is degraded during the night, which releases its inhibitory action on CLOCK/BMAL and allows a new cycle of transcription to take place.
The ROR/Bmal/REV-ERB (RBR) feedback loop is usually seen as adding robustness to the system [31]. Rors and Rev-Erbs are transcribed during the subjective day. Following translation, ROR and REV-ERB proteins compete for ROR regulatory element (RRE) binding sites in the promoter region of Bmal and regulate its transcription. ROR acts as an activator of Bmal and REV-ERB as an inhibitor which results in a fine-tuning of Bmal transcription [33]. Once in the nucleus the BMAL proteins form heterodimer complexes with CLOCK and initiate transcription of target genes (Figure 1).
Minimal models such as the Goodwin oscillator were the first to describe a negative feedback oscillator involving three components [34], [35]. Several kinetic models of the mammalian circadian clock have been subsequently developed [36], [37], [38], [39]. Early models miss essential components such as the nuclear receptor ROR or posttranslational modifications. Other models are rather large and thus the estimation of kinetic parameters becomes exceedingly difficult. Still, many issues regarding the clock remain unknown or not completely understood.
We propose here a single cell model for the mammalian mouse clock of intermediate complexity but containing the most essential biologically relevant processes. Our model allows an independent study of the two loops (PC and RBR). It is biologically comprehensive, emphasizes a parameterization based on biochemical observables, and reflects the current state of research. Although much is known about the circadian clock network, the kinetics of many reactions is not known which makes the parameterization process complex. We have explored known phases and amplitudes among the model components and made use of control theory's principles [40], to obtain estimations for many of the unknown parameters. The resulting model is tested using published data on genome-wide RNAi experiments [41], [42] and transcriptional inhibition data [43].
Our model was applied to address open questions in circadian rhythm biology: firstly, what are the possible reasons for the observed two-loop design? Mathematically, one negative feedback loop with a time delay would be enough to generate stable oscillations. There is evidence from published data showing that overexpression of components of the PC loop does not destroy oscillations [44], [45] which together with remarkable phenotypic effects for members of the RBR loop [46] motivated us to investigate the role of the RBR loop in detail. Secondly, how does degradation kinetics affect the period? We emphasize that such questions cannot be answered intuitively but require quantitative models. The period of the system depends on the timing of gene expression, accumulation and decay, and since clock protein degradation can influence all these processes, intuitive predictions are difficult.
Our simulations show that faster degradation of clock proteins can indeed lead to shorter and longer periods under certain circumstances. In addition, our model predicted that overexpression of members of the RBR loop would lead to damped or even to the loss of oscillations. We could verify these predictions experimentally by constitutively overexpressing Ror and Rev-Erb RNAs in U2OS cells.
Our study represents a step forward towards a fully parameterized model holding significant predictive value. Moreover, this work brings valuable insights into circadian clock biology and helps to understand apparently contradictory results.
We developed a model for the mammalian circadian clock, which allows the study of the two main feedback loops: ROR/Bmal/REV-ERB (RBR) and PER/CRY loop (PC). The model can also be used to study mechanisms critical for the tuning of the circadian system including transcription, translation, import/export, degradation and phosphorylation. We decided to focus on the main pacemaker in the SCN which is assumed to be responsible for the synchronization of the circadian system. Furthermore, the SCN clock might be accountable for general malfunctions and consequent failure of peripheral clocks function, leading to the disruption of normal rhythms [19], [20], [25].
The model was designed based on an extensive literature search and accounts for the available experimental facts (Dataset S1) of the mouse core clock, but is still small enough to allow a systematic parameter determination. For our data collection we gathered available expression data for phases and amplitudes for all the components of the system, regarding the SCN. In order to compare amplitudes of different components found in the literature we normalized the expression level of each component to its mean value. This procedure enables the simulation of expression profiles that oscillate around a base line of 1, for all variables, facilitating the comparison among them. With the developed model we were able to investigate the effects of transcription and degradation on the period of the system and to shed light in a putative role of a two-loop design.
The model contains 19 dynamic variables distributed along two main feedback loops that might be virtually separated (Figure 1, dashed line). Interlocked feedback loops were also reported for Neurospora [47], Drosophila melanogaster [48] and Arabidopsis thaliana [49].
In our model we refer to gene family groups, or gene entities: Per (Per1,2,3) [50]; Cry (Cry1,2) [51]; Ror (Rora,b,c) [52], [53]; Rev-Erb (Rev-Erbα,β) [54], [55]; Bmal (Bmal1,2) [56]. The same principle applies to the proteins and respective protein complexes, represented in Figure 1. The central component, CLOCK/BMAL complex, binds to the promoter regions of clock genes (Rev-Erb, Ror, Per, Cry) activating their transcription [57], [58]. Transcription is controlled by PER/CRY (PER/CRYpool) which possesses an inhibitory effect [59]. In our model the complex represents the pool of all possible PER/CRY complexes present in the nucleus including phosphorylated and unphosphorylated species. We consider in the model the effect of PER/CRY as a transcription inhibitor, regardless of the detailed mechanisms [60], [61].
The circadian core clock network (Figure 1) can be translated into a system of 19 ordinary differential equations (ODEs) with 71 parameters given in Text S1. The system of equations was assembled using mostly the law of mass action [74] and linear degradation kinetics. A Michaelis-Menten degradation kinetics could be used as done in other models allowing smaller Hill coefficients [36], [75]. However, such kinetic laws need more parameters, therefore, we decided to use linear laws in order not to increase the complexity of the model.
Nonlinearities were introduced to describe transcription reactions by means of Michaelis-Menten [76] kinetics and Hill functions [77]. Many parameters could be retrieved from the literature and others were estimated based on known phases and amplitudes using LTI (linear-time-invariant) systems theory (Text S2).
The LTI system theory is often used in electrical engineering, signal processing and control theory. It implies the linearization of the system. Therefore, we created a linear ODE version of the network and applied LTI to our system which allowed a partial determination of the parameters. Linear models allow the analytical calculation of amplitudes and phases as functions of the parameters [40]. Each feedback loop was transformed into a linear open loop system which was then closed, re-establishing the feedback. The parameters were optimized in order to achieve the optimal amplitude and phase-relations. In a subsequent step values for the corresponding parameters of the nonlinear system were determined using a Taylor expansion. After closing the loop the parameters were finally optimized to fine-tune the model.
We based our calculations using key biological assumptions relevant for the mammalian circadian oscillator, such as a period of about 23.5 hours and measured phase/amplitude relations between the components of the model. The parameter estimation procedure is described in detail in Text S2. Still 11 of 71 parameters remain free and their values were adapted to fine-tune the phase and amplitude relations.
The resulting model generates oscillations with a period of 23.5 hours and is able to simulate RNA and protein peaks of expression in the range of the ones found in the literature (Figure 2). The circular graphic shows a comparison between the in silico peaks of expression and the corresponding experimental intervals found in the literature. Represented are 4 mRNA sets (Ror, Rev-Erb, Bmal, Per, Cry) and the nuclear protein complex PER/CRYpool (PER/CRY nuclear pool), covering both parts of the model (RBR loop and PC loop). Bmal mRNA reaches its maximum of expression in the early night. After translation the protein participates in the activation of its target genes in the nucleus. Rev-Erb has its highest expression in the early morning, followed by Ror and Per and finally Cry in the late morning/early afternoon. The heterodimer complex PER/CRY reaches its nuclear expression peak in the late afternoon closing the cycle.
We have tested the predictive capability of the model by comparing results of our simulations with mutation data from knockout mice (Dataset S1) and RNAi data from U-2OS cells [41], [42]. The resulting period of the oscillations was analyzed and given as an output of the simulations. As shown by Brown et al. [78] single cell data might reflect behavioural phenotypes. Thus predictions from our single cell model can be compared with observations from animal mutational phenotypes.
The experimental variability between animal model and cell line data, or even the same system but different publications is higher than the discrepancy between in silico and experimental data (Table 1). This might be due to the fact that the clock system is extremely complex, eventually with more redundancy and further parallel sub-pathways than established so far. It would be conceivable that more feedback loops involving the clock and interconnected networks [79] exist and explain the variability of the phenotypes. Moreover, our control analysis indicates (see Table S1) that variability of some parameters such as degradation rates might be accountable for phenotypic differences between animals and cell lines.
We used the optimized model to analyse the oscillatory potential of each loop as an independent oscillator [80]. Our complete model shows oscillatory expression patterns with a period of 23.5 hours for all components (Figure 3A) and simulates the phase differences and relative amplitudes found in the literature (for comparison see Dataset S1).
Analysing the delays between the different gene species involved in the model (Dataset S1) large delays in the RBR loop can be found. This suggests that this loop could act as an oscillator, also when decoupled from the system. We have therefore hypothesized that the RBR loop should be able to oscillate in the absence of an oscillatory driving force. To test this hypothesis, we replaced the variable PER/CRYpool by its mean value (Text S1, PC = 1.71), creating a constitutive inhibitor. We further wished to analyse the robustness of the model regarding PC and carried out a set of 6 in silico experiments were the PC wild type value is perturbed (PCWT = 1.7) to +/−10%, +/−20%, +/−50% (Supplementary Figure 2). As shown in the figure the oscillations are preserved also under these conditions. This RBR subsystem is a low amplitude oscillator, with a 25.1 hours period (Figure 3B). Interestingly the expression pattern of Ror RNA is almost constant which is consistent with the fact that the inhibitor Rev-Erb might be the driving force of the RBR loop. We aimed to further investigate the independent role of the PC loop. Therefore, we simulated the decoupling of the PC loop by replacing CLOCK/BMAL and REV-ERBN by their mean values (Text S1, x1 = 1.7; x5 = 2.4) generating a constitutive inhibitor and activator respectively. The PC sub-system is a damped oscillator (Figure 3C) with a shorter period (20.7) then the coupled oscillator system.
A negative feedback can induce circadian oscillations if the delay is at least 6 hours [81]. The observed delays between Bmal transcription and its inhibition via REV-ERBN exceed 6 hours (Figure 4). Thus it is conceivable that RBR loop is indeed an oscillator on its own as indicated by our findings.
The nuclear receptors ROR and REV-ERB have been reported to control Bmal expression by competing for RORE elements in the promoter region of the gene and exert an opposite effect on the regulation of Bmal. Our model can simulate the pattern of RORN and REV-ERBN protein expression and its correlation with Bmal RNA expression, thereby illustrating the mechanism of Bmal regulation (Figure 4). The expression curve of both proteins is almost in anti-phase, which was obtained as a result of imposing a specific amplitude and phase for Bmal RNA. From the theoretical standpoint REV-ERB and ROR need to have opposite expression values in order to induce robust transcription of Bmal with a concentration and time for peak of expression according to published data. When the inhibitor REV-ERBN is at its maximum Bmal reaches its minimum expression value and RORN starts increasing its production. Some hours later RORN reaches its maximum and REV-ERBN reaches its minimum level, leading to a peak of Bmal expression. Both REV-ERBN and RORN act antagonistically to enhance Bmal oscillations, which will then regulate the transcription of the genes in the model. These observations are in agreement with experimental findings [33].
Is rhythmic activation of REV-ERB and ROR necessary for Bmal oscillation? To answer this question we carried out further simulations in which we replaced the activator and inhibitor by constitutive ones with corresponding mean value. When REV-ERBN is replaced by its mean value (Text S1, x5 = 2.4) the oscillations of all components in our network are lost. Interestingly, if we increase the concentration of the constitutive inhibitor we recover Bmal oscillations. This could be related to the fact that REV-ERBN acts as an inhibitor of Cry as well. Increasing REV-ERBN induces an inhibitory effect on Cry transcription leading to a decrease of the PER/CRYpool and therefore to a decrease of the inhibition on CLOCK/BMAL. This leads to an increase of RORN and a consequent recovery of Bmal oscillations. If on the other hand, RORN is replaced by its corresponding mean value (Text S1, x6 = 5.8), Bmal still oscillates but with smaller amplitude (data not shown). These results indicate that the amplitude and phase relation between activator and inhibitor is crucial to generate a proper oscillating Bmal with the correct phase and amplitude. Taken together, results from our simulations point to a more important role of the RBR loop on the clock system, than previously assumed.
Published data indicate an influence of the transcription rate on the period of the system [43]. We have addressed this question by perturbing the transcription of each of the five gene entities present in the model and measured the resulting period. A detailed table with all data corresponding to a gradient of the transcription rate from a 10 fold decrease to a 10 fold increase to the wild type is given as Table 3 in Dataset S2.
Analysing the effect of an overall increase in the transcription rate on the system, we observe a direct correlation to the period. As a response to an overall transcription increase, we obtain a longer period revealing a delay of the clock (period measured for the in silico reporter gene Bmal). On the other hand by decreasing the overall transcription rates we obtain a shorter period which accounts for a hastening of circadian oscillations as reported by Dibner et al. [43]. Interestingly, we observe that only perturbations on Cry transcription do not lead to loss of oscillations. The same type of perturbation in the remaining 4 gene entities, leads to loss of oscillations. Moreover, for a decrease of the transcription rate of Per and Cry an increase of the period is observed pointing to their role as inhibitors. The same can be seen when decreasing the rate of transcription for Rev-Erb. This effect is opposite for the activators Ror and Bmal where an increase of the respective transcription rates leads to an increase of the period of the system.
The effects of the degradation of Per on the period are very complex and not yet clarified [82]. This aspect can be exemplified by the following question: Is a faster degradation of a clock element, such as Per2, leading to a shortening or lengthening of the period? Degradation rates are intimately related to the effective delay [65], [83], [84] and consequently one might expect that faster degradation leads to a shorter delay and, subsequently, to a shorter period. This is indeed observed in a cellular model of the FASPS disorder [20]. On the other hand, fast degradation might slow down the nuclear accumulation of the inhibitory PER/CRY complexes leading to a prolonged period. This expectation sounds reasonable as well, thus, intuition alone leads to contradictory predictions and therefore detailed quantitative considerations are required to answer the question raised above.
Mutational phenotypes of Per genes indicate in most cases period shortening or arrhythmic phenotype (Table 1). However, simulation data in Table 1 shows also an increase of the period, with increasing degradation rate. We found these observations remarkable and used the model to find a possible explanation. Our results are quite surprising as can be seen in Figure 5A. We analysed in detail the behaviour of Per when continuously changing its RNA degradation rate (Text S1, dy1) from 0 to 1 (3.3 fold increase of the WT value). Per has a non-monotonic behaviour regarding the degradation rate. This could explain why we see an increase in the period when increasing the degradation rate (Table1) and on the contrary there are published phenotypes showing a decrease.
Figures 5B–5D show simulated time-series of clock genes which can be analysed to understand the underlying mechanisms of non-monotonic period changes. If we choose a value for the degradation rate within the first part of the graphic (marked points B, C) then a decrease of the period with increasing degradation rate would also be seen (Figure 5A). In the second part (points D, E) an increase of the period with the degradation rate is observed. As the degradation rate increases, the system moves from a scenario where the amplitudes of Per and Bmal are small and Per is in phase with Cry (Figure 5B) to another where Per and Bmal amplitudes are larger (Figure 5C). Analyzing the profile of the inhibitor (PER/CRY) it is visible that the shape of the wave varies considerably. The time needed for PER/CRY to reach its inhibitory peak of action (inhibition time, it) and the time needed for it to reach the trough of expression (release time, rt) is different for the 4 points marked (Figure 5A). This might account for the variation in the period. Therefore, we extracted the inhibition times and release times for Figure 5B (it = 11 hours; rt = 16 hours) and Figure 5C (it = 10.5 hours; rt = 15 hours). The values measured (3.4% decrease in it and a 7.4% decrease in rt) together with a sharper peak of the inhibitor complex due to amplitude and phase changes of Per and Cry are correlated with a shorter period (Figure 5C). Due to the earlier phase of Per in Figure 5C compared to Figure 5B the release from inhibition is fastened.
In the second part of the graphic (Figure 5D, 5E) the opposite happens. The period increases with the increase of the degradation rate. The increase in the degradation rate leads to even larger phase shifts between Per and Cry. Following the same methodology, we measured the inhibition and release times for PER/CRY for Figure 5 D, (it = 11.5 hours; rt = 12.1 hours) and for Figure 5E (it = 12,9 hours; rt = 11 hours). These changes correspond to a 10,8% increase of the inhibition time and a 10% decrease of the release time leading to a 1.2% increase of the period. Long inhibition times might be correlated with an increase of the period. This intricate discussion of phase relationships and wave forms helps to understand seemingly counter-intuitive observations. Interestingly, non-monotonic dependencies were found also in much smaller models and with different kinetics than ours [85], [86].
One well studied effect of the degradation on the period of the system is a circadian disorder, FASPS [20], which was the first reported pathology to link known core clock genes to a human disorder. The disease results from a mutation in a casein kinase binding site which affects the phosphorylation of PER and therefore its degradation and results in a circadian oscillator with a shorter period.
We wished to analyse the possible cause for the period shortening knowing that PER's degradation is affected. One possible mechanism could be that FASP mutation reduces the nuclear retention of PER2 but the turnover is not affected [20]. In order to simulate this situation we have increased the export rate of the PER*/CRY nuclear complex to (kex2 = 0.05) and we obtained a decrease in the amplitude of Per and a shorter period, as reported. An alternative situation could be that the turnover of nuclear PER is enhanced, as well as its degradation by the proteasome. This has been previously described as another form of FASPS from the reports on the tau mutation in CKIepsilon [87]. To simulate this hypothesis we have increased the degradation rate for the nuclear protein PER* (dx2 = 0.1). As a result we obtain as well a shortening of the period and a decrease in the amplitude of Per.
These results show that our model can simulate this particular biological problem and is able to illustrate possible alternative scenarios. In this case the model indicates two possible perturbations in the PER* protein, either affecting the degradation rate or the import/export of the phosphorylated protein, both leading to the experimentally observed decrease in the period. With more precise experimental measurements regarding localization and degradation kinetics of Per, the model should be able to discriminate between the two scenarios.
Does a perturbation in the transcription of members of the RBR loop influence the system? To test this property of the upper RBR loop we perturbed the transcription of each gene entity independently and analysed its behaviour and influence on the system. We investigated the robustness of the 4 genes (Rev-Erb, Ror, Per, Cry) represented in the model. In other words we have tested if the oscillations in the expression of these genes are kept in regard to variations in the corresponding transcription rate (Vmax). For each of the genes the transcription rate was varied from 0 to about 3 times the original wild type value, Vmax(WT). This simulates two scenarios: a down regulation of gene expression, for values of Vmax lower than Vmax(WT); an increase in transcription of the RNA, for values of Vmax above Vmax(WT). The results from these simulations are displayed in the form of “rainbow-plots” (Figure 6). These plots provide similar information as bifurcation diagrams [88] with the benefit of allowing the simultaneous visualization of the gene expression dynamics. The rainbow plots allow the detection of sudden qualitative changes in dynamical behaviour of the system, upon small changes of the parameter analysed. We aimed to study the long term effect of the perturbations and hence have simulated 24 days and analysed the last 4 days (Figure 6). For all 4 gene entities there is an optimal parameter range, around Vmax(WT) value which allows the generation of oscillations with the desired phase and amplitude. Furthermore, for Rev-Erb, Ror and Per there is a defined optimal region where the system is able to oscillate, outside this region no oscillations are visible. Interestingly, Cry seems to be very resilient to perturbations. This could be related to the fact that this gene has two inhibitory mechanisms: PER/CRY and REV-ERBN which together could compensate for the increase in transcription levels.
We aimed to explore the role of the RBR loop in more detail and therefore overexpressed in silico both Ror and Rev-Erb. To attain such a perturbation we added a constant RNA, to the endogenous one, for both gene entities independently (Text S1). The constitutive exogenous RNA was taken together with the endogenous one and used for the subsequent protein production. As an output we analysed Bmal patterns of expression. We took care to run our simulations in an in silico set up which would correspond to a real experimental set up, and therefore examined the transient region of the simulations. The result of the simulations for 6 days is given in Figure 7. Our predictions show that Bmal magnitude increases upon Ror constitutive overexpression but the oscillations are lost after 6 days (Figure 7A). Rev-Erb constitutive overexpression leads to a decrease of Bmal magnitude. Similarly to what happens with Ror, the oscillations are also damped and lost after 6 days (Figure 7C).
To experimentally verify our prediction, we have constitutively overexpressed Rora in human osteosarcoma cells harbouring a circadian reporter (Bmal1-promoter-luc) [42], [89] which resulted in loss of oscillations, in agreement with our modelling data (Figure 7B). Higher signals of luciferase activity in Rora overexpressing cells than in GFP controls indicate that overexpression was effective since RORa is a known activator of Bmal1 transcription.
Constitutive overexpression of Rev-Erbα dose-dependently dampened circadian oscillation (Figure 7D). Furthermore, higher levels of Rev-Erbα decreased Bmal1-promoter driven luciferase activity as expected for the transcriptional repressor of Bmal1 [46]. We are aware that a dampening of luminescence signals can reflect both a desynchronization of the cells and a dampening of single cell rhythm. The stronger decay observed after overexpressing both Ror and Rev-Erb (Figure 7b and Figure 7C) is larger than the GFP control. The same can be observed regarding amplitude and magnitude of the oscillations, indicating an effect of the overexpression on the system independent of the desychronization of cells.
We derived and analysed a model for the mammalian circadian clock of intermediate complexity including the most essential cellular processes such as phosphorylation, complex formation, nuclear translocation and transcriptional regulation. Such mathematical models are helpful to gain a quantitative understanding of the dynamical biological system. They aid experimental design and allow the identification of sensitive nodes in the network or the analysis of perturbation effects on the system. Based on available experimental findings on amplitudes and phases (Dataset S1) and using control theory we determined many parameters in a systematic way. However, not all parameters could be calculated in a unique manner. As an example, the phase shift and the amplitude ratio between mRNA and protein allow to specify two parameters but a third one can be varied freely to fine tune the system (Text S2). Moreover, the detailed kinetics of transcriptional activation and inhibition has not been measured. Consequently, our model is consistent with the available data but cannot be regarded as a precise quantitative model of the core clock. Nevertheless, the model can be exploited to gain insight into fundamental questions: How do transcription and degradation control the period? How do feedback loops interact? Is there a more prominent role for the RBR loop than so far described? By posing such questions, simulating and analysing the biological problem, the model can inspire new experiments to test theoretical predictions.
The effect of perturbing the degradation of clock components on the period represents a difficult open question in circadian biology. Due to the complexity of the system results are difficult to predict. We show theoretically that the effect of mRNA degradation of Per on the period is non-monotonic which generates regions of opposite period variation. With a continuous increase of the degradation rate of Per mRNA it is possible to obtain first a decrease in the period and subsequently an increase. Our findings regarding a non-monotonic behaviour of Per can be related and help to explain apparently contradictory reports regarding perturbations on Per and effects on the period.
Indications of opposite change in RNA levels have been found experimentally to induce the same change in the period. Dibner et al. [43] reduced overall transcription and observed a period shortening. On the other hand, Chen et al. [82] reports that, overexpression of Per leads to a period shortening as well. We could simulate both situations with our model (Table 3 in Dataset S2) In these experiments, [43], [82] changes in RNA levels (in both directions) lead to a decrease in the period which resembles the in silico dual behaviour regarding Per. The reported non-monotonic behaviour might as well explain these seemingly opposite experimental results.
Interestingly, Cry seems to be very resilient to perturbations (Figure 6). This could be related to the fact that this gene has two sources of inhibition: PER/CRY and REV-ERBN which together could compensate for the increase in transcription levels. We have tested this hypothesis by removing REV-ERBN inhibition. The new network layout has an effect on Cry amplitude but does not lead to loss of oscillations on Cry. We therefore speculate that this might be related to the fact that Cry, in our system, does not hold posttranslational modifications. These results would be in discrepancy with Ueda et al. [90], where overexpression of Cry leads to loss of oscillation. However in the Ueda study the exogenous Cry was given to the cells introducing additional complexity regarding transcriptional and posttranslational modification. It would be conceivable that CRY protein exhibits posttranslational modifications which would eventually account for the loss of oscillations. On the other hand Fan et al. [44] showed that addition of cell-permeable CRY (CP-CRY) does not lead to loss of oscillations and this biological scenario might be closer to our model and therefore related to our predictions. Further experiments and an extension of the mathematical model to incorporate posttranslational modifications of Cry would be necessary to answer this question, and will be addressed in future work.
The combined activity of the CLOCK/BMAL activator and the P/C inhibitor regulates individual genes with different strengths. Moreover, Bmal itself is fine-tuned by REV-ERB and ROR, allowing the generation of oscillations with the appropriate amplitude and phase. A further fine control of the relative phase relations is subsequently achieved by tuning the degradation rates for each element. This fact raises many questions regarding the role of degradation in the individual control of the concentration and peak of expression of the clock genes. The development of a new model for the mammalian circadian clock (Figure 2) and its fitting to state of the art experimental facts (Dataset S1) rouse our awareness to the importance of the RBR loop. The conventional idea of a single driving core loop might not account for the complexity of the circadian clock and might not be sufficient to explain the redundancy mechanisms reported and the robustness of the system. Our results indicate that the RBR loop might have a more prominent role than previously thought. We present theoretical data that propose the RBR loop as being relevant for the generation of oscillations with appropriate amplitude and phases. The RBR loop can act as an independent oscillator even if we disrupt the oscillations of the lower PC loop (Figure 3B). Moreover, we demonstrate experimentally that the overexpression of elements of this loop (Rev-Erb and Ror) can disrupt oscillations of Bmal mRNA. Additionally, the cross-connection between Rev-Erb and Cry can protect the system from external perturbations of Cry, due to the inhibitory action of REV-ERB.
Our work brings new insight into circadian biology, it points to alternative scenarios able to explain experimental findings. It also raises important questions and might motivate further theoretical and experimental work to explore the RBR loop. The medical consequences of such findings should also not be overseen given that the RBR loop involves nuclear receptors which play a crucial role in hormonal processes and metabolism. Their disruption is connected to many diseases, they can be pharmacologically manipulated by agonists or antagonists and therefore represent an important drug target. Moreover, nuclear receptors bind hormones which could make them key players in synchronization and entrainment of clocks. Elements of the RBR loop might represent the missing link between central and peripheral clocks and could be involved in tissue specific circadian regulation.
The model was designed as a system of 19 ODEs and implemented using Matlab R2010a (Mathworks, Cambridge, UK), with a solver for non-stiff systems (ODE45) which implements a Runge-Kutta method. We have used a relative and absolute tolerance of 10−9, with an integration step of 0.01. The system of equations was assembled using Hill-type kinetics and mass action kinetics (Text S1, Figure S1). Most parameters were derived from the literature or analytically determined using LTI theory (Text S2). The remaining parameters were found by fitting the expression profiles of the variables to published phase and amplitude values.
The rainbow plots in Figures 6 and 7 were produced using Xppaut, version 5.85 (http://www.math.pitt.edu/~bard/xpp/xpp.html) and the Xppaut subsystem Auto.
Lentiviral particles containing hRora, hRev-Erbα or GFP overexpression constructs in a pLenti6 backbone (Invitrogen, Karlsruhe, Germany) were generated as described in published reports [89], [91]. For the high-throughput overexpression analysis of Rora, virus production was performed in 96-well format as described in detail in previous studies [89], [91]. After filtration of the supernatant, U-2 OS cells harbouring a Bmal1-luciferase reporter were transduced in the presence of protamine sulfate (8 µg/ml, Sigma-Aldrich, Hamburg, Germany). Next day, medium was substituted by a blasticidine containing medium (positive selection for 3 days; 10 µg/ml, Invitrogen, Karlsruhe, Germany). For viral transduction of hRev-Erbα and GFP in a larger scale, U-2 OS reporter cells were transduced with 250, 500 or 1500 µl lentiviral containing supernatant including 8 µg/ml protamine sulfate. One day after transduction, cells were selected for 7 days with 10 µg/ml blastidicidin and subsequently seeded into 96-well plates. For online bioluminescence monitoring cells were synchronized by 1 µM dexamethasone (Sigma-Aldrich, Hamburg, Germany). Bioluminescence was recorded for 7 days in a stacker-equipped TopCount luminometer with a sampling rate of about 0.5 hours. Two independent measurements for GFP and hRora (each n = 3) were performed. Note that due to technical variations the first peak shows variable amplitudes. Dose-dependent overexpression of hRev-Erbα and GFP was monitored with an n = 4 per dosage. Raw data were de-trended by dividing a 24 h-running average. Periods and amplitudes were estimated by fitting the cosine wave function via the Chronostar analysis software [92]. For visualization, data were smoothened by a 4 hours-running average. Different basal luciferase levels from raw data were included by the fold change in luciferase activity relative to GFP controls for de-trended and smoothed data. Efficiency of dose-dependent hRev-Erbα overexpression was analyzed via quantitative real-time PCR using QuantiTect primer assays (hRev-Erbα QT00000413 and hGAPDH QT01192646; Qiagen, Düsseldorf, Germany).
|
10.1371/journal.pcbi.0030181 | Organization and Evolution of Primate Centromeric DNA from Whole-Genome Shotgun Sequence Data | The major DNA constituent of primate centromeres is alpha satellite DNA. As much as 2%–5% of sequence generated as part of primate genome sequencing projects consists of this material, which is fragmented or not assembled as part of published genome sequences due to its highly repetitive nature. Here, we develop computational methods to rapidly recover and categorize alpha-satellite sequences from previously uncharacterized whole-genome shotgun sequence data. We present an algorithm to computationally predict potential higher-order array structure based on paired-end sequence data and then experimentally validate its organization and distribution by experimental analyses. Using whole-genome shotgun data from the human, chimpanzee, and macaque genomes, we examine the phylogenetic relationship of these sequences and provide further support for a model for their evolution and mutation over the last 25 million years. Our results confirm fundamental differences in the dispersal and evolution of centromeric satellites in the Old World monkey and ape lineages of evolution.
| Centromeric DNA has been described as the last frontier of genomic sequencing; such regions are typically poorly assembled during the whole-genome shotgun sequence assembly process due to their repetitive complexity. This paper develops a computational algorithm to systematically extract data regarding primate centromeric DNA structure and organization from that ∼5% of sequence that is not included as part of standard genome sequence assemblies. Using this computational approach, we identify and reconstruct published human higher-order alpha satellite arrays and discover new families in human, chimpanzee, and Old World monkeys. Experimental validation confirms the utility of this computational approach to understanding the centromere organization of other nonhuman primates. An evolutionary analysis in diverse primate genomes supports fundamental differences in the structure and organization of centromere DNA between ape and Old World monkey lineages. The ability to extract meaningful biological data from random shotgun sequence data helps to fill an important void in large-scale sequencing of primate genomes, with implications for other genome sequencing projects.
| Alpha-satellite is the only functional DNA sequence associated with all naturally occurring human centromeres. Alpha satellite consists of tandem repetitions of a 171-bp AT-rich sequence motif (called a monomer). In humans, two distinct forms of alpha-satellite are recognized based on their organization and sequence properties. In humans, a large fraction is arranged into higher-order repeat (HOR) arrays (also known as chromosome-specific arrays) where alpha-satellite monomers are organized as multimeric repeat units ranging in size from 3–5 Mb [1]. While individual human alpha satellite monomer units show 20%–40% single-nucleotide variation, the sequence divergence between higher-order repeat units is typically less than 2% [2,3] (Figure 1). The number of multimeric repeats within any centromere varies between different human individuals and, as such, is a source of considerable chromosome length polymorphism. Unequal crossover of satellite DNA between sister chromatid pairs or between homologous chromosomes during meiosis is largely responsible for copy-number differences and is thought to be fundamental in the evolution of these HOR arrays. The organization and unit of periodicity of these arrays are specific to each human chromosome [4,5], with the individual monomer units classified into one of five different suprafamilies based on their sequence properties [5,6]. Interestingly, studies of closely related primates, such as the chimpanzee and orangutan [2,7] indicate that these particular associations do not persist among the centromeres of homologous chromosome, implying that the structure and content of centromeric DNA changes very quickly over relatively short periods of evolutionary time.
In addition to higher-order arrays, large tracts of alpha-satellite DNA have more recently been described that are devoid of any HOR structure [6,8–11]. The individual repeats within these segments show extensive sequence divergence and have been classified as “monomeric” alpha-satellite DNA. Such monomeric tracts are frequently located at the periphery of centromeric DNA [9,11,12]. Consequently, unlike higher-order arrays, some of these regions have been accurately sequenced and assembled because they localize in the transition regions between euchromatin and heterochromatin. Phylogenetic and probabilistic analyses suggest that the higher-order alpha-satellite DNA emerged more recently and displaced existing monomeric repeat sequence as opposed to having arisen by unequal crossing-over of local monomeric DNA [8].
Centromeres and pericentromeric regions are frequently poorly assembled in primate whole-genome sequence assemblies [13–15]. These regions are generally regarded as too difficult to accurately sequence and assemble strictly from whole-genome shotgun (WGS) sequence. However, most WGS sequencing efforts include substantial amounts of alpha-satellite repeat sequence. Indeed, as much as 2%–5% of the sequence generated from the underlying WGS consists of centromeric satellite sequences—such data most often remain as unassembled in public database repositories.
In this study, we develop computational methods to systematically identify and classify alpha-satellite sequences from primate WGS sequence. We predict novel HOR structures from uncharacterized primate genomes and define the phylogenetic relationship of these sequences within the context of known human HOR satellite sequences. Finally, we take advantage of publicly available cloned resources to experimentally validate the dispersal of these newly described alpha-satellite sequences within various primate genomes. The data provide the first genome-wide sequence analysis of alpha-satellite DNA among primates from WGS data and a framework to identify and characterize more repeat-rich, complex regions of genomes as part of genome sequencing projects.
We took advantage of the extensive annotation of human centromeric DNA in the literature to initially construct a non-redundant database of HOR monomeric repeat sequences. We then retrieved WGS sequence data from four primate genomic libraries, identified alpha-satellite monomers using RepeatMasker, and extracted all alpha-satellite repeat units of ∼171 bp in length (Table 1). Our analysis indicated that approximately 1%–5% of all end-sequenced clones generated as part of the WGS libraries represented potential centromeric subclones. Although each library represents only 0.05–0.3 sequence coverage for each genome, human higher-order alpha-satellite arrays are typically 3–5 Mb in length, with hundreds to thousands of copies of each individual unit per chromosome. Consequently, each human HOR unit would be expected to be represented multiple times despite the relatively low coverage of the sequence library.
We compared human WGS alpha-satellite sequences identified within the WIBR2 library to the non-redundant set of HOR sequences by pairwise alignment [16] and Hamming distance [17]. A total of 70% (132 of 188) of human HOR sequences were specifically identified within WGS sequence data (at most 4-bp mismatches), with an average representation of 240 reads per HOR monomer unit. We note that the representation of particular classes was variable and less than the expected number (R2 = 0.13–0.09) as predicted by published minimum and maximum length of each array (Tables S1 and S4, Figure S1). In several cases (e.g., D8Z1, D9Z1, and D16Z1), sequence corresponding to the published HOR arrays was not discovered once within the library (Table 2). We repeated this analysis with additional sources of human WGS sequence and obtained similar results (Tables S1 and S4). The underrepresentation of particular sequences may indicate subcloning biases, variation in copy number, and/or sequence variation between centromeric HOR and published canonical alpha-satellite sequences.
We performed a pairwise analysis of all 135,816 human monomers retrieved from the human WIBR2 library (see Methods). Based on this self-comparison and the sequence similarity to published human HORs, we classified each monomer into one of three categories: (1) those that clustered with our dataset of published higher-order centromeric satellites; (2) those that clustered with each other but did not intersect those in (1); and (3) those that failed to cluster. Since our goal was to recover novel HOR sequences, clusters were established where all members showed at maximum 4-bp differences with any other member in a cluster. This target threshold was chosen because individual alpha-satellite sequences typically exhibit <2% sequence divergence with other paralogous members within a tandem array [18]. By these criteria, 23.3% (31,691 of 135,816) of the recovered monomers clustered with known HORs, with an equivalent proportion (26.2% or 35,499) grouping into 142 HOR clusters not apparently represented in our original dataset. The remaining 68,214 (50%) alpha-satellite monomers represent divergent HOR sequences or putative monomeric alpha-satellite lacking higher-order structure.
WGS sequence reads corresponding to each cluster (type 2, as discussed above) were then retrieved, and each related sequence read was encoded based on its cluster composition (Figure 2). We would expect different monomeric units within different arrays to cluster if they are organized as HOR units. Based on the average read length, a typical WGS read should, then, consist of approximately three distinct HOR monomers. Encoded read compositions were then grouped into larger pattern sets based on a reiterative clustering algorithm. As expected, the pattern set ultimately looped as a result of tandem repetition of the array. We created sequence assemblies (PHRAP; default parameters, -forcelevel=10) [19,20] for all pattern sets that included 30 or more independent WGS sequence reads. A total of 18 distinct sequence contigs were created where the array length (k) ranged from 3–20 subunits.
Each assembled sequence contig was searched against GenBank (nr database) by BLAST (default parameters, p = blastn). We found that 3 of 18 patterns sets corresponded to higher-order alpha-satellite arrays, which had not been included in the original HOR set as part of our literature survey, while another 14 pattern sets showed sequence similarity to other human HOR but were discrepant with respect to published reports either in being more sequence divergent or incomplete with respect to the structure (e.g., D12Z3, D17Z1, D18Z1, etc). In the end, all but one computationally predicted HOR pattern set from the human WGS could be reconciled with published datasets (literature or GenBank).
Our analysis predicted one potentially novel 8-mer HOR unit (HSAHOR8; Table 3) with 92% sequence similarity and 99% query coverage to a clone from Chromosome 22, and only 85% sequence similarity and 94% query coverage to published alpha-satellite sequence D2Z1 (Figure 3). In order to validate its structure, we performed a number of computational and experimental analyses. As a measure of homogeneity, we computed an adjacency statistic that simply calculates the number of times a specified monomer within the WGS sequence read maps adjacently to another specified monomer within the predicted HOR unit (Figure 2). If this repeat were organized as a multimeric tandem array, we would expect encoded monomers to map adjacently at a high frequency. This adjacency statistic for this novel HOR repeat ranged from 97%–100%, indicating considerable homogeneity in the organization of the repeat unit (Figure 3B).
Next, we analyzed mate–pair information associated with the WGS sequence reads. In our model, we would predict that HOR units should be repeated hundreds of times to form a large array of centromeric sequence typically several megabases in length. Consequently, corresponding end sequences from human fosmid clones should both map to the same encoded pattern set even though the two ends are separated by more than 40 kb. For 155 of 156 end-sequence pairs, we observed both the forward and reverse WGS sequences mapping to the same (encoded pattern set) or HOR unit, confirming long-range tandem repeat organization within the clone. As a final test, we performed fluorescence in situ hybridization (FISH) analyses using five different 40-kb fosmid clones representative of this new HOR array, using each as a probe in metaphase hybridizations (Figure 3C). FISH confirmed a typical centromeric HOR pattern, with signals observed on Chromosomes 14 and 22 (Figure 3C) for each of the five probes.
Our initial analysis was biased by triaging alpha-satellite sequences that clustered with known HOR units. As such, we favored accurate reconstruction of these by partitioning the sequence complexity. As a test of de novo alpha-satellite HOR reconstruction, we repeated our computational prediction of new higher-order arrays without excluding repeat units that map to HOR sequence (Table 4). In this blind test, we accurately predicted 12 of 24 known higher-order arrays with more than 92% sequence similarity. If we increase the maximum allowed Hamming distance from 4 to 6, we recover two more arrays with sequence identity greater than 92% (Table 4). This is likely a reflection of underrepresentation of particular classes of HOR sequence within WGS data (Table S1). Although not all classes of human HORs could be recovered, this analysis suggested that the approach could be implemented to discover a subset of previously undescribed HOR structures in uncharacterized genomes.
In an effort to discover novel centromeric HOR units and to compare centromeric DNA in other primate genomes, we repeated our analysis for publicly available chimpanzee, gibbon, and macaque fosmid and bacterial artificial chromosome (BAC) end sequences. We extracted and classified all monomeric alpha-satellite DNA into two groups: monomeric (lacking HOR structure by our criteria) or HOR (evidence for HOR structure within WGS data) (Table 1) for each species. We identified encoded pattern sets in each species and assembled potential higher order repeats (Table 3). Upon analysis of macaque “higher-order” arrays, all potential multimeric repeat units collapsed into a core dimeric repeat structure (see Figure S2). While adjacent monomers showed 30%–45% sequenced divergence, pairwise sequence comparisons of dimeric repeats showed between 2%–5% sequence divergence (Table S5; Kimura 2 parameter). Similar values were obtained based on comparisons between the encoded pattern sets, suggesting considerable homogeneity in the structure and organization of macaque centromeric satellites (as predicted by restriction digest analysis [21].
In contrast, the chimpanzee encoded pattern set showed considerably more diversity in structure, more reminiscent of human centromeric DNA architecture (Table 4). The average chimpanzee paired-end statistic for these pattern sets (37.21%) was similar to accurately predicted HORs in humans, predicting the presence of HORs in chimpanzees. Interestingly, the assembled chimpanzee sequences showed >12% sequence divergence when aligned to human HOR sequences (maximum sequence identity between 78%–88% between human and chimpanzee HORs; Table S3). As a test of our in silico prediction of HOR structure, we retrieved a chimpanzee fosmid clone corresponding to seven of the chimpanzee alpha-satellite HORs. We designed a specific restriction enzyme assay to digest once and only once within the chimpanzee higher-order array (not including the fosmid polylinker multiple-cloning site). Partial and complete restriction enzymatic digestions confirmed the presence of an alpha-satellite HOR structure in all subclones. In six of seven cases, the observed fragment sizes were consistent with that expected based on in silico analyses (Figure 4 and Table 3). Presence of distinct dimeric ladder-sized bands in complete digests suggests a lack of homogeneity or a more degenerate structure in chimp HOR arrays. Similarly, restriction digests of macaque fosmid clones confirmed multiples of the basic dimeric repeat pattern.
As a final test, we selected a fosmid clone representing each of the chimpanzee and macaque HOR units and assessed its chromosomal distribution by metaphase FISH analysis. In humans, it has been shown that centromeric HOR units are grouped into suprafamilies, and that subsets of nonhomologous chromosomes share monomer alpha-satellite sequences from the same suprafamily. Consequently, probes representing a specific HOR unit can cross-hybridize to centromeres from nonhomologous chromosomes under low stringency hybridization conditions. For the chimpanzee HOR, we observed each of the predicted HOR hybridizing to the centromeres of a set of nonhomologous chromosomes (Table 3 and Figure 5A and 5B). Unlike human HORs, we noted several secondary signals mapping to pericentromeric locations on chimpanzee chromosomes. Moreover, even under high-stringency conditions, a single signal to a specific chromosome was seldomly observed. As predicted [2,5–7], hybridization of the chimpanzee probes against human metaphases mapped to the centromeres and pericentromeric regions of nonorthologous chromosomes (Figure S3). We note that not all chimpanzee centromeres were identified in this analysis, indicating that only a fraction of the HORs have been successfully identified. Furthermore, some chromosomes (e.g., Chromosomes 19 and 20) were common to a large number of the probes. Interestingly, even in cases where the FISH patterns appeared virtually identical (PTRHOR 3 and PTRHOR 8), a sequence comparison revealed that the two HORs shared only 78.6% sequence identity, suggesting the presence of two different HOR units on the same chromosome. Fosmids that were used as probes were required to have end sequences matching to the same pattern matching set. We did not FISH those where one end mapped to HOR and the other did not. Such fosmid clones may represent edges of arrays with diverged alpha-satellite.
In contrast to the human and chimpanzee, each probe isolated from the macaque and baboon libraries cross-hybridized equally well to all chromosomes (with the exception of the Y chromosome; Figure 5C and 5D) [21,22]. Reciprocal experiments (where baboon probes were hybridized to macaque, and vice versa) confirmed a long-standing, predominant pancentromeric signal distribution in both species (Figure S3). Despite numerous experiments, no probe could be unambiguously assigned to a specific chromosome in these species. These data suggest fundamental differences in the structure and organization of centromeric DNA between the Old World and great ape primate lineages [2,21,22].
In an effort to assess the evolutionary history of primate alpha-satellite sequence, we examined the phylogenetic relationship between both monomeric and higher-order alpha satellite sequences extracted from primate WGS sequence data. In these analyses, we included all higher-order alpha satellite consensus sequences from human, chimpanzee, and gibbon centromeric regions; dimeric alpha-satellite sequences from macaque and baboon; monomeric alpha satellite sequences from New World monkey [6]; and monomeric alpha-satellite sequence located at the periphery of Chromosome 8 [8]. In light of the large number of sequence taxa of limited length, we performed 100 bootstrap tests for each phylogenetic analysis. Our analysis reveals a tripartite evolutionary relationship among these primate sequences; Old World monkey, ape higher-order, and human monomeric alpha-satellite are each evolutionarily distinct (Figure 6). The data show clear introgression of our predicted chimpanzee HORs, with human suprafamily designations, while our limited survey of gibbon sequences suggest the possibility of a distinct origin from a common set of ape ancestral HOR sequences. The dimeric repeat structure is the fundamental unit of macaque centromeric DNA (Figure 6B). Random sampling, as well as testing of alpha-satellites mapping to encoded pattern sets from the macaque, all show a distinct bifurcation (Figure 6, Figure S4, and unpublished data). Analysis of alpha-satellite sequences identified from random BAC end sequences of the colobus, African green monkey, and baboon confirm that the dimeric repeat structure is common to all Old World monkey species (Figure 6C).
The current model of primate centromere DNA organization has been developed almost exclusively from FISH and restriction enzyme studies of the human genome in the last 25 years [4,5,23]. These efforts required the systematic cloning and sequencing of heterochromatic DNA, frequently from chromosome-specific reagents. Our understanding of the extent of sequence and structural diversity among nonhuman primates is much more limited [2,11,21,22,24–26]. We developed an algorithm to identify, categorize, and reconstruct HOR structures from genome-wide sequence data. In this study, we analyzed more than 1.42 Gb of sequence primarily from three species to identify 265,868 (Table 1) alpha-satellite repeat units corresponding to an estimated 100,000 BAC and fosmid clones. Our results provide a genome-wide perspective on the evolution and structure of these regions and a clone framework for further evolutionary, cytogenetic, and sequence characterization.
We have demonstrated that it is possible to reconstruct known HOR alpha-satellite organization in humans via an algorithm that exploits the multimeric tandem repeat organization and the extensive intrachromosomal sequence homogenization of alpha-satellites. Although many human HOR sequences could be identified (Tables 2 and S1), not all were recovered from analysis of WGS sequence. Although restriction enzyme and subcloning biases are most likely responsible for this, our analysis of different human genome libraries of various insert size, vector type, and subcloning strategies (including WGS from randomly sheared DNA) showed virtually identical biases (Table S1). In addition, not all of those correctly identified as human HORs could be properly assembled into a pattern set that completely corresponded to the known sequence array (Table S2). Due to these limitations, our approach should be viewed as opportunistic at this point, as opposed to comprehensive. Advances in sequencing technology that obviate the need for subcloning may lead to better characterization of centromeric DNA [27].
The most important factor in correctly predicting HOR pattern sets was the Hamming distance choice for clustering of repeats. There is a tradeoff between sensitivity and specificity. A Hamming distance estimate that is too low will fail to cluster related repeats, while increasing the value will lead to overcollapse and a concomitant loss of power to accurately distinguish HOR pattern sets. In humans, we optimally set the Hamming distance to 4 based on paralogous sequence divergence between multimeric units within the human HOR arrays. In a blind study of human WGS sequence, we estimate that approximately 12 of 24 (Table 2) multimeric units can be partially or fully reconstructed at this distance.
The heuristics described to merge pattern sets may also impose problems in HOR array prediction. If there exists two different HOR sets that include monomers of high sequence identity (<2% divergent), the pattern-merging scheme may generate chimeric higher-order structures. For this reason, we only use the HOR structures that are experimentally verified as part of our phylogenetic analysis. In addition, all the HOR structures reconstructed using human WGS reads are either identical to previously published HOR arrays, or validated experimentally. Similarly, all but one computationally predicted HOR structure in the chimpanzee can be experimentally validated.
The availability of paired-end sequence data and corresponding clone reagents provide additional tools for confirmation. Our analysis of human WGS data, for example, identified a previously undescribed HOR sequence structure (HSAHOR8) and corresponding clones for testing. Mate-pair data from human fosmid ends (40-kb inserts) confirm that 99.35% of the pairs map to the same pattern set, confirming tandem reiterations of this multimeric repeat unit. FISH analysis of a corresponding fosmid clone from the library (Figure 3) map the novel higher-order sequence to the primary constriction of Chromosomes 14 and 22. Similarly, analysis of chimpanzee fosmid paired-end sequence data identified seven novel HOR units of various lengths (Table 3), and FISH analysis assigned each of these to specific centromeres on chimpanzee chromosomes (Figure 5A and 5B).
Phylogenetic analyses confirm that human and chimpanzee HOR alpha-satellites share a common origin [23] that is evolutionarily distinct from the flanking peripheral monomeric sequences. Every major human alpha-satellite suprachromosomal family shares homologous sequences with chimpanzee (Figures 6A and S5), despite the fact that they map to nonorthologous chromosomes between the two species (Table 3). A comparison of gibbon alpha-satellites reveals only limited introgression with human–chimpanzee sequence clades. These data suggest that gibbon HORs evolved, in large part, independently from that of the human and chimpanzee. It should be noted however, that the number of gibbon sequences is significantly fewer (Table 1). In addition, the gibbon sequences are derived from a large-insert BAC library where restriction enzyme subcloning biases are thought to be more pronounced. Additional sequencing of the gibbon genome in smaller insert libraries may reveal other, yet unreported sequences and phylogenetic relationships.
Comparisons between ape and Old World monkey alpha-satellite DNA confirm two radically distinct patterns of centromeric organization and chromosome distribution [21,22,25]. Almost all (80% of all monomers at Hamming distance = 10) macaque alpha-satellite sequences are organized around a distinct dimeric repeat structure configuration (Figure 6B). Sampling of different Old World monkey species (including colobus, African green monkey, macaque, and baboon) confirm that the dimeric structure is ancient (15–20 million years old) based on the estimated evolutionary divergence of these species [28]. FISH analysis with either baboon or macaque probes reveal a pancentromeric distribution on metaphase chromosomes (testing of representative clones from each of the ten HOR pattern sets showed no difference; Figure 5). Unlike the great ape higher-order alpha-satellite, HOR structures cannot be assigned to a specific chromosome in these species. These data provide compelling evidence that intrachromosomal homogenization of alpha-satellite DNA has predominated in humans and apes, while transchromosomal exchanges have been the dominant mode among all Old World monkey species.
In summary, we have shown that we can systematically extract evolutionary data regarding centromeric DNA structure and organization from the 2%–5% of WGS sequence data that is typically excluded as part of genome sequencing projects. We provide one of the first genome-wide analyses of centromere structure and evolution from human, chimpanzee, and macaque. Fundamental differences in the structure and organization of centromere DNA between ape and Old World monkey lineages are confirmed [21,22]. The availability of these clone reagents provides a resource for further functional and sequence characterization of primate centromeres and pericentromeric transition regions [29].
We constructed a nonredundant reference set of 254 monomer units from published human higher-order alpha-satellite DNA sequences [6], tracking their suprafamily designation [5,6]. We classified 188 units as canonical human HOR sequence and distinguished an additional 66 as divergent HOR units due to their association with atypical or more divergent centromeric arrays (e.g., Y chromosome and short arm of acrocentric chromosomes). An additional ∼270,000 alpha-satellite monomers were obtained from WGS sequences from various published primate genomic sequencing projects [14,15,30–32]. Sequence and corresponding paired-end sequence annotation was obtained from the National Institutes of Health trace repository (http://www.ncbi.nlm.nih.gov/Traces/trace.cgi) from two human library sources (Fosmid library [WIBR2 ] [31]) and WGS data from Celera [30]) and three nonhuman primate libraries, including chimp (Pan troglodytes) fosmid library (CHORI-1251) [15], rhesus macaque (Macaca mulatta) fosmid library (MQAD) [14], and Northern white-cheeked gibbon (Nomascus leucogenys) BAC genomic library (CH271). A small subsample (300–500 alpha-satellite monomers per species) was obtained from randomly end-sequenced BAC clones from various Old World monkey species, including olive baboon (Papio hamadryas anubis; RPCI-41), vervet monkey (Cercopithecus aethiops; CH252), and black-and-white colobus monkey (Colobus guereza; CH272). We would expect to recover more alpha-satellite sequences from Old World Monkey genomes. However, restriction bias limits subcloning of particular regions, especially in the case of BAC subclones. It is also the likely reason we do not recover all HOR sequences in humans. As a representative of human monomeric DNA lacking higher-order structure, we extracted (360 monomers) from a previously described genomic clone mapping peripherally of higher-order alpha-satellite DNA. We also extracted 71 monomers from another genomic clone mapping peripherally of higher-order alpha-satellite DNA on Chromosome 19 to further validate the phylogenetic relationship of monomeric versus HOR alpha-satellite sequences.
Alpha-satellite DNA sequences were retrieved from WGS data from human, chimpanzee, gibbon, and macaque fosmid and BAC end sequences used as part of genome sequencing projects (Table 1). Reads containing alpha-satellite sequences were initially identified by BLAST sequence similarity searches (p = blastn, v = 10,000), and individual monomer units were extracted using a customized RepeatMasker library [33] with higher-order alpha-satellite consensus sequences in [6] (parameters: -no_is –nolow –lib ‘hor.fa'). We extracted alpha-satellite monomers with the same begin and end positions based on RepeatMasker coordinates [33]. This procedure generated a total of 265,868 alpha-satellite monomer repeat units. For each species, we constructed all possible pairwise alignments for each monomer pair and computed the aligned Hamming distance (defined as the minimum number of substitutions required to change one string into the other) between each pair [17] (not counting indels) as follows: Hamming distance computation is solvable in O(n) time for a pair of sequences of length n. Here, we compute Hamming distance of pairwise alignments; thus, computation of aligned Hamming distance takes O(n2) time for a pair of sequences, and O (k · m · n2) time for m repeat units against k alpha-satellite. To compute the aligned Hamming distance faster, we exploited the fact that the divergence of any pair of alpha-satellite sequences is less than 40%. We first built the multiple sequence alignment of all 188 sequences in the HOR set via Clustal W [34] and used the computed consensus sequence of the alignment as a centroid, where it is aligned pairwise with all WGS repeat units. This step is reminiscent of the “center-star multiple alignment” method described in [35] (pp. 348–350). If any gaps are inserted to the centroid as a result of a pairwise alignment with a WGS repeat unit si, the bases in si that correspond to a gap in the centroid are removed. Thus all the sequences are converted to a new version of the sequence, where all the sequences are of equal length, and the bases that can be optimally aligned to the consensus (therefore conserved in most monomers) are readjusted to the same location within the sequence. As stated above, we only count the number of substitutions during the Hamming distance computation, and indels are not penalized. Any bases inserted in a monomer but not present in the consensus (thus a specific insertion for that monomer) would induce gaps to the other monomer when pairwise alignments are performed. This method of normalizing the sequences precipitates the removal of such bases inserted in a monomer that would not be counted in any case, while aligning the conserved regions (along with substituted bases) to the same coordinates. This ensures that the Hamming distance of any two alignments of repeat units against the centroid would be the same as their aligned Hamming distances. The pairwise alignment of m repeat units and k higher-order consensus sequences with the centroid is completed in O ((m + k) · n2) time, Hamming distances for all pairs of sequences take O (k · m · n) time, and the overall distance computation time is thus reduced to O ((m + k) · n2 + k · m · n). Once the Hamming distance was computed for each pair, we classified monomers into one of three categories: (1) repeat units that have aligned Hamming distance at most four to at least one of the consensus sequences in HOR or divergent HOR unit sets (typical divergence of monomers within an array is <2%; we therefore set the typical Hamming distance to 171 × 2% = 3.41 ≈ 4) subset have aligned Hamming distance of at most four (potential new HOR units); and (3) the remaining repeat units that fail to cluster by this threshold cutoff.
In the human genome, it is usually possible to partition alpha-satellite sequence into blocks of some k monomers (called higher-order alpha-satellite DNA, or HOR where 4 ≤ k ≤ 20). Such patterns can be easily deduced from high-quality sequence using the key string and colorHOR algorithms [36,37]; however, no algorithm has been designed to predict such patterns from unassembled WGS sequence data. We developed a new algorithm, HORdetect, to recognize such motifs from unassembled WGS sequence by a greedy clustering method. Our primary concern at this step is to build clusters of sequences in which the divergence of any pairs of sequences is at most 2%. The following greedy algorithm ensures such a clustering scheme, although it is not perfect and can yield too many numbers of smaller clusters than the optimal number. An optimal clustering that minimizes the number of the generated clusters would be NP-Complete, and most approximation algorithms would still be unfeasible when the large number of input sequences are considered.
As input, we used the set of alpha-satellite repeat units S = {s1, s2, …, sn} and a distance function d (si, sj) that returns aligned Hamming distance of two sequences si, sj.
We generated as output a set of clusters C ={C1, C2, …, Cm}, where the aligned Hamming distance of any pair of sequences in a cluster Ck is at most 4. The greedy clustering method begins with assigning the first sequence in the set to the first cluster. The second sequence is compared with the initial one; if their pairwise Hamming distance is less than 4, it is added to the same cluster. Otherwise, a new cluster is created with the second sequence. This is iterated for all the remaining sequences, requiring that the divergence of any pairs of sequences within a cluster is less than the Hamming distance threshold of 4. Due to this conservative requirement, our aim is to cluster only those sequences that are part of a HOR. We opted to implement a greedy clustering algorithm in order to avoid sorting all pairwise alignment scores in memory. Any algorithm that has to precompute and store all possible pairwise Hamming distances is impractical when a large number of sequences are considered; for 135,816 sequences, such an alignment matrix would require more than 9.2 billion entries (9.2 GB if each entry is implemented as a single byte; 36.8 GB if entries are represented as integers). Furthermore, when a sequence is excluded from a cluster (Hamming distance >4 based on the centroid), that sequence is not compared with the rest of the sequences in the same cluster, thus reducing the computational time.
The algorithm can be formally described as:
1. Set C1 − {si}, and m − 1.
2. i,2 ≤ i ≤ n:
(a) if
where 1 ≤ j ≤ m and d(si, sk) ≤ 4 ; then update Cj − CjÈsi;
(b) otherwise, set m − m + 1 and Cm − {si}..
After the clustering step, all the clusters are assigned a number i = {1…m}. Then, corresponding WGS reads are encoded with the cluster patterns of the repeat units. For example, if a WGS read includes three monomeric repeat units from three different clusters Ck,Cl,Cm, then that read is identified with pattern (k,l,m). WGS reads with the same cluster sets (patterns) are grouped together, and trivial patterns are merged; i.e., patterns (k,l,m) and (l,m,t) are collapsed to (k,l,m,t). The merging process is iterated as long as there are patterns that can be merged. In case of conflicting patterns, i.e., (k,l,m), (l,m,t), and (l,m,z), two separate new pattern sets are constructed as (k,l,m,t) and (k,l,m,z). A schematic representation of our alpha-satellite HOR detection algorithm can be found in Figure 1. Reads with the same pattern sets were then assembled with phrap [19] and consed [20] tools (with default parameters) to generate a consensus sequence contig (GenBank accessions). We validated the new consensus sequences computationally by examining paired-end sequences and adjacency statistics (see text).
FASTA-formatted sequences were obtained corresponding to each of the extracted alpha-satellite monomers, and multiple sequence alignments were constructed using Clustal W (version 1.83) [34]. Due to the large number of sequence taxa, neighbor-joining methods were used to construct unrooted trees (complete deletion parameters, 100 bootstrap iterations). Phylogenetic trees were visualized using HyperTree hyperbolic tree viewer [38].
Fosmid genomic clones corresponding to chimpanzee, human, and macaque HORs were obtained from Children's Hospital Oakland Research Institute (CHORI) or Washington University Genome Sequencing Center (WUGSC). Fosmid insert DNA was purified (1–2 μg) and digested with diagnostic restriction enzymes under partial (0.6 U/30 min) and complete restriction conditions (1 U/1 h). Primate fosmid DNAs were hybridized as FISH probes against metaphase spreads obtained by PHA-stimulated lymphocytes from normal donors and primate metaphase chromosomes (Figures 3 and 5; H. sapiens, P. troglodytes, M. mulatta, and Papio anubis) as previously described [39]. Both high- and low-stringency FISH experiments were performed using the following conditions: high stringency, three washes with 0.1× SSC at a temperature of 60 °C; low stringency, three washes with 50% formamide at 37 °C followed by three washes with 2× SSC at 42 °C. The reported FISH experiments are performed using high stringency.
The GenBank (http://www.ncbi.nlm.nih.gov/GenBank) accession numbers for the structures discussed in this paper are monomeric alpha-satellite DNA on Chromosome 8 (AC026005), monomeric alpha-satellite DNA on Chromosome 19 (AC010523), higher-order repeat sequence D2Z1 (M81229), and a clone from Chromosome 22 (BX294002.19). |
10.1371/journal.ppat.1006774 | The amino-terminus of the hepatitis C virus (HCV) p7 viroporin and its cleavage from glycoprotein E2-p7 precursor determine specific infectivity and secretion levels of HCV particle types | Viroporins are small transmembrane proteins with ion channel activities modulating properties of intracellular membranes that have diverse proviral functions. Hepatitis C virus (HCV) encodes a viroporin, p7, acting during assembly, envelopment and secretion of viral particles (VP). HCV p7 is released from the viral polyprotein through cleavage at E2-p7 and p7-NS2 junctions by signal peptidase, but also exists as an E2p7 precursor, of poorly defined properties. Here, we found that ectopic p7 expression in HCVcc-infected cells reduced secretion of particle-associated E2 glycoproteins. Using biochemical assays, we show that p7 dose-dependently slows down the ER-to-Golgi traffic, leading to intracellular retention of E2, which suggested that timely E2p7 cleavage and p7 liberation are critical events to control E2 levels. By studying HCV mutants with accelerated E2p7 processing, we demonstrate that E2p7 cleavage controls E2 intracellular expression and secretion levels of nucleocapsid-free subviral particles and infectious virions. In addition, our imaging data reveal that, following p7 liberation, the amino-terminus of p7 is exposed towards the cytosol and coordinates the encounter between NS5A and NS2-based assembly sites loaded with E1E2 glycoproteins, which subsequently leads to nucleocapsid envelopment. We identify punctual mutants at p7 membrane interface that, by abrogating NS2/NS5A interaction, are defective for transmission of infectivity owing to decreased secretion of core and RNA and to increased secretion of non/partially-enveloped particles. Altogether, our results indicate that the retarded E2p7 precursor cleavage is essential to regulate the intracellular and secreted levels of E2 through p7-mediated modulation of the cell secretory pathway and to unmask critical novel assembly functions located at p7 amino-terminus.
| Viroporins are small transmembrane viral proteins with ion channel activities modulating properties of intracellular membranes, which impacts several fundamental biological processes such as trafficking, ion fluxes as well as connections and exchanges between organelles. Hepatitis C virus (HCV) encodes a viroporin, p7, acting during assembly, envelopment and secretion of viral particles. HCV p7 is produced by cleavage from the HCV polyprotein but also exists as an E2p7 precursor, of poorly defined properties. In this study, we have explored how the retarded cleavage between E2 glycoprotein and p7 viroporin could regulate their functions associated to virion assembly and/or perturbation of cellular membrane processes. Specifically, we demonstrate that p7 is able to regulate the cell secretory pathway, which induces the intracellular retention of HCV glycoproteins and favors assembly of HCV particles. Our study also identifies a novel assembly function located at p7 amino-terminus that is unmasked through E2p7-regulated processing and that controls the infectivity of different types of released viral particles. Altogether, our results underscore a critical post-translational control of assembly and secretion of HCV particles that governs their specific infectivity.
| Hepatitis C virus (HCV) infection is a major cause of chronic liver diseases worldwide. With 180 million people persistently infected, chronic HCV infection induces liver diseases such as liver cirrhosis and hepatocellular carcinoma. Although new direct antiviral agents are now able to eradicate the virus in most patients, no protective vaccine currently exists against HCV and it remains major challenges in basic, translational and clinical research [1, 2].
HCV is a plus-strand RNA enveloped virus. Its genome is translated as a single polyprotein that is processed by cellular and viral proteases in 10 mature viral proteins [3] consisting of: i) an assembly module (core-E1-E2-p7-NS2) encompassing the capsid protein (core) as well as the E1 and E2 surface glycoproteins that are incorporated in viral particles, and the p7 and NS2 proteins that support virion assembly, and ii) a replication module encompassing the nonstructural proteins NS3, NS4A, NS4B, NS5A and NS5B that are sufficient to support viral RNA replication but that also contribute to virion production through ill-defined processes. As HCV proteins arise from a shared polyprotein, several post-translational modifications control their expression rates within infected cells. In addition, at least three precursors, i.e., tandem proteins with delayed cleavage, are also detected and may implement functions different than their cognate individual proteins. They consist of immature core protein, associated to the D3 trans-membrane peptide, whose removal allows core targeting to lipid droplets (LDs) [4, 5]; NS4B-5A, which promotes the formation of replication vesicles [6]; and E2p7 [7–11], whose properties are explored in this report.
Assembly of viral particles occurs at endoplasmic reticulum (ER)-derived membranes in close proximity to LDs and virus replication complexes [12], with NS2 and p7 being key players in gathering virion components [13–17]. Particularly, NS2 associates with E1E2 glycoproteins and NS3 as well as with NS5A, which interacts with HCV RNA and core [18–20] and promotes genome encapsidation. Virion assembly begins with the formation of a nucleocapsid, formed by core/RNA complex, and is coupled with its envelopment and acquisition of the E1E2 glycoproteins as well as lipids and apolipoproteins [18, 21]. The pathway of secretion of virions remains to be elucidated and may occur through a non-canonical route [22, 23]. Of note, HCV produces different types of particles in addition to infectious virions, including nucleocapsid-free subviral particles [24, 25], E2-containing exosome vesicles [26, 27], naked nucleocapsids [28], and a range of more or less lipidated infectious viral particles [29]; yet, the regulation of their production is poorly understood.
The p7 protein is a small, 63 amino-acid-long protein, consisting of a “hairpin-like” topology involving three helices inducing two trans-membrane segments connected by a hydrophilic, positively-charged cytosolic loop [30–32], though alternative folds and topologies have been proposed [33–35], e.g., with the p7 C-terminus exposed to the cytosol [33]. As it is able to form an ion channel in either hexameric or heptameric form [30, 35, 36] exhibiting a funnel- or flower-like shape [35, 37], it was classified as a viroporin, like M2 of influenza virus [38]. Importantly, p7 is dispensable for replication but essential for both assembly and secretion of infectious particles [11]. First, p7 modulates the formation of NS2 complexes with E2, NS3 and NS5A [9, 13, 14, 16, 39], allowing clustering of assembly components and regulation of early assembly events. Second, p7 allows, in concert with NS2, the regulation of core localization at lipid droplets vs. ER-derived membranes [17], from where viral particles are released in the secretory pathway. Third, p7 modulates the envelopment of nascent virions [40]. Fourth, p7 may regulate the pH of some intracellular compartments, which could be essential for the protection and secretion of infectious particles [41, 42].
A recent study indicated that residues of the first helix of p7 that are predicted to point toward the channel pore are important for assembly [10]. Noteworthy, viroporin ion channel activities modulate properties of intracellular membranes and, thereby, impacts several fundamental biological processes such as trafficking, ion fluxes as well as connections and exchanges between organelles [38, 43]. While several biophysical studies showed that p7 can change ionic gradients in reconstituted membrane assays in vitro [30, 44–47], few reports have addressed the relevance of such properties in cellulo [41, 42, 48, 49], although, by analogy with viroporins from alternative viruses, this may have diverse proviral functions [38]. For example, the 2B viroporin from coxsackievirus modulates calcium homeostasis, which leads to the suppression of apoptotic host cell responses [50]. Likewise, p7 may promote immune evasion by antagonizing the antiviral IFN function [51].
Interestingly, while p7 is released from the viral polyprotein through cleavage at E2-p7 and p7-NS2 junctions by the cellular signal peptidase [7, 52], it also exists in infected cells as an E2p7 precursor of poorly defined properties [7–11]. Intriguingly, virus mutants that exhibit either only E2p7 precursor (i.e., using a point mutation in E2-p7 cleavage site) or, conversely, no E2p7 expression (i.e., using an IRES sequence between E2 and p7) are both impaired for production of infectious particles [9, 14, 39, 53], suggesting that timely liberation of E2 and/or p7 are critical events for assembly/release of HCV particles.
Here, we explored how the retarded cleavage between E2 and p7 could regulate their functions associated to virion assembly and/or perturbation of cellular membrane processes. We demonstrate that p7 is able to regulate the cell secretory pathway, which induces intracellular retention of HCV glycoproteins, and to control release of nucleocapsid-free subviral and infectious viral particles. Specifically, through biochemical, imaging and functional analysis of a series of mutant viruses with modified E2-p7 junction as well as through p7 transcomplementation assays, our data uncover different mechanisms by which p7 regulates the proportion of different types of secreted HCV particles and determines their specific infectivity.
As HCV p7 is released through inefficient E2p7 cleavage, we first sought to address its function by increasing its expression levels in HCV-infected cells. We found that co-expression of individual p7 with JFH1 HCVcc RNAs in Huh7.5 hepatoma cells decreased the levels of extracellular particle- associated E2 proteins, resulting in ca. 3-fold reduced secretion (Fig 1A).
Since some viroporins from alternative viruses alter the canonical secretory pathway [38, 43], we then asked whether p7 could impact the secretion of VSV-G tsO45 (VSV-Gts), a temperature-dependent folding mutant of VSV-G glycoprotein commonly used as model cargo of protein secretion [54]. At 40°C, this protein remains unfolded, resulting in its accumulation in the ER, whereas its folding can be restored at 32°C, which allowed its transfer from the ER to the Golgi and then the plasma membrane (Fig 1C). We transfected in Huh-7.5 cells VSV-Gts with p7 constructs from different HCV strains and using different signal peptide configurations, which resulted in ca. 60% of cells co-expressing both proteins among the transfected cells (Fig 1B). As monitored by flow cytometry analysis, p7 co-expression significantly reduced the kinetics and levels of VSV-Gts cell surface expression at permissive temperature of 32°C (Fig 1C).
Next, to address how p7 alters traffic through the secretory pathway, we measured the resistance of intracellular VSV-Gts to endoH digestion, used as a marker of ER-to-Golgi traffic [54, 55]. While at 0h, all VSV-Gts glycans remained endoH-sensitive, reflecting ER retention at 40°C, they progressively became resistant to endoH cleavage upon 1-3h incubation at 32°C (Fig 1D–1F), underscoring VSV-Gts transfer to the Golgi apparatus. Importantly, p7 co-expression resulted in dose-dependent decrease of the kinetics of VSV-Gts endoH-resistance (Fig 1D–1F), in a manner similar to influenza virus M2 (Fig 1E), indicating that p7 slowed down the rate of VSV-Gts ER-to-Golgi traffic or, alternatively, favored its retention in the ER. We noticed that p7 proteins from different HCV genotypes/strains mediated this effect (Fig 1E) with that of H77 strain appearing less efficient for inhibiting Golgi transfer. We also found that p7 associated to its own signal-peptide, i.e., the E2 amino-terminus (∆E2p7 construct), induced the strongest VSV-Gts ER retention.
We then thought that p7-mediated alteration of the secretory pathway could induce the retention of HCV glycoproteins at ER membranes, which may favor assembly of HCV particles [21]. Thus, we expressed E1E2 glycoproteins, alone or with p7, to promote their secretion in the cell supernatant as subviral particles (SVP), i.e., nucleocapsid-free enveloped particles [24], that peaked at a buoyant density of ca. 1.05–1.06 g/ml (Fig 1G). As detected by E2 immunoblots from the pellets after ultracentrifugation of the cell supernatants (Fig 1H and 1I), we found that p7 co-expression induced up to 60% increase of E2 intracellular expression and concomitant 55% decrease of SVP release (Fig 1H). This resulted in a ca. 2–3 fold reduced E2 secretion (Fig 1I), which, combined with the results of Fig 1A, indicates that p7 can induce the retention of HCV glycoproteins.
By controlling the release of free p7, E2p7 cleavage efficiency may adjust the levels of p7 expression, which could regulate the extent by which p7 slows down the cell secretory pathway and, concomitantly, the traffic and thus secretion of HCV E2 glycoproteins (Fig 1A, 1H and 1I). Hence, we introduced mutations at the E2-p7 junction in Jc1 and/or JFH1 viruses in order to design HCVcc mutants that exhibit increased E2p7 cleavage scores (Fig 2A), as predicted by SignalP method (http://www.cbs.dtu.dk/services/SignalP/). This strategy was preferred over the use of an IRES sequence between E2 and p7, as reported before [9, 10, 14], because it induces a natural, i.e., signal peptidase-mediated liberation of p7 from the HCV polyprotein. First, we inserted linkers of various sizes at the N-terminus of p7 (HAHALp7 and ASGGSp7 viruses), which left p7 sequence unchanged; the former linker consisting of a double HA tag allowing the detection of p7 [56]. Second, we introduced a substitution at position 2 of p7 (p7-L2S). Third, we generated p7 mutants having insertions of a single residue, either a threonine (p7-T2) at position 2 [8] or an alanine (Ap7) at position 1 of p7 (Fig 2A). None of these mutations–termed hereafter p7 ATMI (Amino-Terminus Membrane Interface) mutants, introduced before the first p7 helix (shown as grey box in Fig 2A) [30, 35, 36], are expected to change p7 structure or opening of its channel pore, as shown by molecular modeling (S1 Fig). Of note, we could not identify mutations at E2 carboxy-terminus that accelerated E2p7 cleavage. Finally, we also introduced a control mutation abrogating E2p7 cleavage (E2-A367R; Fig 2A) [9].
Next, we investigated the rate of E2p7 cleavage by treating lysates of Huh-7.5 cells expressing mutant virus RNAs with EndoH to remove E2 glycans, which improved E2 vs. E2p7 electrophoretic separation (Fig 2B and 2C). Except for the E2-A367R mutant that was not cleaved, as expected, all mutants displayed almost complete E2p7 cleavage, which compared to the ca. 40% and 50% uncleaved E2p7 precursor detected with parental Jc1 and JFH1 viruses, respectively. Using a HA-tag antibody to reveal the HAHALp7 protein, we confirmed that the accelerated cleavage detected for the JFH1 HAHALp7 and Jc1 HAHALp7 viruses induced the release of p7 at the expected molecular size with poor if not undetectable E2p7 (i.e., E2HAHALp7) expression (Fig 2D), though such analysis could not be extended to the other p7 ATMI mutants owing to the unavailability of antibodies against native p7. In addition, as previously reported [8, 11], we also detected small amounts of E2p7NS2 precursor for both wt and mutant viruses (Fig 2D; S2A Fig).
Interestingly, when we investigated the infectivity of these mutant viruses, we found that, relative to parental viruses, the p7 ATMI mutant viruses had decreased extracellular infectivity, by ca. 3-fold to over 100-fold (Fig 2E), depending on the p7 modifications and the virus backbones (Fig 2A). Particularly, the Ap7 insertion and the p7-L2S substitution mutants in JFH1 virus induced ca. 3- and 10-fold decreased infectivity, respectively, whereas the p7-T2, ASGGSp7, and HAHALp7 JFH1 insertion mutants exhibited complete loss of infectivity (Fig 2E). Likewise, the Jc1 HAHALp7 mutant virus displayed a 10–20 fold reduced infectivity, as compared to parental virus (Fig 2E). Similar defects were observed for intracellular infectivity (Fig 2E), indicating that these p7 ATMI mutant viruses were not impaired at the stage of secretion of viral particles. Finally, we found that co-expression of wt p7 (though not mutant p7 such as HAHALp7) restored the production of both extracellular and intracellular infectious particles to levels detected for wt viruses (Fig 2F and 2G; S2B Fig). Hence, since all p7 ATMI mutants displayed nearly complete E2p7 cleavage (Fig 2B and 2C) and since HAHALp7 co-expressed with mutant viruses did not restore infectivity (Fig 2G), these results indicated that the integrity of the N-terminal end of p7 itself is crucial for assembly of infectious particles.
Thus, in the subsequent experiments, we compared more particularly the Jc1 HAHALp7 virus to its parental Jc1 counterpart since this mutant retained some infectivity levels, which allowed us to further characterize this novel phenotype; yet, the most salient results described below could be extended to the other p7 ATMI mutant viruses (see supplemental figure set).
Since p7 ATMI mutant viruses displayed augmented E2p7 cleavage rates (Fig 2), we wondered whether they had altered E2 expression levels. As compared to wt viruses, we observed a 2–3 fold increased intracellular expression of total E2 (i.e., free E2 + E2p7) for all mutant viruses exhibiting increased E2p7 cleavage (Fig 3A–3C; S3A Fig). No modification of E2 half-life could be detected for mutant vs. wt virus (S3C Fig). Specifically, taking into account that ca. 40% E2 was detected as E2p7 precursor for wt JC1 virus (Fig 2B and 2C), we estimated that the actual ratio of free E2 expression for mutant vs. wt viruses is ca. 4–5 fold. We also found that core expression was increased by ca. 1.5–2 fold (Fig 3B and 3D; S3B Fig), whereas expression levels of NS2 and NS5A non-structural proteins (Fig 3B and 3D) were unchanged. As similar results were obtained for all p7 ATMI mutant viruses, this indicated that these mutations modulated the expression levels of structural proteins without altering viral replication and/or translation. Importantly, co-expression of either wt p7 or HAHALp7 did not revert E2 expression to wt levels (Fig 3B and 3C), which indicated that increased E2p7 cleavage (rather that p7 N-terminus modification per se) induce up-regulation of E2 glycoproteins.
Next, we hypothesized that the increased expression of structural proteins could be due to a blockage of their secretion, which may also explain the losses of mutant virus infectivity (Fig 2E–2G). Thus, we quantified the total secretion of virion components, i.e., E2 glycoproteins, core and viral RNAs, in the supernatants of cells expressing p7 ATMI mutant virus relative to wt virus (Fig 4). Using immuno-precipitation (IP) assays of cell supernatants with GNA lectins, which bind glycans present on HCV E1E2 glycoproteins [57], we detected a ca. 4–5 fold increased secretion of E2 protein (Fig 4A), which matched the 4–5 fold elevated levels of intracellular free E2 (Fig 3C, combined with Fig 2B and 2C). Although the ratio of extracellular vs. total intracellular E2 expression levels indicated a 2–3 fold difference between wild-type and mutant viruses (Fig 4D), taking into account that 40% of intracellular E2 species were in the form of non-secreted E2p7, we deduced identical ratios of extracellular vs. intracellular free E2 for wt and mutant viruses. Similar results were obtained for the other p7 ATMI mutants (S4A Fig), suggesting that the increased E2 secretion from cells expressing the p7 ATMI mutant viruses is directly linked to the augmented intracellular E2 expression. As co-expression of p7 in trans did not significantly restore E2 expression (Fig 3C) and secretion (Fig 4A and 4D) to wt levels, this indicated that the delayed cleavage of E2p7 is essential for the control of intracellular and extracellular E2 levels.
Strikingly, in contrast to E2, we observed that the p7 ATMI mutant viruses had decreased secretion of both core and viral RNAs in the cell supernatants, by ca. 2–5 fold (Fig 4B and 4C; S4B and S4C Fig). Furthermore, we found that wt p7 restored normal secretion levels of nucleocapsid components (Fig 4B and 4C; S4D Fig) though not those of E2 glycoproteins (Fig 4A). This indicated that viruses harboring p7 ATMI mutations display differential alterations of pathways leading to trafficking and/or secretion of viral glycoproteins vs. nucleocapsid components, i.e., HCV core and RNA. Finally, since mutant p7, e.g., HAHALp7, co-expression did not restore normal secretion levels of core and RNA (Fig 4B and 4C), these results indicated that p7 itself, rather than its cleavage from E2, modulates the secretion of viral nucleocapsids.
Altogether, our results suggest that altered p7 expression, as induced by accelerated cleavage and release from E2 as well as by its N-terminal modification, influence the proportion of secreted virion components. Indeed, relative to core and/or viral RNAs, a 15–25 fold higher expression of HCV glycoproteins was detected in the supernatants of cells infected with Jc1 HAHALp7 virus as compared to wt virus (Fig 4E and 4F).
Interestingly, we found that the p7 ATMI mutant viruses exhibited strongly decreased specific infectivity relative to their content in core protein or viral RNA, from 4–5 fold for Jc1 HAHALp7 virus (Fig 4H and 4I) to over 50-fold for other mutants (S5A and S5B Fig). Likewise, relative to E2 glycoprotein, the specific infectivity of the Jc1 HAHALp7 virus was decreased by ca. 35-fold, as compared to the parental virus (Fig 4G). Importantly, co-expression of wt p7, but not of ATMI mutant p7, restored (though not completely), the specific infectivity of the mutant viruses (Fig 4G–4I; S5C Fig). Altogether, this pointed out to altered ratios of different secreted forms of HCV-derived particles incorporating these components or, alternatively, altered composition of the viral particles themselves.
To demonstrate if the viral components were secreted in particulate forms, we centrifuged the supernatants of infected-cells in conditions allowing sedimentation of particles. We detected in the pellets a ca. 3-fold increase of E2 levels (Fig 5A; S6A Fig) and a 2–3 fold decrease of both core and RNA (Fig 5C) while comparing Jc1 HAHALp7 mutant vs. parental viruses. This augmentation of secreted particle-associated E2 levels could be detected for all p7 ATMI mutant viruses (S6B Fig). Interestingly, we also observed an increased secretion of E1-containing particles (S6A Fig), likely owing to secretion of HCV glycoproteins as E1E2 heterodimers. Furthermore, ectopic expression of wt p7 restored wt levels of particle-associated HCV glycoproteins, core and RNA (Fig 5A and 5C). Since ectopically-expressed ATMI mutant p7 did not rescue the above levels (Fig 5C), this indicated that p7 N-terminus modifications rather than accelerated E2p7 cleavage induced changes in secretion levels of mutant viral particles, perhaps by altering the ratios between SVPs and (infectious) viral particles.
Then, we aimed at characterizing the different types and proportions of secreted particles for wt vs. p7 ATMI mutant viruses.
First, to confirm that E2 detected in the pellets of ultracentrifuged cell supernatants was secreted as particles, we treated the supernatants of virus-expressing cells with Triton-X100 before ultracentrifugation. We found that such treatment decreased the presence of E2 in the pellets for both wt and mutant viruses (Fig 5D; S6C Fig), hence indicating that a substantial part of secreted E2 proteins were in a sediment form, likely vesicular.
Second, since the p7 ATMI mutant viruses secreted higher E2 amounts with poorer infectivity compared to wt viruses (Fig 4G), we quantified the association of their glycoproteins with other virion components, i.e., core and RNA. When we pulled down E2 using GNA lectins, we found a 5–6 fold decreased association of core and RNA with E2 (Fig 6A). Moreover, ectopic expression of wt p7, though not mutant p7, restored, though partially, the association of E2 with viral core and genome (Fig 6A).
Finally, we investigated the degree of envelopment of the secreted core proteins within a lipid bilayer. Hence, we treated the supernatants of virus-expressing cells with proteinase K and subsequently determined the amounts of proteinase K-resistant core, which indicates its full protection by a lipid membrane or, conversely, its secretion as naked or badly enveloped core particles. Interestingly, as compared to their corresponding parental viruses, we detected decreased amounts of membrane-protected core for Jc1 HAHALp7 virus and, more dramatically, for JFH1 HAHALp7 virus (Fig 6B), which correlated with their respective losses of infectivity (Fig 2E–2G). Furthermore, we found that co-expression of wt p7, though not mutant p7, restored the lipid membrane envelopment of their secreted core proteins to almost wt levels (Fig 6B).
Altogether, these results pointed out to a disruption induced by p7 amino-terminus changes of the degree of association between HCV glycoproteins, viral envelopes and nucleocapsids, which could be due to an excess of E2 vs. core and RNA forms secreted independently, such as SVPs vs. naked/partially enveloped core particles, respectively, or, alternatively, to an increased density of E2 glycoproteins on the surface of secreted viral particles.
To investigate this further, we separated virus sub-populations using buoyant density-gradient fractionation. Jc1 and JFH1 HCVcc physical particles had more than 90% of viral RNA and 95% of core protein in fractions of densities of 1.10–1.15 g/ml, with a peak at 1.11 g/ml (Fig 6C; S7A–S7C Fig). As shown before [58–62], core and RNA could also be detected at higher densities (to up to 1.36 g/ml) and at lower densities, until 1.02 g/ml (S7 Fig). Jc1 HCVcc particles had more than 80% of their infectivity in fractions of densities of 1.08–1.13 g/ml, with a peak at 1.11 g/ml (Fig 6C; S7A and S7B Fig), in agreement with recent reports [57, 60–64]. Lastly, we found that E2 glycoproteins were detected in lower density fractions, with ca. 90% in fractions of densities of 1.03–1.08 g/ml and a peak at 1.05–1.06 g/ml (Fig 6C; S7A Fig) representing SVPs (Fig 1G). Importantly, less than 10% of E2 could be detected at densities of 1.08–1.15 g/ml, in which physical and infectious particles were prominent and corresponded to enveloped viral particles.
Interestingly, the density profile of Jc1 HAHALp7 virus was qualitatively similar to that of wt virus (Fig 6C; S7A and S7B Fig) and did not reveal any alteration of the distribution of the different types of particles along the gradient. However, quantitatively, we found the same alterations of the ratios of E2, core, RNA and infectivity in the different fractions for the mutant vs. parental virus (S7A Fig), as compared to unfractionated viral particles (Fig 5A and 5C). Specifically, 2–3 fold augmented E2 levels were detected in the SVP fractions of 1.03–1.08 densities (S7A Fig). Likewise, 2–3 fold reduced core or RNA levels were detected in all fractions whereas infectious titers were decreased by ca. 10-fold in these fractions (S7A Fig). Furthermore, we found that, whatever the density, the co-expression of Jc1 HAHALp7 virus with ectopic p7 restored wt infectivity concomitantly to restoration of the wt levels of core and RNA (S7B Fig).
Finally, we found that the loss of infectivity was accentuated for the JFH1 HAHALp7 virus, which displays a stronger phenotype than the Jc1 HAHALp7 virus (Fig 2E–2G), in fractions of densities of 1.08–1.15 g/ml containing most viral particles (Fig 6C; S7C Fig). Particularly, while core levels were reduced by ca. 2–3 fold for both Jc1 HAHALp7 and JFH1 HAHALp7, 6- and over 100-fold reductions of infectivity levels were detected for the Jc1 HAHALp7 and JFH1 HAHALp7 mutant viruses, respectively, relative to parental viruses (Fig 6D).
Altogether, these results suggested that the p7 N-terminus controls both the secretion and the specific infectivity of secreted virus particles, likely via a process involving the completion of their envelopment, as indicated by the PK-sensitivity of core from secreted p7 ATMI mutant particles.
We then aimed at dissecting how p7 modulates the composition of viral particles by addressing HCV assembly mechanisms, from intracellular clustering of virion components to their envelopment.
First, since the HCV virion assembly rate is linked to core subcellular localization [17, 65–67] and since p7 alters this event in concert with NS2 [17, 40], we investigated by confocal microscopy analysis whether core from p7 ATMI mutant viruses could be relocated from lipid droplets to ER membranes, where envelopment and release of viral particles occur [3, 21]. While individually expressed Jc1 core had predominant distribution around lipid droplets, as previously described [17], its co-expression with either wt p7 or HAHALp7 protein induced full targeting at ER membranes (S8A Fig). Similar results were obtained with full-length viruses, since no difference between Jc1 and Jc1 HAHALp7 viruses could be detected regarding the prevalent ER-localization of their core proteins (S8B Fig). These data indicated that the p7 ATMI mutations and/or the accelerated E2p7 cleavage did not prevent p7-mediated early assembly events leading to core targeting to the ER membrane, but rather, impaired later assembly events.
Using HA-tag antibodies, we confirmed that HAHALp7 and core co-localized at the ER in cells infected with the Jc1 HAHALp7 virus (Fig 7A), as reported before [56]. Interestingly, we found that HAHALp7 and core co-localization could also be detected in infected cells treated with 5μg/ml digitonin (Fig 7A and 7C), which permeabilizes plasma but not ER membranes [68, 69]. Similar results were also obtained with JFH1 HAHALp7 virus as well as with HAHALp7 expressed individually, with or without signal peptide (S9A–S9C Fig). Since E2/core co-localization could be detected in Triton-treated cells but not in digitonin-treated cells (Fig 7B and 7C; S9A–S9C Fig), as expected owing to the luminal exposition of E2 ectodomain, these results indicated that the N-terminus of HAHALp7 points towards the cytosol. Note that these results do not exclude that p7 may also adopt the reverse topology, i.e., with N- and C-termini exposed toward the luminal side of the ER [70].
We then investigated the sites of HCV assembly, which are represented by ER-derived areas where structural and non-structural viral proteins co-cluster with HCV RNA [66]. As compared to parental virus, we did not find significantly altered clustering of core, E2, NS4B, and NS5A for the Jc1 HAHALp7 virus (Fig 8A and 8B), suggesting identical rates of early assembly events. Likewise, we did not observe strong differences in the number of core structures co-localizing at assembly sites with HCV positive strand RNA for wt vs. mutant viruses (Fig 8C and 8D). Altogether, these results indicated that the initiation of early assembly events, i.e., allowing clustering of the HCV structural components at the ER membrane, were not impaired by p7 ATMI mutations.
Next, since the NS2 non-structural protein is thought to serve as a scaffold for gathering virion assembly components through its interaction with E1E2 [9, 13–16, 71] and since p7 regulates this association [9, 14, 39], we tested if our p7 ATMI mutants had impaired E1E2/NS2 association. Using confocal microscopy, we did not detect altered co-localization of core, E2 and NS2 (Fig 9A and 9B), further highlighting that the assembly sites of viral particles were not grossly changed. However, when we analyzed E1 and NS2 co-immuno-precipitation with E2 antibodies, although the E1/E2 association was unchanged (Fig 9C), we detected a 3–4 fold decreased association of E2 with NS2 for p7 ATMI mutant vs. wt viruses (Fig 9E; S10A Fig). These results indicated that, relative to wt virus, the higher amounts of intracellular HCV glycoproteins detected for the p7 ATMI mutant viruses (Fig 3B and 3C; S3 Fig; S7 Fig) were not associated to NS2. This implied that only a fraction of the up-regulated E2 levels interact with the NS2 assembly platform, and suggested that part of the pool of HCV glycoproteins not associated to NS2 or to assembly sites induce the formation of SVPs. Thus, we performed the reverse co-immuno-precipitation with NS2 antibodies to more directly address interactions between assembly proteins at assembly sites. Strikingly, when we analyzed E1 and E2 co-immuno-precipitation with NS2, we found ca. 2–3 fold increased E1E2 association to NS2 in cells expressing Jc1 (Fig 9F) or JFH1 (Fig 9D; S10B Fig) p7 ATMI mutant viruses, as compared to parental viruses. Furthermore, these altered interactions with NS2 were reversed upon ectopic expression of wt p7 (Fig 9D; S10B Fig). Thus, since ectopically-expressed p7 did not restore wt intracellular E2 expression (Fig 3C) and since NS2 expression was unchanged for the mutant virus compared to wt (Fig 3D), this indicated that p7 N-terminus modulates NS2 association with HCV glycoproteins.
Finally, since NS5A interacts with HCV RNA [72, 73] and core [18, 19] as well as with NS2 [13, 15], which likely transfers core and HCV RNA to assembly sites or to nascent viral particles [18–20], we investigated NS2 association with NS5A. No significantly altered co-localization of core, E2, NS2 and NS5A could be detected while comparing mutant vs. parental viruses (Fig 10A–10D), again underscoring the proximity of these different factors at assembly sites that appeared unaltered qualitatively. However, as compared to parental virus, we found a reduced NS5A co-immuno-precipitation with NS2 in cells expressing the Jc1 HAHALp7 or other p7 ATMI mutant viruses (Fig 10E; S9C Fig), which was restored upon co-expression with wt p7 (Fig 10E).
Altogether, these results suggested that the p7 amino-terminus determines the fine-tuning of the interactions between HCV glycoproteins, NS5A and NS2 required for envelopment of viral particles at assembly platforms.
We report here novel functions of the HCV p7 viroporin, which appears to modulate i) the cell secretory pathway, ii) the assembly and proportion of different secreted HCV-derived particle forms, including SVPs, partially enveloped core particles and infectious virions, and iii) the specific infectivity of the latter type of particles. Overall, our results provide novel insights in the properties of p7, particularly regarding the role of its junction segment with E2 and its retarded cleavage from E2. This distinguishes different functions of p7 between those that only depend on delayed E2p7 cleavage from those that reveal the role played by the p7 amino-terminus itself in envelopment and production of infectious viral particles.
Because of their intrinsic capacity to be routed to the cell surface, owing to their localization in the cell secretory pathway, the traffic and distribution of envelope glycoproteins need to be controlled, by e.g., retention in specific intracellular compartments, in order to avoid immune detection of infected cells. Moreover, as these glycoproteins are synthesized in the ER lumen, counteracting ER stress activation in infected cells is important to avoid subsequent cell death. Since HCV proteins are expressed from a polyprotein, implying that all proteins are initially expressed at the same rate, its well-ordered cleavage may regulate the stability and functions of some proteins, such as for E2 and p7 that coexist with a E2p7 precursor (Fig 2) of ill-defined functions [7–11]. Here, by designing mutants at E2-p7 junction, we show that the augmentation of E2p7 cleavage, which is mediated by signal peptidase [7, 52], induced an up-regulation of the levels of E2 in infected cells, by ca. 4–5 fold.
This original phenotype is not caused by increased rates of replication or translation of such mutant viruses, judging from similar levels of viral RNAs or of non-structural proteins, respectively, for mutant vs. wt viruses. That both short peptide extensions and single alanine insertion before p7 structure induced E2p7 increased cleavage and E2 up-regulation at similar levels (Figs 2C and 3C; S3A Fig) argued against the possibility that p7 N-terminal modifications could per se change E2 expression. Consistently, co-expression of wt p7 with these mutant viruses did not restore wt E2 intracellular levels (Fig 3C).
Different possibilities may explain how E2p7 processing could modulate E2 expression. On the one hand, liberated E2 could be stabilized by its partners, such as e.g., E1 [74, 75] or SPCS1 [71]. On the other hand, E2 and E2p7 could activate or block different degradation pathways, such as autophagy or proteasome/lysosome. Indeed, HCV glycoproteins are known to be activators of the unfolded protein response (UPR) in HCV-infected cells [76]. Finally, it is also possible that the amounts of free p7, which are likely higher for the p7 ATMI mutant viruses, may regulate these degradation pathways. In support of these assumptions, a previous study [77] indicated that a mutation of p7 reported to block cleavage between E2 and p7 (p7-R(K)GR33-35AAA) [9, 11, 13] induced E2 degradation. Likewise, mutants abrogating E2p7 cleavage, such as E2-A367R (Fig 2) as well as p7-A1W, p7-E3W, p7-K4W, p7-A10W or p7-S12W [10], exhibited poor E2p7 expression despite wt rates of replication, relative to parental viruses. Further studies will be necessary to clarify this issue.
Regulation of the intracellular quantities of surface glycoproteins as well as their recruitment at virion assembly sites is crucial for production of infectious particles, which require optimal E1E2 incorporation levels to mediate entry into cells. Several cellular factors promoting the different steps of HCV assembly have been identified [3] and include factors allowing initial core and NS5A targeting at the LDs [78–82], HCV particle assembly [71, 83], or fission of enveloped nucleocapsids [84, 85]. As for viral factors, NS2 gathers assembly components at ER-localized sites near LDs and replication complexes [13, 15, 16, 86], at detergent-resistant membranes (DRM) areas [9, 39]. Specifically, NS2 interaction with E2 and E1 as well as with other viral factors—p7, NS3 and NS5A, is believed to be key for virion biogenesis [9, 13–15, 71, 84]. Yet, independent of this capture mechanism, E1E2 glycoproteins have the intrinsic capacity to induce SVP formation (Fig 1) [24], which implies a competition for their recruitment at the assembly sites of infectious particles. Noteworthy, E2/NS2 interaction depends on SPCS1, one of the 5 subunits of signal peptidase, and abrogation of E2/NS2/SPCS1 triple interaction via SPCS1 silencing markedly reduced HCV assembly [71].
Our co-IP assays indicated that increasing E2p7 cleavage, concomitantly with augmented E2 expression and SVP formation, stimulated the interaction between E1E2 and NS2 (Fig 9D and 9F; S10B Fig). A first, simple possibility to explain this stronger E1E2/NS2 association may involve the increase of E2 intracellular expression, which, incidentally, would lead to greater opportunities for E2/NS2 association. A second possibility could involve either a concentration of SPCS1 at the vicinity of E2 and NS2, as a result of SPCS1 recruitment by the signal peptidase complex during E2p7 and p7NS2 processing, or, alternatively, of a preferential interaction of cleaved, liberated E2 with SPCS1 and hence, with NS2. However, both possibilities would be difficult to reconcile with the finding that wt p7 co-expression, which did not restore wt E2 intracellular levels (Fig 3C), restored normal levels of E1E2/NS2 interaction (Fig 9D and 9F; S10B Fig). A third possibility is that the alteration of p7 N-terminus in our E2-p7 junction mutants may affect NS2 capacity to interact with some of its other partners, as discussed below. Unexpectedly, our results indicated that the increased E1E2/NS2 interaction correlated with reduced formation of infectious particles, in agreement with lowered secretion of nucleocapsids (Fig 5C). In this respect, it is likely that a loss of viral particle formation would translate in an increase of E2 density on NS2 platforms (Fig 9D and 9F) because E2 would not be consummated in assembled and released virions.
Our results indicate that the p7 N-terminus also determines HCV infectivity by controlling the secretion of enveloped vs. naked/partially enveloped core particles (Fig 6B and 6D). This is in agreement with a previous report showing that p7 regulates the envelopment of nascent viral particles [40]. A recent study indicated that the first helix of p7 harbors a key determinant of HCV infectivity (e.g., V6, H9, S12), as underscored by mutagenesis of these residues pointing toward the p7 channel pore [10]. Intriguingly, we reveal here for the first time, a novel determinant at the extreme amino-terminal end of p7, i.e., before its first helix (S1 Fig), that strongly modulates infectivity (Fig 2) and the relative amounts of enveloped vs. non/partially-enveloped core particles (Fig 6B). Furthermore, we demonstrate that changes in this amino-terminal p7 determinant (p7 ATMI mutants), via e.g., short peptide extensions, single amino-acid insertions or substitutions that did not alter p7 structure (S1 Fig), induced a stronger reduction of infectivity (Fig 2) than of secretion of enveloped core particles (Fig 6D), resulting in reduced specific infectivity of the mutants, by 4- to over 50-fold depending on ATMI mutant types (Fig 4G–4I; S5 Fig.
Envelopment of viral particles pertains to a series of events that likely occur rapidly once two components, i.e., surface glycoproteins and nucleocapsids, encounter at assembly sites following mobilization from their respective storage pools, i.e., respectively, NS2 platforms apposed to LDs [9, 39] and LDs/replication complexes [12]. Such events are difficult to catch experimentally, because they are transient by nature as they lead to quick release of assembled viral particle from such assembly sites within the ER lumen. How HCV core and RNA are transferred from LD surface to ER assembly sites to initiate the release of infectious, enveloped viral particles remains poorly defined [3, 87], although it involves concerted actions of p7 and NS2 [17] and of NS5A [18, 20]. Accordingly, previously described assembly-defective mutants, such as p7-KR33/35QQ and core-C69-72A [40], ∆p7 [65] or ∆E1E2 and NS5A-∆2328–2435 [20], display strong core-LD accumulation, which correlates with their loss of infectivity. Strikingly, in contrast to these previous assembly mutants but similar to parental viruses, our Jc1 p7 ATMI mutants readily targeted core at the ER membrane (S8 Fig) despite reduced infectivity and did not significantly alter the co-clustering of structural and non-structural proteins with HCV RNA (Figs 8–10), both of which events previously shown to be critical for achieving efficient assembly [17, 65, 66]. Moreover, as shown by others, p7 regulates NS2 subcellular localization at punctate sites near LDs and its association with DRMs along with other viral proteins, including core, E2, and NS3 [9, 16, 39, 86]; yet, while other assembly-defective virus mutants, such as the p7-KR33/35QQ and p7-KR33/35AA in these previous reports, disrupted NS2 localization and/or E2/NS2 association, the p7 ATMI mutants displayed increased E1E2/NS2 interaction, compared to parental viruses. Along with the finding that our mutants exhibited wt capacity to slow down the cell secretory pathway (Fig 1E), this underscores that the ATMI class of p7 mutants retains most p7 properties and inhibits viral assembly though a novel mechanism.
Our results imply an envelopment defect caused by inadequate mobilization and/or transfer of core and RNA at E1E2-containing NS2 assembly platforms. This is reflected by our findings that such p7 ATMI mutants exhibited altered E1E2/NS2 and NS2/NS5A interactions (Figs 9 and 10) but also that failure to mediate correct particle envelopment resulted in secretion of partially enveloped, proteinase K-sensitive core particles (Fig 6B). Since co-expression of our mutant viruses with wt p7 restored the above alterations to almost normal levels, this questions about the role of p7 N-terminus in this mechanism and raises the possibility that it regulates core and RNA transfer to assembly sites and/or to assembling viral particles (Fig 11). Interestingly, a previous report suggested that p7 genetically interacts with some regions of NS2 as well as of NS5A [88], strengthening the notion of a functional dialog between p7, NS2 and NS5A. Our findings that the N-terminus of HA-tagged p7 points towards the cytosol (Fig 7) support the likelihood that it mediates critical interactions with cytosolic factors promoting assembly. A possibility is that modifications of p7 N-terminal surface, before the first p7 helix (S1 Fig), disrupted such interactions (Fig 11). In this respect, co-localization and association of NS2 with NS5A, which is decreased upon p7 deletion or alterations [13, 16], is thought as a crucial event mediating core/RNA transfer to assembly sites [18, 20, 66]. Thus, since our mutant viruses did not display altered co-localization of these assembly factors, it is likely that NS2 complexes formed with p7 ATMI mutants failed to efficiently mediate the encountering of nucleocapsids with NS2-bound viral surface components and/or to induce the release of fully enveloped, infectious viral particles. Why such failure resulted in a proportional release of partially enveloped core particles (Fig 6B vs. Fig 6D), which could be similar to those that have been detected in the serum of infected patients [28], is intriguing. It raises the possibility that the transfer of core and RNA to E1E2 glycoproteins liberated from NS2 complexes at assembly sites is associated to the recruitment of a mechanism that closes nascent particles (Fig 11). Such a mechanism, which likely involves components of the ESCRT pathway, as previously described for HCV [84, 85], could be disrupted by p7 ATMI mutants is such a way that, rather than being correctly closed, the budding membrane capsule would detach, allowing escape of imperfectly enveloped nucleocapsids in the ER lumen, which could explain their reduced specific infectivity. Indeed, as p7 ATMI mutants impair the interaction between NS2 and NS5A, this may prevent the recruitment of HRS, an ESCRT-0 component that interacts with p7, NS2 and NS5A [84] and subsequently, of all ESCRT components required for correct envelopment.
Our data indicate that the regulation of the amounts of free p7 could modulate the production of viral particles. Indeed, we found that p7, which localizes at the ER (Fig 7) [89] slows down the cell secretory pathway in a dose-dependent manner, likely at the stage of ER-Golgi transport (Fig 1). While this property is not intuitive, given that HCV, like other Flaviviridae, is thought to exit the cells through the secretion pathway [90], this could either induce the concentration of its glycoproteins at virion assembly sites in the ER lumen through their active retention or reflect the cooptation of another pathway of secretion for HCV particles [22, 23]. Furthermore, as HCV glycoproteins can be secreted as SVPs independently of other viral proteins (Fig 1G) [24], this feedback loop may ensure that excess glycoproteins, arising from their release upon E2p7 cleavage, could be appropriately controlled so as to prevent activation of immune responses. Additionally, as indicated by the delayed transport of VSV-Gts used as a model cargo (Fig 1B–1F), it is possible that p7 expression could alter the secretion of cellular proteins such as, e.g., immune effectors, as shown for viroporins of other viruses [43, 91].
The mechanism used by p7 to slow down the secretion of glycoproteins needs further investigation. Viroporins of alternative viruses have previously been involved in modulation of the secretory pathway, though through a variety of mechanisms [38]. For example, the M2 protein from influenza virus has a direct effect on late steps of plasma membrane delivery by delaying late Golgi transport, which indirectly affects the efficiency of earlier transport steps by altering the ionic content of the Golgi apparatus and the endosomes [92, 93]. Alternatively, Coxsackievirus 2B proteins modify ER membranes, which inhibits protein processing and sorting by decreasing calcium homeostasis in ER and Golgi [43]. Likewise, p7 can change ionic gradients in both reconstituted membrane assays in vitro [30, 44–47] and in cellulo [41, 42], which could affect anterograde transport and/or modify intracellular compartments.
In conclusion, our report underscores the function of E2p7 delayed processing in modulating i) the intracellular E2 levels, ii) the retention of E2 through the slowing down of the secretion pathway, and iii) the unmasking of functions of p7 amino-terminus in assembly and envelopment.
Huh7.5 cells (kind gift of C Rice, Rockefeller University, New York, USA) and 293T kidney (ATCC CRL-1573) cells were grown in Dulbecco’s modified minimal essential medium (DMEM, Invitrogen, France) supplemented with 100U/ml of penicillin, 100μg/ml of streptomycin, and 10% fetal bovine serum.
pFK-JFH1wt_dg, pFK-JFH1/J6/C-846_dg plasmids encoding full-length JFH1 and Jc1 HCV [94] were kind gifts from R Bartenschlager (Heidelberg University, Germany). pFK-JFH1J6XbaIC-846HAHA-L-p7_dg encoding a Jc1 virus with the HAHALp7 linker peptide between E2 and p7 [56] was kindly provided by T Pietschmann (Twincore, Germany). JFH1 virus-derived constructs encoding the p7-T2, p7-L2S, Ap7 and ASGGSp7, HAHALp7, and E2-A367R mutants were derived from the pFK-JFH1wt_dg plasmid. The E1 glycoprotein was also point-mutated in the pFK-JFH1wt_dg and pFK-JFH1 HAHALp7 constructs to introduce the A4 epitope, resulting in plasmids encoding JFH1 E1(A4) and JFH1 E1(A4) HAHALp7 viruses, respectively [57]. Constructs were created by PCR mutagenesis (oligonucleotide sequences are available upon request).
The plasmid pEGFP-N3-VSV-Gts was a kind gift from K Konan (Albany Medical College, USA). The plasmids encoding noSPp7 (JFH1), ΔE2p7 (JFH1), ΔCp7 (JFH1), ΔCp7 (H77), ΔE2p7 (J6), ΔE2HAHALp7 (JFH1), and noSPHAHALp7 (JFH1) allow individual expression of wt, variant or mutant p7 under different signal peptide configurations. The plasmids pTG 13077-HCV-ΔC-E1-E2-J6, pTG 13077-HCV-ΔC-E1-E2-JFH1 and pTG 13077-HCV-ΔC-E1-E2-p7-JFH1 contain retroviral vector genomes encoding E1E2 and/or E1E2p7 proteins from J6 and JFH1 viruses.
All constructs were expressed in Huh7.5 cells using procedures reported before [17].
Mouse anti-actin (clone AC74, Sigma-Aldrich), mouse anti-E1 A4 (kind gift from HB Greenberg), rat anti-HA (clone 3F10, Roche), mouse anti core C7-50 (Thermo Fisher Scientific), rat anti-E2 clone 3/11 (kind gift from J McKeating), mouse anti-NS2 6H6 and mouse anti-NS5A 9E10 (kind gift from C Rice), rabbit anti-NS2 (kind gift from B Lindenbach), human anti-E2 AR3A (kind gift from M Law), mouse anti-GFP (Roche), anti-VSV-G 41A1, mouse anti-E2 antibody AP33 (kind gift from A Patel) were used according to the providers’ instructions.
Huh7.5 cells were seeded 16h prior to transfection with pEGFP-N3-VSV-Gts and p7-encoding plasmids using GeneJammer transfection reagent (Agilent). Medium was changed 4h post-transfection and cells were incubated overnight at 40°C. 24h post-transfection, cells were chased at 32°C. For western blot analysis, cells were lysed at indicated time points in wells cooled on ice before clarification and western blot analysis. For flow cytometry analysis, cells were harvested and put in suspension at 32°C. At indicated time points, cells were fixed with 3% paraformaldehyde.
The plasmid popol-ΔE1sE2 (JFH1)-H6 (kind gift from Epixis SA) encoding soluble E2 (JFH1) with a 6xHis tag was used to purify E2 in order to assess the sensitivity of E2 quantifications by western blots. 293T cells grown in 10 cm-plates were transfected with 15μg of popol-ΔE1sE2 (JFH1)-H6. 16h post-transfection, the medium was replaced by OptiMEM. 24h and 48h later, supernatant was harvested and purified using a HisTrap column. Fractions were pooled and then dialyzed. A sample was analyzed by SDS-PAGE. Concentration of sE2 was obtained by measurement of OD at 280nm and purity was analyzed by LC-MS/MS.
HEK293T cells were seeded 24h prior to transfection with VSV-G plasmid, pTG-5349 packaging plasmid, and either pTG 13077-HCV-ΔC-E1-E2-JFH1, pTG 13077-HCV-ΔC-E1-E2-J6 or pTG 13077-HCV-ΔC-E1-E2-p7-JFH1 plasmids using calcium phosphate precipitation. Medium was replaced 16h post-transfection. Vector supernatants were harvested 24h later, filtered through a 0.45 μm filter, and were titrated by flow cytometry using AP33 antibody against E2.
Lentiviral vectors were used to transduce Huh7.5 cells (MOI = 2). 72h post-transduction, cell supernatants were centrifuged at 25,000 rpm for 4h at 4°C using SW41 rotor and Optima L-90 centrifuge (Beckman). Pellets were suspended in PBS prior to use for western blot analysis.
For gradient analysis, 1 ml of supernatant concentrated 40x by Vivaspin columns (MW cut-off 100-kDa (Sartorius)) was loaded on iodixanol density gradients. 12 fractions were collected from the top and used for refractive index measurement and precipitation of proteins before western blot analysis.
Methods for in vitro transcription of HCV RNA and its electroporation into Huh-7.5 cells have been described [17, 61]. When p7 was co-expressed with viral RNA, 2μg of plasmid DNA encoding p7 or control DNA were co-electroporated with 10μg of viral RNA.
To determine the percentage of HCV-positive producer cells following electroporation, cells were fixed and stained using Cytofix/Cytoperm (BD) according to manufacturer's instructions. NS5A staining was achieved with 9E10 antibody (kind gift from C. Rice, Rockefeller University, New York, USA) and cells were analyzed using MacsQuant VYB (Milteny Biotech).
Electroporated cells were counted and 100,000 cells were lysed in lysis buffer (20 mM Tris [pH 7.5], 1% Triton X-100, 0.05% sodium dodecyl sulfate, 150 nM NaCL, 5‰ Na deoxycholate) supplemented with protease/phosphatase inhibitor cocktail (Roche) and clarified from the nuclei by centrifugation at 13,000×g for 10 min at 4°C for quantitative western blot analysis (see below).
HCV core protein was also quantified by CMIA—Chemiluminescent Microparticle ImmunoAssay (Architect, Abott). The extracellular E2 protein was quantified by Western Blot after precipitation of E2-containing cell supernatants with Galanthus Nivalis lectins (GNA) bound to agarose beads (Vector Laboratories). The extracellular HCV RNAs were quantified as described previously [61]. Infectivity titers were determined as focus-forming units per milliliter [17]. Serial dilutions of supernatants were used to infect Huh7.5 cells and focus-forming units were determined 3 days post-infection by counting NS5A-immunostained foci. For determining intracellular infectivity, electroporated cells were washed with PBS, harvested with Versene and centrifuged for 4 min at 400xg. Cell pellets were suspended in medium and subjected to 4 cycles of freeze and thaw, using liquid nitrogen.
For purification of particles, supernatants were harvested and filtered through a 0.45μm filter and centrifuged at 25,000 rpm for 1h45 at 4°C with a SW41 rotor and Optima L-90 centrifuge (Beckman). Pellets were resuspended in PBS prior to use for western blot to quantify E2 or for quantification of core and RNAs.
Viral supernatants were i) left untreated, ii) treated with Proteinase K (PK, 50 μg/mL) in 10x PK buffer as described in [84] for 1h on ice, or iii) pre-treated with Triton X-100 5min at room temperature prior to treatment with PK. PK activity was stopped by adding 10 mM PMSF and protease inhibitors cocktail (Roche). The core protein was quantified with CMIA.
1mL of viral supernatant was loaded on top of a 3–40% continuous iodixanol gradient (Optiprep, Axis Shield). Gradients were centrifuged for 16h at 4°C in Optima L-90 centrifuge (Beckman). 16 fractions of 750 μl were collected from the top and used for refractive index measurement infectivity titration, core quantification and RNA quantification, as described above. For E2 protein analysis, HCV particles were produced in OptiMEM (Invitrogen) and concentrated 40x by Vivaspin molecular weight cutoff 100-kDa columns (Sartorius). 1 ml of concentrated virus suspension was loaded on density gradients. 12 fractions were collected from the top and used for refractive index measurement, titration, core quantification and RNA quantification, as described above. The remaining volumes of fractions were used for protein precipitation with 4 volumes of acidified acetone/methanol buffer and left at -20°C overnight. Proteins were pelleted at 16,000xg for 15min and dried before resuspension in lysate buffer, denaturation in Laemmli buffer, and Western Blot analysis.
Endoglycosidase Hf (Endo-Hf; NEB) treatment was performed according to the manufacturer's recommendations. Briefly, protein samples were mixed to denaturing glycoprotein buffer and heated at 100°C for 5 min. Subsequently, 1,000 units of Endo-Hf were added to samples in a final volume of 25 μl and the reaction mixtures were incubated for 1 h at 37°C, before western blot analysis.
Proteins obtained in total lysates or after digestion or immunoprecipitation, were denatured in Laemmli buffer at 95°C for 5min and were separated by sodium dodecyl sulfate polyacrylamide gel electrophoresis, then transferred to nitrocellulose membrane and revealed with specific primary antibodies, followed by the addition of IRdye secondary antibodies (Li-Cor Biosciences), followed by imaging with an Odyssey infrared imaging system CLx (Li-Cor Biosciences).
For NS2/E2 interaction, 1 million electroporated cells were lysed with buffer (50 mM Tris-Cl (pH 7.5), 150 mM NaCl, 1% Nonidet P-40, 1% sodium deoxycholate, and 0.1% SDS). Lysates were cleared by centrifugation at 16,000xg for 10 min at 4°C and were incubated overnight at 4°C with AR3A antibody against HCV E2 or with rabbit NS2 antibody. Protein A/G-coated agarose beads were added to samples for 2h at room temperature. Immune complexes were then washed and eluted with Laemmli buffer for 5 min at 95°C before western blot analysis.
For NS2/NS5A interaction, 1 million electroporated cells were cross-linked with 1mM dithiobis(succinimidyl propionate) (DSP) (ThermoFisher) 30min at room temperature. Tris (pH 7.5) was added up to 200 mM to quench unreacted DSP. Cells were resuspended in lysis buffer (50mM Tris pH 7.4, 150mM NaCl, 1mM EDTA, 0.5% n-dodecyl-β-maltoside) and treated as for NS2/E2 interaction with incubation with rabbit NS2 antibody.
Experimental procedures were previously described [66]. Briefly, Huh7.5 cells grown on glass coverslips and were infected at MOI of 0.2. 72h post-infection, cells were washed with PBS, fixed with 3% paraformaldehyde in PBS for 15min, quenched with 50mM NH4Cl and permeabilized with 0.1% Triton X-100. Fixed cells were then incubated with primary antibodies in 1% BSA/PBS, washed and stained with the corresponding fluorescent Alexa-conjugated secondary antibody (Alexa-488, Alexa-555 and Alexa-647, Molecular Probes) in 1% BSA/PBS. LDs were stained with 10μg/mL Bodipy 493/503 (Molecular Probes) according to the manufacturer’s instructions. Cells were washed with PBS, stained for nuclei with Hoechst (Molecular Probes) and mounted with Mowiol 4–88 (Sigma-Aldrich) prior to image acquisition with LSM-710 (Zeiss) confocal microscope. When stated, the combined detection of HCV RNA by FISH and viral proteins was done as previously described [66].
For digitonin permeabilization, the staining procedure was the same except that cells were permeabilized with 5μg/ml Digitonin (Sigma-Aldrich) for 10 min. Cells permeabilized with Triton X-100 were acquired first; then, cells permeabilized with Digitonin were acquired in order to use the same laser settings.
Images were analyzed and quantified with the ImageJ software as previously described [66]. The Pearson’s and Manders’ correlation coefficients were calculated by using the JACoP plugin [95]. For the Digitonin vs. Triton permeabilization experiments, the relative fluorescence intensity of each channel was quantified by using the integrated density measurement of ImageJ software.
Three-dimensional homology models of p7 hexamers and their mutants were constructed using the NMR/MD p7 model of Chandler and colleagues [36] and the NMR p7 structure of OuYang and colleagues [35] (PDB accession number 2M6X) as templates. Models of p7 were constructed with the Swiss-Model automated protein structure homology modeling server (http://www.expasy.org/spdbv/ [96]) using the HCV JFH1 strain p7 sequence as input. JFH1 p7 homology model derived from OuYang et al. was directly obtained as a hexamer by the automated procedure. For JFH1 p7 homology model derived from Chandler et al., raw amino-acid sequence of p7 from strain JFH1 was first loaded in Swiss-PdbViewer software [96] and fitted to the NMR/MD p7 hexamer model [36] before submission for model building to Swiss-Model using the SwissModel Project Mode. All p7 JFH1 mutants were constructed using the latter protocol, i.e., fitting of the raw amino-acid sequence of p7 mutants to wild type hexamer models from the JFH1 strain. Coordinates of homology models derived from the automated model building were used without further minimization or manual manipulation. For mutants Ap7 and ASGGSp7 exhibiting N-terminal extensions, additional residues were added manually assuming a random conformation and were minimized using Swiss-PDB Viewer tools.
Significance values were calculated by applying the paired t-test using the GraphPad Prism 6 software (GraphPad Software, USA). For confocal analysis, a two-tailed, unpaired Mann-Whitney test was applied. P values under 0.05 were considered statistically significant and the following denotations were used: ****, P≤0.0001; ***, P≤0.001; **, P≤0.01; *, P≤0.05; ns (not significant), P>0.05.
|
10.1371/journal.pbio.1001823 | In Vivo Time-Resolved Microtomography Reveals the Mechanics of the Blowfly Flight Motor | Dipteran flies are amongst the smallest and most agile of flying animals. Their wings are driven indirectly by large power muscles, which cause cyclical deformations of the thorax that are amplified through the intricate wing hinge. Asymmetric flight manoeuvres are controlled by 13 pairs of steering muscles acting directly on the wing articulations. Collectively the steering muscles account for <3% of total flight muscle mass, raising the question of how they can modulate the vastly greater output of the power muscles during manoeuvres. Here we present the results of a synchrotron-based study performing micrometre-resolution, time-resolved microtomography on the 145 Hz wingbeat of blowflies. These data represent the first four-dimensional visualizations of an organism's internal movements on sub-millisecond and micrometre scales. This technique allows us to visualize and measure the three-dimensional movements of five of the largest steering muscles, and to place these in the context of the deforming thoracic mechanism that the muscles actuate. Our visualizations show that the steering muscles operate through a diverse range of nonlinear mechanisms, revealing several unexpected features that could not have been identified using any other technique. The tendons of some steering muscles buckle on every wingbeat to accommodate high amplitude movements of the wing hinge. Other steering muscles absorb kinetic energy from an oscillating control linkage, which rotates at low wingbeat amplitude but translates at high wingbeat amplitude. Kinetic energy is distributed differently in these two modes of oscillation, which may play a role in asymmetric power management during flight control. Structural flexibility is known to be important to the aerodynamic efficiency of insect wings, and to the function of their indirect power muscles. We show that it is integral also to the operation of the steering muscles, and so to the functional flexibility of the insect flight motor.
| A blowfly's wingbeat is 50 times shorter than a blink of a human eye, and is controlled by numerous tiny steering muscles—some of which are as thin as a human hair. To visualize the movements of these muscles and the deformations of the surrounding exoskeleton, we developed a technique to allow us to look inside the insects during tethered flight. We used a particle accelerator to record high-speed X-ray images of the flying blowflies, which we used to reconstruct three-dimensional tomograms of their flight motor at ten different stages of the wingbeat. We measured the asymmetric movements of the steering muscles associated with turning flight, together with the accompanying movements of the wing hinge—arguably the most complex joint in nature. The steering muscles represent <3% of total flight muscle mass, so a key question has been how they can modulate the output of the much larger power muscles. We show that by shifting the flight motor between different modes of oscillation, the fly is able to divert mechanical energy into a steering muscle that is specialized to absorb mechanical energy. In general, we find that deformations of the muscles and thorax are key to understanding this remarkable mechanism.
| Insects are the smallest and most agile of all flying animals. These attributes are taken to extremes in the dipteran flies, whose single pair of wings enable a range of dramatic flight manoeuvres, from turning on the spot, or flying backwards, to even landing on ceilings. The blowfly Calliphora vicina routinely pulls up to four times its body weight during turns [1], and its seemingly simple reciprocal wingbeat belies the complexity of the flight motor that drives it [2]. Each of the two wings is powered by four stretch-activated muscles that undergo self-induced oscillations at a frequency in excess of 100 Hz. Rather than attaching directly to the wings, these indirect power muscles drive small amplitude deformations of the thorax, which are then amplified through the intricate wing hinge [3]. This arrangement leaves little scope for the indirect power muscles to create the wing kinematic asymmetries that are required for asymmetric flight manoeuvres [4],[5]. Instead, kinematic asymmetries are produced by the 13 steering muscles [6],[7]. Collectively the steering muscles have <3% of the total mass of the indirect power muscles, which leads to a key, unresolved question. How are the tiny steering muscles able to shape the vastly greater—and essentially symmetric—output of the indirect power muscles [4],[5], so as to produce the large wingbeat asymmetries that enable fast flight manoeuvres?
The wing articulates with the thorax through a complex arrangement of cuticular structures called sclerites, which the steering muscles actuate [3],[6],[7]. Here, we provide a brief overview of the sclerites and associated muscles. A more detailed anatomical description of the muscles and their attachment points in blowflies can be found in [6], and these are described for the dipteran flight motor in general in [3]. The wing hinge is formed by four axillary sclerites, but only the first, third, and fourth of these have steering muscles attached. In flies, the fourth axillary sclerite is fused to the thoracic wall, and is usually referred to as the posterior notal wing process. The wing base is also connected by ligaments to the external head of a lever-like control linkage called the basalare sclerite, which projects into the thorax. In summary, ten steering muscles insert on the axillary sclerites; a further three insert on the basalare sclerite [3],[6]. Rather little is known about how the steering muscles modify the motions of these sclerites, but electrophysiological studies have correlated the activation states of some of these muscles with variation in wingtip kinematics, principally focussing upon variation in stroke amplitude. For example, during visually stimulated roll responses in Calliphora, activity of the first and second basalare muscles, and activity of at least some of the third axillary muscles, is associated with increased stroke amplitude. However, because several steering muscles are active during roll manoeuvres [3],[8],[9], the individual function of each muscle cannot readily be inferred from electrophysiological and wing kinematic data alone. Other work has attempted to identify the effects of different groups of steering muscles upon the aerodynamic forces and moments [10], but the specific mechanisms through which the steering muscles manipulate the wing hinge sclerites remain elusive.
Elucidating muscle function fully requires measurements of stress, strain, and activation, combined with knowledge of the mechanism the muscle actuates [11]. These measurements can be made simultaneously in larger vertebrates [12], but this has not yet been achieved in insects. Most of our current understanding of steering muscle function comes from anatomical [6],[7],[13] and electrophysiological [8],[9],[14]–[19] studies, and we know surprisingly little about the mechanics of how the steering muscles control the wingbeat. This is due in part to the extraordinary difficulty of measuring micrometre-scale muscle movements in vivo at frequencies in excess of 100 Hz. Indeed, although patterns of muscle activation [4],[8]–[10],[14],[16]–[19] and stresses produced under work-loop conditions [20] have been characterised for some insect steering muscles, almost nothing is known about the associated muscle strains and the resulting thoracic movements. Techniques used to measure and visualize muscle strains in vertebrates, such as sonomicrometry [21],[22] and stereo X-ray imaging [23], are unsuited to insects. Strains have been measured in insect power muscles using vivisective microscopy [24], external markers [25], and X-ray diffraction [26], but the smaller size of the steering muscles and their close interaction with the wing hinge makes them inaccessible even to these methods. To study the kinematics of the steering muscles, we therefore developed a new imaging technique allowing high-resolution, time-resolved microtomography of blowflies (C. vicina) in tethered flight (Figure 1).
Microtomography has previously been used in vivo to make time-resolved measurements of mouse hearts and lungs [27],[28], but to resolve the actuation of the insect flight motor we have extended the spatial and temporal resolutions of the technique by an order of magnitude each. This allowed us to produce tomographic visualizations of the instantaneous state of the flight motor for ten evenly spaced phases of the wingbeat (Movie S1, view here). We used these data to measure and compare the muscle strains and thoracic movements associated with different wingbeat kinematics. Taken together, our results emphasise the importance of muscular and cuticular deformations in modulating and controlling the kinematics of flapping flight.
We undertook time-resolved microtomographic imaging of the thorax of tethered blowflies flying in the TOMCAT beamline of the Swiss Light Source [29]. We used single exposure phase retrieval to increase contrast by an order of magnitude over standard absorption-based imaging [30]. This was important to enable the high acquisition rates and short exposure times required to resolve the wingbeat cycle. The insects were tethered to a rotating stage that underwent four complete revolutions per recording, thereby allowing radiographs to be taken from multiple evenly spaced viewing angles whilst the insect was flying (Figure 1). We simultaneously captured the three-dimensional wingtip kinematics using stereo high-speed photogrammetry [31] and grouped the radiographs according to the wingtip position. Each group contained multiple radiographs corresponding to the same phase of the wingbeat, but taken from different viewing angles. This allowed us to reconstruct tomograms for each group separately, producing tomograms for ten evenly spaced phases of the wingbeat. Each tomogram pools radiographs from c. 600 wingbeats and therefore represents the average state of the flight motor at the corresponding phase of the wingbeat.
The flies were rotated during radiographic acquisition (332° s−1 or 347° s−1), producing a left-handed visual and inertial roll stimulus in the brightly lit lab environment (Figures 1, 2A, and 3). The left wing had consistently higher stroke amplitude than the right wing (141±7° versus 100±9°; mean ± standard deviation), and a shallower stroke plane (47±4° versus 68±10°), typical of a stabilizing roll response [13],[19],[32],[33]. The results of our experiments therefore allow us to compare the muscle strains and thoracic movements associated with simultaneous high versus low amplitude wingbeats in each individual. We analysed all three muscles inserting on the basalare sclerite (b1, b2, b3), and the two largest muscles (I1, III1) inserting on the first and third axillary sclerites (Figures 4 and 5). Together, these make up most of the mass of the steering muscles [6],[7]; the other eight steering muscles are smaller and could not be distinguished reliably from the surrounding tissues.
We first used our visualizations to describe the motions of the thoracic mechanisms that the steering muscles actuate (Movie S1, view here; Movie S2, view here; Movie S3, view here). The muscles that attach to the first axillary sclerite insert on its internal arm, which projects into the thorax and moves in opposition to the wing [7]; in contrast, the third axillary sclerite moves rather little relative to the base of the thorax (Movie S2, view here). The lever-like internal arm of the basalare sclerite oscillates back-and-forth (Figure 6; Movie S2, view here), while its external head articulates with a moving part of the thoracic wall called the pleural plate (Figure 7; Movie S3, view here). This hardened region of thoracic wall swings antero-ventrally on the downstroke, accommodated by the alternate opening and closing of two orthogonal clefts at its borders [34]. Rotation of the pleural plate was clearly responsible for driving oscillations of the basalare sclerite, which were of greater amplitude on the high-amplitude wing (Movie S3, view here).
The wingbeat asymmetries that we measured were associated with bilateral asymmetries in steering muscle kinematics (Figures 4–6; Movie S2, view here), which we quantified by measuring strains directly from the tomograms (Figures 5, 6, and 8). We were unable to measure muscle resting length for the purposes of normalizing muscle strains because the flies were flying continuously and reacted to the roll stimulus throughout each recording. Instead, we referenced the strain of each muscle in a pair to the pooled mean length of both muscles, which allowed us to compare muscle strains within each pair and between flies. Mean muscle strain was bilaterally asymmetric within each muscle pair (Figure 8C): higher on the high-amplitude wing for muscles I1, III1, and b3; but lower on the high-amplitude wing for muscles b1 and b2 (Figure 6). All of the muscles except III1 displayed detectable strain oscillations at wingbeat frequency, but we could only detect statistically significant bilateral amplitude asymmetries in muscles b1 and b3 (Figure 8A). The amplitude of these strain oscillations was twice as high on the low-amplitude wing for b1 (Figures 6C and 8A), and four times as high on the high-amplitude wing for b3 (Figures 6A and 8A). The b1 strain oscillations also displayed a statistically significant phase asymmetry, with the oscillations on the low-amplitude wing delayed by a quarter of a wingbeat (Figure 8B).
Muscle strains need not always be caused by contraction of the muscle itself. For example, work-loop measurements have shown that b1 is specialized to do negative work (i.e., to absorb rather than impart kinetic energy), and is unable to cycle fast enough to drive oscillations at wingbeat frequency [20]. The measured b1 oscillations must therefore have been driven by oscillations of the basalare sclerite forced by movement of the wing and thorax (Movie S1, view here; Movie S2, view here; Movie S3, view here). We cannot say unequivocally why the b1 strain oscillations were bilaterally asymmetric, but in principle this must reflect either asymmetric loading or asymmetric stiffness. Electrophysiological studies have shown that b1 is activated earlier with increasing wingbeat amplitude, which increases both its stiffness and the amount of negative work done under a given strain [9],[10],[19]. It has therefore been hypothesised that this increased stiffness should cause the amplitude of the b1 muscle's oscillations to be lower when the wingbeat amplitude is higher. Our strain measurements support this hypothesis, but our visualizations show that the explanation is incomplete. This is because the lower amplitude oscillations of b1 on the high-amplitude wing are actually associated with larger oscillations of the basalare sclerite (Figure 7; Movie S3, view here). The picture is further complicated by the fact that b3, which is expected to act antagonistically with b1, also has higher amplitude oscillations on the high-amplitude wing (Figure 6A).
To resolve this puzzle, we examined the movements of the basalare sclerite in greater depth. Our visualizations show that movement of the basalare sclerite is dominated by rotation about its external head on the low-amplitude wing, but by dorso-ventral translation of the whole sclerite on the high-amplitude wing (Movie S2, view here; Movie S3, view here). Consequently, the internal tip of the basalare sclerite traces an orbit that is aligned with b1 on the low-amplitude wing, but with b3 on the high-amplitude wing (Figure 5C; Movie S2, view here). These different modes of oscillation of the basalare sclerite explain why the strain amplitude is higher on the low-amplitude wing for b1, but higher on the high-amplitude wing for b3. We cannot determine how this is brought about, but one possibility is that the variable stiffness of the b1 muscle alters the impedance of the system anisotropically. Another possibility is that the orientation of the basalare sclerite is altered by the large b2 muscle [17],[19], which, like b1, has a lower mean strain on the higher amplitude wing (Figure 6D).
Turning manoeuvres are associated with asymmetric aerodynamic power requirements, which cannot be met by varying the output of the power muscles asymmetrically [35]. We hypothesise that changing the mode of oscillation of the basalare sclerite serves to increase the amount of kinetic energy transferred to b1 on the low-amplitude wing, thereby absorbing excess muscle output. To test the plausibility of this hypothesis, we combined our measurements of b1 muscle strain with the results of a previous work-loop study [20], to estimate the amount of negative work being done by b1. Unlike the other steering muscles, b1 is typically active on both wings, although it is not necessarily activated on every wingbeat. We estimate that b1 would have done negative work at a rate of 0.04–0.06 mW on the high-amplitude wing (0.02 mW if inactive) and 0.18–0.30 mW on the low-amplitude wing (0.06 mW if inactive). These intervals bracket the entire range of possible activation phase, and show that the b1 muscle could have been doing negative work at a rate up to 0.28 mW higher on the low-amplitude wing. This would be sufficient to manage anything up to a 24% asymmetry in the time-averaged aerodynamic power requirements of Calliphora, which have been estimated to be 1.58 mW per wing on the downstroke, and 0.81 mW per wing on the upstroke [36]. Our results therefore demonstrate that the b1 muscles could play a significant role in asymmetric power management, although it remains an open question whether the activation phase of b1 is controlled appropriately for this function.
Our visualizations reveal a completely unexpected behaviour in another steering muscle, showing that the long tendon that connects the I1 muscle to the first axillary sclerite buckles when the wing is elevated above the wing hinge. This behaviour was observed on both wings in all four individuals, and was always greater on the high-amplitude wing (Figure 9; Movie S2, view here). Buckling only occurs under compressive loading, so it follows that both I1 muscles must be under compression in the upper part of the wingbeat. Consequently, I1 contraction cannot possibly increase stroke amplitude by exerting tensile stress on the first axillary sclerite at the top of the upstroke, contrary to what has been inferred previously from static anatomy [6],[7]. Instead, I1 contraction must limit the movement of the wing at the bottom of the downstroke, thereby reducing stroke amplitude. Consistent with this interpretation, I1 muscle strain was always lower on the low-amplitude wing. This includes those points in the stroke cycle at which the tendon transitioned between its taut and buckled states. Since the I1 tendon must have been unloaded at these transition points, the fact that the muscle was shorter on the low-amplitude wing necessarily implies that I1 must have been contracted on the low-amplitude wing. This conclusion is consistent with the correlations observed in previous electrophysiological studies, which have found that I1 is only active at reduced stroke amplitude [8]–[10].
Buckling of the I1 tendon is important for two reasons. First, it accommodates higher amplitude movements of the first axillary sclerite than would otherwise be possible, because the effective strain measured along the straight line joining the origin of the tendon to the origin of I1 (Figure 9C) has four times the amplitude of the actual strain that the I1 muscle experiences on the high-amplitude wing (Figure 6E). Second, it means that I1 contraction will always be intermittent in its effects within each stroke cycle, even if—like b1—the I1 muscle is unable to cycle at wingbeat frequency. Tendon buckling is not unique to I1. Although we were unable to visualize the second muscle of the first axillary sclerite (I2) fully, our visualizations show that the long tendon of this muscle also buckles on every wingbeat. Tendon buckling also occurs to a lesser extent in b3 (Movie S2, view here). This previously unknown phenomenon of tendon buckling may therefore be a rather general mechanism in the operation of the blowfly flight motor.
The fast, complex, three-dimensional movements of the insect flight motor are powered and controlled by several tens of linear actuators, each individually producing only a low-amplitude contractile strain. Here we have presented the first time-resolved visualisations of the workings of this extraordinary mechanism. Our results clearly show that the function of the steering muscles in controlling the wing kinematics can only be understood by placing them in the context of the deforming thoracic structures to which they attach. Deformations of the thoracic wall are not only responsible for transmitting forces from the power muscles to the wings, but are also important in accommodating qualitative changes in the modes of oscillation of the wing articulations. Likewise, deformations of the tendons connecting the steering muscles to the wing articulations are important in accommodating large excursions of the wing articulations, whilst permitting the steering muscles to curtail the wing's movement at certain stages of the stroke cycle. Structural flexibility is known to be important to the aerodynamic efficiency of insect wings [37], and to the function of their indirect power muscles. We have now shown that it is integral also to the operation of the steering muscles, and so to the functional flexibility of the insect flight motor. We anticipate that the insights from this work will inspire the design of future micromechanical systems, and the technique that we have developed is of course applicable to other biological systems exhibiting periodic motion.
Blowflies (C. vicina) were collected from a permanent breeding colony at the Department of Bioengineering, Imperial College London and kept on a 24 h (12∶12) light-dark cycle. All individuals were used within two weeks of emergence at ambient lab temperature. Insects were cold-anesthetized at 4°C for 10 minutes and fixed dorsally by the scutum to a wooden tether, using a mixture of beeswax and colophonium. The scutum is a stiff, reinforced thoracic structure [7], and is the standard mounting point for tethered flight preparations in flies. The wooden tethers were attached to a rotation stage using a custom-made holder to align the anteroposterior axis of the animals with the rotational axis of the end station (Figure 1). The insects (n = 4) were placed in a 2 ms−1 airstream and left to settle into flight for >30 s before recording radiographs.
The X-ray source was a superbending magnet located 25 m from the sample. Monochromatic and polychromatic beam configurations were available, and we ran experiments using both types of configuration for comparison. In the monochromatic configuration (n = 2), a double crystal multilayer monochromator was placed 7 m downstream of the source to extract monochromatic X-rays with a bandwidth of 2% at 18 keV photon energy (wavelength = 0.7 Å) and flux of 8×1011 ph/s−1 mm−2 at the sample site. The monochromator was removed in the polychromatic configuration (n = 2), which increased total photon flux by two orders of magnitude and increased the mean photon energy to 35 keV. However, the polychromatic beam was filtered to optimize the bandwidth and the peak wavelength value of the X-rays, which reduced the beam power to an estimated 2×1012 ph/s−1 mm−2 and mainly attenuated longer wavelengths. The beam was 10 mm wide and 4.1 mm high at the sample site under the monochromatic configuration, but was increased in height to 5.7 mm under the polychromatic configuration, which enabled visualization of the entire thorax (Figure 2). The polychromatic beam therefore offers the advantages of a higher flux and larger sampling volume compared to the monochromatic beam, but the algorithms used to reconstruct tomograms from the radiographs assume a specific beam energy, which is better defined for the monochromatic beam. In practice, we found no qualitative difference in the contrast or detail of the radiographs or tomograms between beam configurations, and conclude that both beam configurations allowed comparably good imaging. Results from both configurations are pooled in the analyses which follow.
A 100 μm thick, Ce-doped LuAG scintillator was placed at a distance of 350 mm (monochromatic configuration) or 150 mm (polychromatic configuration) behind the sample to convert the transmitted X-rays into visible light. The scintillator distance was chosen to maximize the phase contrast of the radiographs and was dependent upon the mean photon energy (18 keV for the monochromatic beam and 35 keV for the polychromatic beam). The resulting edge-enhanced image was magnified using a custom-made, high numerical-aperture microscope (Elya solutions, s.r.o) offering continuously adjustable 2- to 4-fold magnification. Projection images were acquired with a pco.Dimax 12-bit CMOS detector system recording at 2,500 Hz for the monochromatic beam and 1,840 Hz for the polychromatic beam, while the insects were rotated at 347° s−1 or 332° s−1, respectively.
The laboratory environment provided a rich, high-contrast, visual scene, which would have stimulated the visual system of the insects strongly during rotation. The rotation rates of 347° s−1 and 332° s−1 were an order of magnitude higher than the lowest rates known to induce visually stimulated turning reactions in Diptera [9],[14]. The angular velocity of the insect during rotation was three orders of magnitude lower than the mean wingtip velocity, so any bilateral asymmetries in the wing kinematics must have been due to changes in flight motor output in response to the roll stimulus, rather than passive aerodynamic effects due to rotation.
Two synchronized Photron SA3 cameras (Photron Ltd) with 180 mm Sigma macro lenses were used to film the blowflies, recording at 4,000 Hz with a 33.3 μs exposure time and at 448×384 pixel image size (Figure 1). Illumination for the cameras was provided by a custom-built infrared LED light source directed onto white card below the insect. The cameras were calibrated using fully-automated calibration software running in Matlab (The Mathworks Inc.) [31]. We tracked the wingtips using background subtraction and manual thresholding to isolate the outlines of the wings in each camera view. The tip of each wing was determined as the point along the outline that was furthest from the wing hinge. The three-dimensional coordinates of the wingtip were then calculated using the camera calibration parameters.
A data acquisition module (National Instruments USB-6211 DAQ), sampling at 80 kHz, was used to record the exposure times of the Photron SA3 cameras and the pco.Dimax detector system for the purposes of grouping the radiographs. The flies had a mean wingbeat frequency of 145 Hz, so each 4 s recording consisted of approximately 600 wingbeats (Figure 3). We used the measured wingtip kinematics to group radiographs taken from different angles but at identical phases of the wingbeat. We identified the beginning and end of each wingbeat from the wingtip kinematics, and selected the radiographs closest in exposure time to ten evenly spaced phases of each wingbeat for analysis. This allowed us to combine data from all of the wingbeats measured for a given fly, despite the fact that their period was somewhat variable (Figure 3). Our tomographic reconstruction technique therefore produced one composite wingbeat for each individual, comprising ten time steps, where every time step pools radiographs from c. 600 wingbeats.
The projections were despeckled to remove bright pixels caused by scattered X-rays hitting the detector, and were flat field corrected with the average flat-beam images (i.e., images taken with no sample) and dark images (i.e., images taken with no beam) acquired immediately after the scan. Phase retrieval was performed in a qualitative manner using the ANKAPhase implementation [38] single image phase retrieval algorithm under the assumption that the object consisted of a homogeneous soft tissue material [39]. We assumed that the steering muscles had a refractive index equal to that of water [40]. For the monochromatic beam, the real and imaginary parts of the deviation from one of the complex refractive index of the material were 7×10−7 and 5×10−10, respectively. For the polychromatic beam, we assumed that the mean X-ray energy was 35 keV and used values of 2×10−7 and 10−10 for the real and imaginary parts, respectively, of the decrement from one of the index of refraction. Tomographic reconstruction was performed using a Fourier transform-based algorithm [41]. The resulting voxels had an isotropic spacing of 3.3 μm, with no discernible difference between tomograms collected using the monochromatic or polychromatic beam.
The tomographic data were visualized and segmented using Amira (VSG). We segmented the data using a manual threshold that separated the muscles and cuticle from the surrounding material (Figure 5). The manual threshold was chosen at a level approximately double that of the background noise (Figure 10). The end points of the muscles were manually tracked using natural features as markers to ensure that the same parts of the muscles were tracked from one frame to the next and between individuals (Figure 5A). These end points were then used to calculate the lengths for each steering muscle (Figure 5B). Both b3 and I1 exhibited tendon buckling during parts of the wingbeat. To take account of this, we used three-dimensional skeletonization [42] to find the line running through the centre of the tendon, which was then connected to the muscle ends to form a continuous line (Figure 5B).
A sinusoid of arbitrary mean, amplitude, and phase can be expressed as a linear combination of a sine function, a cosine function, and a constant. For each pair of steering muscles, we used a single linear model to regress the strains that we had measured for both wings on the sine and cosine of the wingbeat phase, comparing the fitted coefficients between wings. We did not control separately for fly identity, because the strain measurements had already been normalized by the mean value for each fly, such that the mean strain was the same for all flies (i.e., equal to zero). We used a Monte Carlo method to transform the 95% confidence intervals for the parameter estimates of the linear model into 95% confidence intervals for the mean, amplitude, and phase of the strain oscillations. This allowed us to test statistically for differences in the mean, amplitude, and phase of the strain oscillations between the high- and low-amplitude wings (Figure 8).
Tu and Dickinson [20] measured the negative work done by the b1 muscle at different amplitudes of oscillatory strain, and with different phases of muscle activation, using the work loop technique. We interpolated their data to estimate the range of negative work that would be done by the muscle with the measured strain amplitudes of 2.3% and 5.5%. This allowed us to estimate that the net negative work done per wingbeat would have been in the range 0.25–0.41 μJ for the high-amplitude wing, and in the range 1.23–2.06 μJ for the low-amplitude wing, depending upon the unknown phase of muscle activation. The mean wingbeat frequency in our data was 145 Hz, so the b1 muscle would have been absorbing kinetic energy at a rate of 0.04–0.06 mW on the high-amplitude wing and 0.18–0.30 mW on the low-amplitude wing. If the muscle were inactive on either the high- or low-amplitude wing, kinetic energy would have been absorbed at a rate of 0.02 mW and 0.06 mW, respectively.
Tethering is known to affect wing kinematics in other dipteran species [43], but there is a paucity of free-flight data for Calliphora with which to compare our tethered wing kinematics, particularly during the roll manoeuvres that we have simulated. The mean wingbeat frequency (145±11 Hz) and mean stroke plane angle on each wing (46.8°±4.1° low-amplitude wing, 68.0±9.6° high-amplitude wing), were within ranges observed in a free-flying Calliphora [44], with similar wing length (9.2±0.5 mm free-flight data versus 8.7±0.4 mm in our data). Mean stroke amplitude on the high-amplitude wing (141°±7°) was also within the range of free-flying Calliphora (123°–150°), but the mean stroke amplitude on the low-amplitude wing (100°±9°) was slightly lower than previously recorded. However, free-flight kinematics have only been measured in symmetric flight conditions, and Calliphora typically reduce the stroke amplitude on the ipsilateral side during roll manoeuvres, rather than increasing it on the contralateral side, consistent with our measured kinematics [32]. Thus, we cannot discount an effect of tethering on our insects, but their wing kinematics appear to be broadly representative of those used during free-flight.
A concern with using high-power X-rays to examine the biomechanics of the insect flight motor is that the radiation may affect the physiology of the insects during recording [45]. All four individuals continued flying after recording stopped, but although their measured wing kinematics fluctuated during recordings, there was no systematic change in the wing kinematics over the recording period (Figure 3). Stroke amplitude was bilaterally asymmetric throughout each recording, and was consistent with the asymmetry expected during a compensatory roll response, indicating that the flies were responsive throughout to the roll stimulus that we provided.
Further evidence of the consistency of the flies' behaviour is provided by the quality of the tomograms themselves, because the tomographic reconstruction process will only be successful if the pose of the sample is consistent within each group of radiographs. Any significant variation in steering muscle kinematics between wingbeats would result in blurring of the reconstructed tomograms, which each represent the average state of the flight motor at a given phase of the wingbeat. The edge detail of the rigid scutum had similar edge sharpness to the steering muscles (Figure 10), which indicates that the steering muscle kinematics were consistent through each recording.
Notwithstanding the consistency of their wing and muscle kinematics during the recordings, and the fact that the flies continued to fly immediately following exposure, all four individuals died a short while after. We therefore calculated the radiation dose received by the flies to assess the severity of exposure. Most of the X-rays produced by the beamline pass through the insects, but the amount will be dependent on both the individual (due to variation in size and hydration) and beam energy. We determined the proportion of X-rays absorbed by the insects by measuring the difference in image intensity between flat-beam images and radiographs where the insect was in the beam, using a region of interest containing the thorax, but not the mount. Using this method, we estimated that the mean absorption was 23% for the monochromatic beam and 13% for the polychromatic beam. The absorbed dose (D) was calculated as the absorbed power per unit mass:where a is the proportion of the beam absorbed by the insect, f is the beam flux, w is the width of the insect exposed to the beam (estimated from the radiographs to be 3.3 mm), h is the height of the beam, m is the mass of the insect (assumed to be 82 mg [20]), and t is the recording duration (4 s). The estimated total dose was 350 Gy for the monochromatic beam and 1,300 Gy for the polychromatic beam.
These total doses are similar to or less than the doses that have been applied to other insects in previous work without any measurable long-term effect [45]. However, our dose rates (90 Gy s−1 and 325 Gy s−1, for the monochromatic and polychromatic beam, respectively) were at least an order of magnitude higher than those used in previous work [45]. We therefore attribute the adverse effects of radiation following exposure to the high rate at which the dose was supplied.
|
10.1371/journal.pgen.1001227 | The DNA Damage Response Pathway Contributes to the Stability of Chromosome III Derivatives Lacking Efficient Replicators | In eukaryotic chromosomes, DNA replication initiates at multiple origins. Large inter-origin gaps arise when several adjacent origins fail to fire. Little is known about how cells cope with this situation. We created a derivative of Saccharomyces cerevisiae chromosome III lacking all efficient origins, the 5ORIΔ-ΔR fragment, as a model for chromosomes with large inter-origin gaps. We used this construct in a modified synthetic genetic array screen to identify genes whose products facilitate replication of long inter-origin gaps. Genes identified are enriched in components of the DNA damage and replication stress signaling pathways. Mrc1p is activated by replication stress and mediates transduction of the replication stress signal to downstream proteins; however, the response-defective mrc1AQ allele did not affect 5ORIΔ-ΔR fragment maintenance, indicating that this pathway does not contribute to its stability. Deletions of genes encoding the DNA-damage-specific mediator, Rad9p, and several components shared between the two signaling pathways preferentially destabilized the 5ORIΔ-ΔR fragment, implicating the DNA damage response pathway in its maintenance. We found unexpected differences between contributions of components of the DNA damage response pathway to maintenance of ORIΔ chromosome derivatives and their contributions to DNA repair. Of the effector kinases encoded by RAD53 and CHK1, Chk1p appears to be more important in wild-type cells for reducing chromosomal instability caused by origin depletion, while Rad53p becomes important in the absence of Chk1p. In contrast, RAD53 plays a more important role than CHK1 in cell survival and replication fork stability following treatment with DNA damaging agents and hydroxyurea. Maintenance of ORIΔ chromosomes does not depend on homologous recombination. These observations suggest that a DNA-damage-independent mechanism enhances ORIΔ chromosome stability. Thus, components of the DNA damage response pathway contribute to genome stability, not simply by detecting and responding to DNA template damage, but also by facilitating replication of large inter-origin gaps.
| Loss of genome integrity underlies aspects of aging and human disease. During DNA replication, two parallel signaling pathways play important roles in the maintenance of genome integrity. One pathway detects DNA damage, while the other senses replication stress. Both pathways activate responses that include arrest of cell cycle progression, giving cells time to cope with the problem. These pathways have been defined by treating cells with compounds that induce either replication stress or DNA damage, but little is known about their roles during unperturbed DNA replication. They may be important when several adjacent replication origins fail to initiate and forks from flanking origins must replicate longer regions of DNA than normal to complete replication. We have used a derivative of budding yeast chromosome III lacking all efficient replication origins to identify mutants that preferentially destabilize this chromosome fragment, which mimics a chromosome with a large inter-origin gap. We found that the DNA damage response pathway, but not the replication stress response pathway, plays an important role in maintaining this fragment. The signal recognized in this case may be replisome failure rather than forks stalled at endogenous DNA damage.
| In eukaryotic chromosomes, DNA replication initiates at multiple origins, specified by cis-acting sequences called replicators. In the budding yeast, Saccharomyces cerevisiae, replicators are termed ARS elements and were identified by their ability to promote extrachromosomal maintenance of plasmids. Chromosomal replication origins coincide with ARS elements, which contain the binding site for the six-subunit initiator complex, ORC. During G1, ORC recruits additional proteins to form pre-replicative complexes (pre-RCs) that initiate replication during S phase [1]. The average distance between active replication origins in S. cerevisiae is approximately 40 kb, based on both electron microscopic analysis of replicating DNA molecules [2] and whole genome analysis [3], [4]. In fission yeast, a similar range of estimates was obtained from whole genome analysis and DNA combing [3], [5].
The presence of multiple origins on chromosomes raises the question of whether replicators are activated according to a fixed temporal program or whether their use is stochastic, i.e. different replicators are activated in different cells or in successive S phases. In budding yeast, 2D-gel analyses and replication timing studies suggested that replicators are activated according to a program, although some variability is inevitable because some replicators fire inefficiently [3], [4], [6]–[12]. Recent single-molecule studies in budding yeast ([13], [14] Wang and Newlon, manuscript in preparation), and in fission yeast [5], [15]–[18] reflect this stochasticity in initiation.
Stochastic activation of replicators should occasionally produce large inter-origin gaps caused by failure of adjacent origins to initiate, referred to as the random gap problem [19]. Recent theoretical analysis of the replication dynamics of the fission yeast genome based on data that describe the positions and firing probabilities of replicators and the rate of fork movement suggests that long inter-origin gaps occur frequently in fission yeast [20]. In 88% of 2000 simulations using this stochastic hybrid model, at least one region of the genome contained an inter-origin gap more than 6-fold longer than the average inter-origin spacing; replication of such a gap would require about twice the known length of S phase. These results suggest that completion of DNA replication requires most of the normal G2 period of the cell cycle, and in some fraction of the population, regions of the genome would still be replicating at the normal time of mitosis. The problematic regions included about 5% of the genome, and each individual region appeared infrequently in the simulations, making such regions difficult to detect experimentally. It is not known how cells cope with this issue.
One possibility is that ongoing replication activates a checkpoint response to prevent cells from undergoing mitosis prior to completion of S phase. Two intertwined checkpoints function during S phase (Figure 1). The DNA damage response is activated by a signal transduction cascade in response to stalling of replication forks encountering DNA damage (reviewed by Branzei and Foiani [21], [22]). Experimentally, this response is activated by treatment with MMS or UV; unperturbed cells probably activate this pathway in response to forks encountering endogenous DNA damage. The replication stress response is activated experimentally by hydroxyurea treatment, which slows replication forks by inhibiting ribonucleotide reductase; it is not known what endogenous signal(s) activate(s) it.
Activation of an S phase response may occur in cells coping with long inter-origin gaps. Rad53p, the ortholog of the mammalian and fission yeast effector kinase, Chk2 (Figure 1), becomes hyperphosphorylated late in S phase in mutants that fail to fire some replication origins, indicating activation of a checkpoint [23], [24]. In addition the stability of a yeast artificial chromosome (YAC) carrying human DNA sequences from which origins had been deleted depended on RAD9, the mediator in the DNA damage response pathway [25]. However, other evidence suggests that cells do not monitor either the initiation or completion of DNA replication. For example, strains carrying tight alleles of cdc6 (cdc18+ in S. pombe), which encodes a pre-RC component, or of dbf4, the regulatory subunit of the Cdc7p kinase required for origin firing, proceed directly from G1 to mitosis despite failing to replicate any DNA [26]–[28]. Even ongoing replication may not prevent anaphase entry [29], [30].
We have created a derivative of yeast chromosome III lacking efficient replicators as a tool to detect mechanisms that contribute to replication of large inter-origin gaps (the 5ORIΔ-ΔR fragment - Figure 2). This fragment is composed entirely of yeast sequences with the exception of plasmid sequences at the fragmentation point. It replicates efficiently, with a loss rate per division of 2.1×10−3 [31] and is much more stable than the YAC [25]. We carried out a genetic screen for mutants specifically defective in maintenance of this ORIΔ derivative on the premise that mutations that caused destabilization of the 5ORIΔ-ΔR derivative, but had little or no effect on maintenance of the corresponding 0ORIΔ-ΔR derivative (Figure 2) would identify genes required for the replication of long inter-origin gaps, or perhaps new replication initiation mechanisms. This screen identified three originless fragment maintenance (Ofm) mutants, one dominant, OFM1-1, and two recessive, ofm6-1 and ofm14 (an allele of RAD9) [32]. The rad9 mutation increased the loss rate of the 5ORIΔ-ΔR fragment, but did not cause the frequent rearrangement that was seen with the YAC [32].
Here we report the results of a modified synthetic genetic array (SGA) screen [33], [34] used to identify additional Ofm mutants. Deletions of several genes in the DNA damage response pathway caused an Ofm phenotype. Further analysis indicated that this pathway contributes to the replication of large inter-origin gaps. In contrast, the replication stress response pathway does not contribute to the stability of the 5ORIΔ-ΔR fragment. Surprisingly, genes in the homologous recombination pathway, which are believed to contribute to the restart of collapsed replication forks, do not contribute to the maintenance of the fragment.
Our previous visual screen for Ofm mutants was labor intensive, during both the initial visual screening of colonies grown from the mutagenized culture and in subsequent attempts to identify mutations responsible for the phenotype. Thus, we adapted synthetic genetic array (SGA) technology [33], [34] for use in a colony sectoring screen to identify additional Ofm mutants in the S. cerevisiae viable deletion collection. One limitation of this screen is that essential genes could not be tested.
In the primary screen, as detailed in Methods, we used SGA technology to create ade2Δ::natR xxxΔ::kanR haploid MATa progeny. We then used chromoduction [35] to introduce the 5ORIΔ-ΔR fragment of chromosome III marked with ADE2 into each strain (Figure 2). Chromoductants, each carrying the 5ORIΔ-ΔR fragment were then streaked on plates with limiting adenine. Loss of the ADE2-marked fragment during growth of a colony results in a red sector. If a mutant has a low 5ORIΔ-ΔR fragment loss rate, such sectors will be rare; conversely, a mutant with an elevated loss rate will yield highly sectored colonies, providing a semi-quantitative estimate of loss rates. Examples of sectoring patterns are shown in Figure 3. The majority of the 5171 strains screened showed a low rate of sectoring as illustrated by the aro7Δ mutant used as a control. Ninety strains had an elevated rate of sectoring, as shown by the spe1Δ and ctf8Δ strains.
The elevated sectoring observed for the 90 strains selected from the primary screen could reflect either defects in transmission of all chromosomes, e.g. a defect due to the loss of a component of the kinetochore, or defects specific to 5ORIΔ-ΔR fragment transmission. To distinguish between these possibilities, we identified a colony from each of the 90 strains that had lost the 5ORIΔ-ΔR fragment, then separately introduced by chromoduction the 0ORIΔ-ΔR and the 5ORIΔ-ΔR fragments (Figure 2), and compared the sectoring phenotypes of two independent chromoductants carrying each of these fragments by estimating the number of red sectors per colony seen in chromoductants. Our previous measurements of loss rates of these chromosome III derivatives by fluctuation analysis allowed us to make semi-quantitative estimates of loss rates based on sectoring patterns [31], [32]. The loss rate of the 5ORIΔ-ΔR derivative is ∼2×10−3 losses per division in wild type cells, and this loss rate results in 0–3 sectors per colony in the SGA strain background. Colonies of strains carrying the 0ORIΔ-ΔR derivative, which has a loss rate of about 2×10−5 losses per division, rarely have a red sector. Mutant strains with loss rates of the 5ORIΔ-ΔR fragment in the range of 10−2 losses per division form colonies with 5–10 sectors per colony, and strains with loss rates in the range of 10−1 losses per division form colonies with ≥10 sectors per colony. The results of this secondary screen are detailed in Table S1. For example, spe1Δ was classified as an Ofm mutant because cells carrying the 5ORIΔ-ΔR fragment gave rise to colonies with 5–10 sectors per colony, while those carrying the 0ORIΔ-ΔR fragment yielded colonies that were rarely sectored; ctf8Δ was called a non-Ofm mutant because cells carrying either fragment gave rise to colonies with >10 sectors per colony (Figure 3). Overall, the 71 deletion strains in which the high sectoring phenotype of the 5ORIΔ-ΔR fragment was reproduced in the secondary screen were divided into high confidence Ofm mutants (52 strains), possible/probable Ofm mutants (14 strains) and non-Ofm mutants (5 strains) (Table 1). In the high confidence Ofm mutants, the two chromoductants carrying the 5ORIΔ-ΔR derivative were estimated to have at least 5–10 sectors per colony, and the two chromoductants carrying the 0ORIΔ-ΔR derivative rarely gave rise to a colony with a sector. In the case of the probable/possible Ofm mutants, either the two 5ORIΔ-ΔR chromoductants or the two 0ORIΔ-ΔR chromoductants showed different sectoring patterns. In the non-Ofm mutants, the 0ORIΔ-ΔR chromoductants all showed a sectoring pattern consistent with at least a 100-fold increase in the loss rate of this derivative.
A gene ontology (GO) analysis was performed on the 52 genes whose deletion caused Ofm phenotypes and on the 5 genes whose deletion caused non-Ofm phenotypes (http://db.yeastgenome.org/cgi-bin/GO/goTermFinder.pl). The three highest scoring clusters among the Ofm mutants (P = 8×10−5–4×10−3) share many genes and correspond to the annotations “cell cycle checkpoint”, “DNA damage response, signal transduction”, and “DNA damage checkpoint”. The cell cycle checkpoint cluster (SGS1, BFA1, MAD2, MAD3, RAD9, RAD17, and RAD24) included all of the genes present in the other two clusters. When the possible/probable Ofm mutants were included in the analysis the highest scoring cluster was still “cell cycle checkpoint” (p = 8×10−7). In addition to the 7 genes above, the cluster included BIM1, BUB1, BUB2, BUB3, CSM3 and TOF1. RAD9, RAD17, and RAD24 function in the DNA damage response pathway while MAD2 and MAD3 function in the spindle checkpoint, though some results have suggested an additional role in the DNA damage checkpoint [36]–[38]. BFA1 and BUB2 are required to prevent mitotic exit in both the DNA damage and spindle checkpoint pathways [39]. The highest scoring cluster (P = 3×10−6) among the non-Ofm mutants corresponded to the annotation “mitotic cell cycle”. This cluster included all five mutants identified as non-Ofm mutants.
Results of the GO analysis and identification of a null allele of RAD9 in our forward mutation screen [32] led us to examine the DNA damage response pathway in more detail. We moved the deletions of interest into the YKN10 strain background (Table S2) as described in Methods. Analysis of these strains allowed us to confirm that each deletion caused an Ofm phenotype and to quantitate the effects of the mutations in the strain background with which we had the most experience.
Our premise in undertaking this screen is that problems with the replication of the 5ORIΔ-ΔR derivative may be qualitatively different than the problems sustained by the 0ORIΔ-ΔR derivative by virtue of the presence of a long inter-origin gap. Therefore, we wanted to be able to make a quantitative comparison of loss rates that are very different. We reasoned that a comparison of the number of additional loss events sustained by the 5ORIΔ-ΔR and 0ORIΔ-ΔR derivatives in a given mutant would provide a measure of the strength of the Ofm phenotype. We define the “Ofm index” as the number of additional loss events per 105 divisions for the 5ORIΔ-ΔR derivative divided by the number of additional losses for the 0ORIΔ-ΔR derivative (Table 2). Two examples illustrate our reasoning. Suppose that in a wild type cell the loss rate of the 0RIΔ-ΔR derivative is 1 and the loss rate of the 5ORIΔ-ΔR derivative is 100. In one case, a mutation causes both derivatives to sustain an additional 400 loss events per 105 cell divisions. In this case the Ofm index = (500−100)/(401−1) = 1. This is the outcome we might expect for a mutation in a kinetochore component, and we would not consider the mutant to be an Ofm mutant. In another case, a mutation causes the 0ORIΔ-ΔR derivative to sustain 10 additional loss events and the 5ORIΔ-ΔR derivative to sustain an additional 900 loss events. In this case the Ofm index = (1000−100)/(11−1) = 90. We would consider this high Ofm index to indicate a specific defect in maintenance of the 5ORIΔ-ΔR fragment.
We first wished to distinguish the roles of the DNA damage and replication stress response pathways in the maintenance of the 5ORIΔ-ΔR derivative. In budding yeast, these pathways are best distinguished by the effects of mutations in the mediators because the pathways share both upstream and downstream components (Figure 1). The DNA-damage-specific mediator, Rad9p, an ortholog of mammalian 53BP1, is phosphorylated by the PIKK Mec1. Hyper-phosphorylated Rad9p binds the effector kinase Rad53p, an ortholog of Chk2, and facilitates both phosphorylation of Rad53p by Mec1p and activation of Rad53p kinase activity by autophosphorylation [40]–[44]. We previously found that both our original rad9 allele and the rad9Δ allele cause Ofm phenotypes, with mutants strains having Ofm indices of 81 and 65, respectively (Table 2 and [32]). These results indicate the DNA damage response pathway contributes to the maintenance of the 5ORIΔ-ΔR derivative.
The corresponding mediator in the replication stress response pathway is Mrc1p, a homolog of mammalian claspin. Mrc1p plays roles in both the replication stress response and normal replication fork progression [45]–[50]. Analysis of the role of MRC1 in the maintenance of the 5ORIΔ-ΔR derivative was complicated by its location on chromosome III and its dual role in S phase. We constructed both recipient and donor strains carrying the mrc1Δ allele; the 5ORIΔ-ΔR mrc1Δ and 0ORIΔ-ΔR mrc1Δ fragments were then separately transferred into the mrc1Δ recipient strain by chromoduction. Both 5ORIΔ-ΔR and 0ORIΔ-ΔR fragments were destabilized in the homozygous mrc1Δ strain, resulting in a low Ofm index (Table 2); the mrc1Δ strain is not an Ofm mutant. A deletion that removed the C-terminal half of the MRC1 ORF (the allele included in version 1 of the systematic deletion collection) caused a similar loss rate of the 5ORIΔ-ΔR fragment, but the 0ORIΔ-ΔR loss rate was about 10-fold lower than in the complete ORF deletion strain, suggesting that the N-terminus of Mrc1p may contribute to maintenance of the 0ORIΔ-ΔR fragment (data not shown).
To distinguish between the roles of the replication stress response and fork progression functions of Mrc1p in the maintenance of the 5ORIΔ-ΔR derivative, we made use of a separation of function allele, mrc1AQ, made by mutating six consensus Mec1p phosphorylation sites [47]; this allele lacks the replication stress response function of MRC1, but retains the fork progression function. Plasmids carrying either wild type MRC1 or mrc1AQ complemented the high loss rate of the 5ORIΔ-ΔR fragment in the mrc1Δ strain (Table 3). These results indicate that it is the loss of the fork progression function of Mrc1p that destabilizes the 5ORIΔ-ΔR fragment, not the loss of replication stress signaling. Therefore, mutations that impair DNA damage signaling, but not replication stress signaling, cause an Ofm phenotype.
We further tested the role of MRC1 in replication fork progression in our YKN10 background by examining the activation of dormant origins on chromosome III using 2D gel electrophoresis. These origins are inactive in the wild type strain because they are replicated by a fork from an adjacent origin before they can fire. Dormant origins can be activated by deletion of adjacent origins, which causes a delay in the time at which forks from the nearest remaining origins reach them, giving them an opportunity to fire [31], [51]. The dormant origin ARS304 is also activated in an mrc1Δ strain [49] in which forks progress slowly [48], [49]. To explore the generality of this phenomenon, we examined the activation of three dormant origins on chromosome III: ARS301, ARS304 and ARS314. As shown in Figure 4, replication initiation at ARS301 and ARS314, revealed by the presence of bubble-shaped intermediates, was detected in the mrc1Δ mutant, but not in the MRC1 strain; ARS304 was also active in the mutant (data not shown). Thus, activation of dormant origins is a general phenomenon in mrc1Δ strains that most likely reflects slow fork progression.
Deletions of other components of the DNA damage and replication stress response pathways also caused Ofm phenotypes. Deletions of genes encoding sensors shared by both pathways, including RAD17, which encodes a subunit of a PCNA-like clamp, and RAD24, which encodes the large subunit of its clamp loader (see Figure 2), caused Ofm phenotypes with Ofm indexes of 100 and 85, respectively (Table 2). Although it was not scored as a potential Ofm mutant in the primary screen, further examination revealed that deletion of DDC1, which encodes another subunit of the clamp, caused colonies of strains carrying the 5ORIΔ-ΔR fragment to sector similarly to the rad17Δ strain (Figure S1). Genes encoding other shared sensors were not examined because they are essential, including RFC2, RFC3, RFC4, RFC5, DDC2 and DPB11 (Figure 1).
Sensors activate PIKKs shared by both pathways. In S. cerevisiae, the ATR homolog, Mec1p, plays a more important role in the detection and repair of DNA damage than does the ATM homolog, Tel1p [52]. MEC1 is essential and was not in our screen; however the lethality caused by the mec1Δ allele can be suppressed by deletion of the ribonucleotide reductase inhibitor encoded by SML1 [53]. The sml1Δ mutation did not increase the loss rate of the 5ORIΔ-ΔR derivative, though it did slightly elevate the loss rate of the 0ORIΔ-ΔR derivative (Table 2). Since the sml1Δ strain is not an Ofm mutant, we examined mec1Δ in the sml1Δ background. The mec1Δ allele confers an Ofm phenotype indicating by its Ofm index of 40 (Table 2). The other PIKK, Tel1p, does not contribute to maintenance of the 5ORIΔ-ΔR fragment. The loss rate of this fragment in the tel1Δ mutant was 2.3±0.4×10−3 per division, similar to its loss rate in the wild type strain, and its loss rate in the mec1Δ tel1Δ double mutant was 1.3±0.2×10−2, similar to its loss rate in the mec1Δ mutant.
Downstream of the mediator, Rad9p, are the two effector kinases, Chk1p and Rad53p, homologues of the mammalian kinases, Chk1 and Chk2, respectively. The chk1Δ strain was not scored as a potential Ofm mutant in the primary screen; however, further examination revealed that this strain had an Ofm phenotype, with an Ofm index of 33 (Table 2). This result implicates Chk1p in transducing the signal from Rad9p to downstream targets. The rad53Δ mutant was not in the screen because it is inviable, but its inviability is suppressed by deletion of SML1. We found that the rad53Δ sml1Δ double mutant did not have an Ofm phenotype (Ofm index = 7) because the rad53Δ mutation caused an increase in the loss rate of the 0ORIΔ-ΔR fragment (Table 2). The increased loss rate of the 0ORIΔ-ΔR fragment in the rad53 strain indicates that Rad53p contributes to the maintenance of chromosomes with a normal complement of replication origins and is consistent with its well-documented role in response to DNA damage [22]. However, the loss rate of the 5ORIΔ-ΔR fragment was increased about 3-fold relative to the sml1Δ control, raising the possibility that Rad53p also contributes to the maintenance of this fragment. We examined the loss rate of the 5ORIΔ-ΔR fragment in a sml1Δ rad53Δ chk1Δ strain and found that its loss rate in the triple mutant was 880±140×10−5, approximately equal to the sum of the loss rates in the sml1Δ rad53Δ and chk1Δ mutants and nearly as high as the loss rates in strains carrying deletions of upstream components of the checkpoint pathway (Table 2). The Ofm index of the triple mutant was similar to that of the chk1 strain. Taken together, these results are consistent with the idea that Chk1p is primarily responsible for transducing the signal from Rad9p to downstream effectors, with Rad53p making a relatively small contribution to the maintenance of the 5ORIΔ-ΔR fragment as long as Chk1p is active, but becoming important in the absence of Chk1p.
RAD52 is required for virtually all homology-based double-strand break repair mechanisms, including break-induced replication and single-strand annealing (reviewed by Symington [54]). Our previous work showed that a rad52 mutant does not have an Ofm phenotype [31]; for this analysis we measured the stabilities of the 5ORIΔ-ΔR and 0ORIΔ-ΔR fragments (Figure 2) in wild type and rad52 strains in the CF4-16B strain background (Table S2), which differs slightly from the YKN10 background used in experiments summarized in Table 2. The 0ORIΔ-ΔR fragment was lost at a rate of 7×10−5 in the wild type strain and 9.5×10−4 in the rad52 strain, while the 5ORIΔ-ΔR fragment was lost at a rate of 1.5×10−3 in the wild type strain and 3.1×10−3 in the rad52 strain, leading to an Ofm index of 1.8 [31]. Confirming and extending these results, strains carrying deletions of ten genes in the RAD52 epistasis group (RAD50, RAD51, RAD52, RAD54, RAD55, RAD57, RAD59, RDH54, MRE11, and XRS2) all showed wild type sectoring in our primary screen (Figure S2). These results indicate that, in otherwise wild type strains, recombinational repair is not required for maintenance of ORIΔ chromosome derivatives.
By deleting the five efficient origins from the 5ORIΔ-ΔR fragment, we altered both the positions at which replication most likely initiates and the distances that individual replication forks travel. The high loss rates of the 5ORIΔ-ΔR fragment seen in the DNA damage response mutants could result from difficulty in initiating replication, difficulty in replication fork progression, or both. To address this issue, we examined stabilities of two additional derivatives of chromosome III, the full-length 5ORIΔ chromosome and the ΔL-6ORIΔ fragment (Figure 2), in these mutants. The 5ORIΔ-ΔR fragment used in our mutant screen is truncated to the right of the ARS310 deletion. The full-length 5ORIΔ chromosome carries the same deletions of the five efficient origins as the 5ORIΔ-ΔR fragment, but retains origins distal to the ARS310 deletion; the inefficient origin, ARS313, is located about 20 kb distal to the ARS310 deletion, and the efficient origin, ARS315, is located about 50 kb distal [55]. This derivative is as stable as the 0ORIΔ-ΔR derivative in the wild type strain and the sml1Δ mutant. The ΔL-6ORIΔ fragment was derived from the full-length 5ORIΔ chromosome by removing the centromere-associated inefficient origin, ARS308, and fragmenting the chromosome to the right of ARS304, which removed ARS304, the dormant origins associated with HML and the left telomere. This derivative is as stable as the 5ORIΔ-ΔR derivative in the wild type strain and the sml1Δ mutant (Table 2). In both 5ORIΔ and ΔL-6ORIΔ derivatives, the origin-deleted region to the left of ARS313 can be replicated by forks that initiate at ARS313 or at origins further to the right. In 5ORIΔ, but not in ΔL-6ORIΔ, there also exists the potential for the origin-deleted region to be replicated by forks that initiate at one of the normally-dormant HML-associated origins.
If a mutant has an initiation defect, then the presence of additional origins on 5ORIΔ and ΔL-6ORIΔ derivatives should suppress the Ofm phenotype. Conversely, if a fork progression defect creates difficulty in completing replication of a large inter-origin gap, the presence of additional origins should not suppress the defect. The ΔL-6ORIΔ fragment provides a particularly stringent test of fork progression and/or fork stability, because a collapsed leftward-moving fork initiated at ARS313 or ARS315 cannot be rescued by a fork initiated at one of the HML-associated dormant origins.
We first examined the stability of these larger gapped constructs in the mrc1Δ mutant because it has a known fork progression defect [48], [49]. MRC1 was deleted from the full-length 5ORIΔ chromosome to avoid complementation; MRC1 is distal to ARS304 so, like the dormant origins, it is absent from the ΔL-6ORIΔ fragment. In mrc1Δ mutants, loss rates of the full-length 5ORIΔ chromosome and the ΔL-6ORIΔ fragment were similar, and were about 2.5-fold lower than the loss rate of the 5ORIΔ-ΔR fragment (Figure 5, Table 2). These results are consistent with our expectation that the additional origins on these two derivatives would not suppress the fork progression defect of mrc1Δ. Activation of HML-associated dormant origins does not appear to contribute to the stability of the full-length 5ORIΔ chromosome in the absence of Mrc1p, because the ΔL-6ORIΔ fragment, which lacks HML-associated dormant origins, showed a loss rate similar to 5ORIΔ. The 2.5-fold higher rate of loss of the 5ORIΔ-ΔR fragment likely reflects the fact that replication of this fragment is at least partially dependent upon activation of HML-associated dormant origins, and that these origins are less efficient than the origins present on the right arm in the full-length 5ORIΔ chromosome and the ΔL-6ORIΔ fragment (Figure 5).
Consistent with the observation of Cha and Kleckner [56] that Mec1p stabilizes forks in slow replication zones, we found that the mec1Δ mutant behaved similarly to the mrc1Δ mutant. The 5ORIΔ chromosome was unstable in a mec1Δ strain (Figure 5, Table 2), suggesting a fork progression defect. The loss rate of the ΔL-6ORIΔ fragment was less than three-fold higher than that of the full-length 5ORIΔ chromosome, suggesting that the HML-associated dormant origins make only a small contribution to the stability of the full-length 5ORIΔ chromosome in the absence of Mec1p.
Results obtained with the rad9 and rad24Δ mutants contrasted sharply with the mrc1Δ and mec1Δ results. The full-length 5ORIΔ chromosome was substantially more stable than 5ORIΔ-ΔR or ΔL-6ORIΔ in the absence of Rad9p or Rad24p, with loss rates about 40-fold lower than the 5ORIΔ-ΔR fragment and only two-fold higher than the 0ORIΔ-ΔR fragment (Table 2 and Figure 5). By contrast, in mrc1Δ and mec1Δ strains, the full-length 5ORIΔ chromosome is 10- to 20-fold less stable than 0ORIΔ-ΔR.
The relative stability of 5ORIΔ-ΔR in rad9 and rad24Δ mutants might indicate that the presence of efficient origins to the right of the origin-deleted region could suppress the Ofm phenotype of these mutants. If this were the case, then the loss rate of the ΔL-6ORIΔ fragment should also be low. However, the loss rates of this fragment were as high as or higher than the 5ORIΔ-ΔR fragment in both mutants. The high loss rates of both the 5ORIΔ-ΔR fragment and the ΔL-6ORIΔ fragment indicate that maintenance of the full-length 5ORIΔ chromosome in rad9 and rad24Δ strains requires the presence of replication origins on both sides of the ORIΔ gap, and suggest that a single fork cannot traverse the gap in these strains.
One explanation for the lower stability of the ΔL-6ORIΔ fragment in the rad9Δ strain than in a mec1Δ sml1Δ strain is that in the absence of Rad9p, Mec1p kinase activity is deleterious. If this were the case, the loss rate of ΔL-6ORIΔ fragment in a rad9Δ mec1Δ sml1Δ triple mutant should be the same as in the mec1Δ sml1Δ strain. Alternatively, a second pathway, possibly Tel1p-dependent, could activate Rad9p in the absence of Mec1p, or Rad9p could have a DNA-damage-response-independent function that contributes to the maintenance of the ΔL-6ORIΔ fragment. In both of these cases, the triple mutant should have a loss rate similar to the rad9Δ strain. The loss rates of the ΔL-6ORIΔ fragment were 6.9±0.6×10−2 in a rad9 sml1Δ strain and 5.1±0.4×10−2 in a rad9 mec1Δ sml1Δ strain, suggesting that a second pathway activates Rad9p. Alternatively Rad9p has a function that is independent of its role in the DNA damage response pathway in maintenance of the ΔL-6ORIΔ fragment (see Discussion).
Finally, the behavior of the ΔL-6ORIΔ fragment in the effector kinase mutants provides strong support for idea that Rad53p becomes important for the maintenance of ORIΔ chromosomes in the absence of Chk1p. The loss rates of the ΔL-6ORIΔ derivative in the chk1Δ and rad53Δ sml1Δ strains were similar and elevated approximately 2-fold relative to the 5ORIΔ-ΔR derivative. The loss rate in the chk1Δ rad53Δ sml1Δ mutant was approximately 10-fold higher and was equal to the very high loss rate seen in the rad9Δ mutant (Table 2).
The loss rate of the full-length 5ORIΔ chromosome was much higher in the mrc1Δ and mec1Δ strains than in the rad24Δ and rad9 strains. It appears that the dormant origins associated with HML near the left end of the full-length 5ORIΔ chromosome contribute to the maintenance of this chromosome in wild-type, because derivatives truncated to remove HML-associated dormant origins showed higher loss rates than derivatives containing them (Figure 5,Table 2 and [31]). Increased activation of these dormant origins in rad9 and rad24Δ, as compared to in mrc1Δ and mec1Δ, could explain the differences in stability of the full-length 5ORIΔ chromosome in these two sets of mutants. Therefore, we examined the activation of the dormant origins ARS301, ARS302/ARS303/ARS320 (three closely-spaced ARS elements), and ARS304 on the full-length 5ORIΔ fragment by 2D gel analysis (Figure 6). Both bubble- and Y-shaped replication intermediates were detected at ARS301 in mec1Δ and rad9Δ strains, indicating that this origin is activated in a subset of the cells in both strains. A fortuitous restriction-site polymorphism allowed us to distinguish the signal arising from the balancer chromosome from that arising from the 5ORIΔ chromosome. Bubble-shaped intermediates were detected only in strains where the 5ORIΔ chromosome was present, indicating that ARS301 fires only on the 5ORIΔ chromosome. Similarly, we found bubble arcs arising from the ARS302/ARS303/ARS320 cluster in mec1Δ and rad9Δ strains, but only when the 5ORIΔ chromosome was present. ARS304 was not detectably active in either mutant (Figure 6). In all cases, the intensity of the bubble arc was less than that of the Y arc, indicating that in the majority of cells each ARS was passively replicated.
We quantitated the percent of bubble-shaped replication intermediates produced by the 5ORIΔ chromosome, using two approaches to quantitate the signal (Methods and Table S3). ARS301 initiated replication in 1.7–6.3% of the population in the rad9Δ strains, and in 7.4–14.6% of the population in mec1Δ sml1Δ strains. The range of values for the ARS302/ARS303/ARS320 cluster was similar, 2.3–7.6% in rad9Δ strain and 9.4–15.7% in the mec1Δ sml1Δ strains. Thus the dormant replicators are 2-to 3-fold more active in mec1Δ strains than in rad9Δ strains, indicating that the higher stability of the 5ORIΔ chromosome in rad9Δ strains cannot be explained by increased activation of dormant origins.
The activation of HML-associated origins in the rad9Δ strain may account for the differences in stability of the 5ORIΔ chromosome and the ΔL-6ORIΔ derivative. The HML-associated origins fire only late in S phase [51], [57]. Leftward-moving forks normally reach them before they are programmed to fire. In the rad9Δ strain, approximately 10% of cells activate either ARS301 or the ARS302/ARS303/ARS320 cluster in the full length 5ORIΔ chromosome. About 10% of rad9 cells lose the ΔL-6ORIΔ fragment (Table 2), suggesting that about 10% of the forks initiated to the right of the gap fail to traverse the gap in rad9 mutants. In this situation, the ΔL-6ORIΔ fragment, which lacks the HML-associated dormant origins, would be lost as a result of incomplete replication. In contrast, only 0.03% of rad9 cells lose the 5ORIΔ chromosome (Table 2) because, in the 10% of cells in which leftward-moving forks fail to traverse the gap, firing of one of the HML-associated dormant origins allows the replication of this chromosome to be completed.
Unlike the 5ORIΔ-ΔR fragment and the full-length 5ORIΔ chromosome, the ΔL-6ORIΔ fragment was structurally unstable. Stable derivatives that had lost the cloNAT-resistance marker present at left-hand end of the fragment (Figure 2) arose in the rad9, rad24, and mec1 mutants. The rates of production of these stable derivatives were similar to the loss rates of the ΔL-6ORIΔ fragment measured in these strains (Table S4). Twelve stable derivatives of the ΔL-6ORIΔ fragment produced by the rad9 strain migrated on pulsed-field gels with the full-length balancer chromosome, suggesting that chromosome III sequences distal to the fragmentation point had been restored (data not shown). One possible mechanism for the production of these stable derivatives is that replication forks collapse and are processed into double-strand breaks that are repaired by break-induced replication [58] using the balancer chromosome as a template.
We employed a novel modification of the SGA method to screen for mutations that preferentially destabilize a chromosome III derivative lacking efficient replication origins. The modification utilized a single chromosome transfer technique, chromoduction, to transfer the 5ORIΔ-ΔR fragment into an ordered array of the viable ORF deletion collection. Yuen et al. [59] carried out similar colony-sectoring screens of the viable deletion collection using two chromosome fragments. Of the 66 chromosome transmission fidelity (ctf) mutants identified in these screens, 14 were also identified in our screen. As expected, given that the ctf mutants were identified using chromosome fragments carrying a normal complement of replication origins, the majority of the ctf mutants we re-identified were found in the non-Ofm or possible/probable Ofm classes. The two scored as Ofm mutants are ctf18Δ and mad2. It seems likely that many ctf mutants were not identified in our screen because they caused only small increases in the rate of loss of the 5ORIΔ-ΔR fragment. Approximately 60% of the loss rates measured for chromosome fragments in ctf mutants are less than or equal to the loss rate of the 5ORIΔ-ΔR fragment; increases of that magnitude would not have been detected in our visual screen.
Our results indicate that the DNA damage signaling pathway, but not the replication stress signaling pathway, contributes to the maintenance of the 5ORIΔ-ΔR fragment. While the DNA damage and replication stress response pathways share many components (Figure 1), mutation of the DNA-damage-tocheckpoint-signaling mediator, Rad9p, preferentially destabilized the 5ORIΔ-ΔR fragment, but a checkpoint-deficient mutation in the replication-stress-specific signaling mediator, Mrc1p, did not. Mutations in many of the shared signaling components also caused Ofm phenotypes.
We found unexpected differences in the contributions that the DNA damage signaling pathway makes to maintenance of ORIΔ chromosome derivatives and the contributions that it makes to DNA damage resistance. First, the DNA damage signaling pathway detects and stabilizes forks stalled at sites of damage and facilitates repair or bypass of the damage; studies with DNA damaging agents [60]–[63] indicate that this function is more strongly dependent on RAD53 than on CHK1. Based on the results presented here, the DNA damage signaling pathway also contributes to the replication of large inter-origin gaps, which can arise when several adjacent origins fail to fire. Such gaps appear commonly during the replication of the rDNA array [14]. The 5ORIΔ-ΔR fragment, the full-length 5ORIΔ chromosome and the ΔL-6ORIΔ fragment mimic these large gaps, and the pathways identified by the Ofm mutants may have arisen to facilitate the replication of large inter-origin gaps. Interestingly, this function appears to be facilitated primarily by CHK1 with a contribution from RAD53 evident in the absence of CHK1.
Second, we found that mec1Δ and mrc1Δ mutations have different effects than rad9Δ and rad24Δ mutations on the stabilities of the ΔL-6ORIΔ and full-length 5ORIΔ derivatives. Dormant origins near the left end of chromosome III are more strongly activated in a mec1Δ mutant than in a rad9 mutant (Figure 6), suggesting that in the mec1Δ strain the HML-associated dormant origins have more time to fire. However, removing the dormant origins, as in the ΔL-6ORIΔ fragment, caused a 16-fold greater increase in the rate of chromosome loss in the rad9 strain than in the mec1Δ strain (Table 2), suggesting that forks fail to reach the left end more often in the rad9 strain. One explanation for this disparity is that an alternative pathway activates Rad9p in mec1Δ cells, which results in stabilization of replication forks and allows them to progress, albeit slowly, in the absence of Mec1p [56]. In mec1 mutants, we suggest that slow fork progression through the long ARS305Δ – ARS310Δ gap allows time for the activation of either ARS301 or the ARS302/ARS303/ARS320 cluster in ∼20% of the cells (Figure 6). However in the absence of the dormant origins, as in the ΔL-6ORIΔ fragment, these slow-moving forks are able to complete replication through the gap to the telomere in >99% of the cells (Table 2). Tel1p, which is also a PIKK, is a candidate for activation of Rad9p, in this situation. However, our observation that Tel1p did not contribute to the stability of the 5ORIΔ-ΔR fragment in either the presence or absence of Mec1p (see Results) suggests that Tel1p does not contribute to this pathway. In the absence of Rad9p, we suggest that forks initiated to the right of the ARS305Δ – ARS310Δ gap simply fail to traverse the gap approximately 10% of the time (Table 2), and that, in the absence of the dormant origins, replication of the chromosome is not completed, leading to segregation of the partially replicated molecule and chromosome loss. An alternative explanation for the disparity is that Rad9p has a function that is independent of its role in the DNA damage response pathway.
Finally, we found that strains carrying deletions of ten genes in the RAD52 epistasis group did not show elevated loss rates of the 5ORIΔ-ΔR fragment. Since genes in this epistasis group are required for all homology-dependent repair processes, including double-strand break repair, break-induced replication and replication fork restart, these results suggest that replication of this ORIΔ derivative does not require repair of DNA damage or double-strand breaks.
Our favored model for the role of the DNA damage response pathway in the replication of ORIΔ chromosome derivatives is based on the idea that replication forks age, i.e. that the probability of fork arrest due to failure of a replisome component increases with the distance the fork has traveled. We refer to these forks as crippled, to distinguish them from forks that are stalled (arrested by DNA damage or nucleotide depletion with replisome intact) or collapsed (replisome disassembled), and to reflect the need for some replisome component to be replaced or modified in order to continue elongation. These crippled forks are then recognized and restored by a RAD9-and CHK1-dependent pathway. The restart of these crippled forks is independent of homologous recombination because there is no DNA damage to be bypassed, and, therefore, double-strand breaks are therefore not formed. If a fork were arrested due to failure of a replisome component, there would be no impediment to elongation once the replisome is reconstituted.
There are alternative models to explain the role of the DNA damage response pathway in maintaining the 5ORIΔ-ΔR fragment, which has a large inter-origin gap. The simplest is that the DNA damage response monitors the completion of replication. However, the evidence for such a checkpoint is not compelling (see Introduction). Debate over the existence of a replication completion checkpoint is ongoing; our observations provide only circumstantial evidence in favor of such a checkpoint.
Another model to explain the role of the DNA damage response pathway in maintaining fragments with large inter-origin gaps suggests that forks stall at sites of endogenous DNA damage and are stabilized by this pathway. The 5ORIΔ-ΔR and ΔL-6ORIΔ fragments would be especially sensitive to such events in the absence of the DNA damage response because the stalled forks would collapse. In the case of the 0ORIΔ-ΔR fragment, which has a full complement of replication origins, a collapsed fork could be rescued by a converging fork from an adjacent origin. In contrast, the 5ORIΔ-ΔR fragment has fewer initiation events, so a collapsed fork would be rescued less often by a converging fork, resulting in an elevated loss rate in a DNA damage checkpoint mutant. Consistent with this suggestion, our analysis of individual 5ORIΔ-ΔR molecules in wild type cells suggests that replication initiates at only one or two places per molecule, but at different places on different molecules (Wang et al., manuscript in preparation).
The enhanced stability of the full-length 5ORIΔ chromosome compared to the ΔL-6ORIΔ fragment in the rad9 and rad24Δ mutants is also consistent with this endogenous damage model, as a collapsed leftward-moving fork in the 5ORIΔ chromosome can be rescued by a fork initiating at one of the dormant origins near HML. Our finding that mec1Δ confers an Ofm phenotype while tel1Δ does not is also consistent with this model because MEC1 plays a more important role in the tolerance of DNA damage than does TEL1 [52].
However, this endogenous damage model is challenged by findings that fork stabilization at sites of DNA damage and survival are more strongly dependent on RAD53 than on CHK1 [60]–[66], whereas CHK1 makes a more important contribution than RAD53 to 5ORIΔ-ΔR fragment maintenance, suggesting that the DNA damage response is not simply stabilizing forks in response to damage. While Segurado and Diffley [61] have suggested a role for CHK1 in stabilizing replication forks, that function was detected only in the absence of both RAD53 and EXO1, which encodes a nuclease responsible for fork collapse in the absence of RAD53. Thus, it seems unlikely that this explains the contribution of CHK1 to 5ORIΔ-ΔR fragment maintenance. Another problem is that deletions of genes, whose products are required for mismatch repair, repair of UV damage, and homologous recombination, did not increase the loss rate of the 5ORIΔ-ΔR fragment in the primary screen, as would have been expected if DNA damage-provoked fork collapse was responsible for loss of this fragment.
Replication fork aging also suggests an explanation for the close spacing of replication origins in S. cerevisiae. A median inter-origin distance of 36 kb was estimated from visualization of replicating molecules by electron microscopy (reviewed by Newlon [67]), and a similar median distance, 34 kb, was estimated using the genome-wide replication timing data of Raghuraman et al. [4]. Based on a median fork rate of 2.3 kb per minute and an S phase of 55 minutes [4], a single fork from the earliest firing origin would be able to replicate ∼120 kb and a fork from an origin activated in the middle of S phase would be able to replicate ∼60 kb. Thus, origins are spaced more closely than predicted by the median origin activation time and rate of fork movement. The observed high density of origins may insure that gaps too long to be reliably replicated do not occur, even if several adjacent origins fail to fire.
Pan et al. described a DNA Integrity Network of 78 genes on the basis of synthetic fitness or lethality defects [68]. Sixteen of these genes are believed to have roles in S phase checkpoints. Deletions of eight of these genes cause an Ofm phenotype: RAD9, RAD17, RAD24, CTF18, MEC1, DDC1, CHK1, and RAD53. Deletions of two other genes in this group, csm3Δ and tof1Δ, were scored possible Ofm mutants.
In addition to the checkpoint genes, our Ofm mutants included deletions of two other genes from the DNA Integrity Network, HST3 and POL32, both of which have links to DNA replication. HST3 encodes a NAD+-dependent histone H3 lysine-56 deacetylase [69]–[71]. Our analysis of hst3 mutants will be presented elsewhere; it indicates that the Ofm phenotype of hst3Δ results from a fork progression defect (Irene et al. manuscript submitted). pol32Δ mutants, which lack a nonessential subunit of DNA polymerase Δ, also show fork progression defects, which may explain their Ofm phenotype [72]–[75].
In summary, we have identified a set of genes whose products facilitate replication of large inter-origin gaps. This set is enriched in components of the DNA damage and replication stress signaling pathways. Replication of large inter-origin gaps shows several surprising features: Dependence on the DNA-damage-specific mediator, Rad9p, rather than the replication-stress-specific mediator, Mrc1p; a stronger dependence on the effector kinase, Chk1p than Rad53p, and no dependence on homologous recombination
Yeast strains are listed in Table S2. All strains are isogenic with YPH499 [76], except the full-length and fragmented chromosome donor strains, which are in the CF4-16B background [31], and YJT242 (and its parent Y7029) and the viable ORF deletion collection, which are related to S288C [77]. SGA selection media were prepared as described in [78]. Chromoductants for the SGA screen were selected on -Ade -Leu -Lys -Arg dropout plates containing 60 µg/ml canavanine (Sigma) and 10 µg/ml thialysine (Sigma). Chromoductants in the YKN10 background were selected on -Leu-Trp -Arg dropout plates containing 60 µg/ml canavanine and 10 µg/ml cycloheximide (Sigma), except that chromoductants of the ΔL-6ORIΔ fragment were selected on -Leu -Ade -Arg dropout plates containing 100 µg/ml CloNAT (Werner Bioagents, Germany), 60 µg/ml canavanine, and 10 µg/ml cycloheximide. Limiting adenine medium was purchased from US Biologicals.
YJT242 was created by transforming Y7029 with a PCR product carrying the natMX gene, amplified from pAG25 [79], flanked by homology to the ADE2 locus; sequences of primers are available upon request. Individual G418-resistant knockouts were moved into the YKN10 background by transformation with a PCR product amplified from the appropriate strain from the ORF deletion collection (Open Biosystems) using the locus specific A and D primers (www-sequence.stanford.edu/group/yeast_deletion_project/Deletion_primers_PCR_sizes.txt). The mrc1Δ::NAT allele was introduced into the YKN10 background using primers and a template generously provided by K. Sugimoto (UMDNJ). This allele was converted to mrc1Δ::KAN by transforming YJT294 with NotI-cut pFA-KanMX4 [80] and selecting for G418-resistance yielding YJT551. The his3-Δ367 alleles were generated by fusion PCR and introduced by two-step gene replacement [81]. Primers are available upon request. The bar1-Δ1327 allele carries a BglII-BsrGI deletion that removes 1327 bp within the open reading frame.
In our version of the screen, a strain carrying an ade2Δ::natMX mutation, which causes the accumulation of a red pigment in colonies grown on limiting adenine and confers nourseothricin resistance, was mated to the array of viable deletion mutants, each marked with kanMX, which confers G418 resistance. The resulting diploids were then sporulated, and double mutant ade2Δ::natR xxxΔ::kanR haploid MATa progeny were selected. The array of double-mutant strains was mated to F510αA1–4, the donor strain, carrying the 5ORIΔ-ΔR derivative of chromosome III marked with ADE2 (Figure 2). Because the donor strain carries the kar1-Δ15 mutation, normal karyogamy is inhibited, resulting in inefficient production of diploid cells [82]. During the transient heterokaryon stage, single chromosomes are transferred at low frequency between the two nuclei, a process called chromoduction [35]. The strains were marked to allow selection for rare chromoduction events in which the 5ORIΔ-ΔR fragment was transferred into the ade2Δ::natR xxxΔ::kanR nucleus. The 5ORIΔ-ΔR fragment carries LEU2 at its endogenous locus and an ectopic copy of ADE2 inserted near the ARS307 deletion (Figure 2). The corresponding donor strain carrying the 5ORIΔ-ΔR fragment is Leu+ and Ade+, but canavanine-sensitive and thialysine-sensitive because it carries the wild type CAN1 and LYP1 alleles. The double mutant (ade2Δ::natR xxxΔ::kanR) strains generated by SGA analysis are Leu−, Ade−, canavanine-resistant, and thialysine-resistant. Any diploids that form between the donor strain and the ade2Δ::natR xxxΔ::kanR double mutant strains are Leu+ and Ade+, but canavanine-sensitive and thialysine-sensitive because the can1Δ and lyp1Δ mutations are recessive. The desired chromoduction event results in cells that are Leu+ and Ade+, because they carry the 5ORIΔ-ΔR fragment, and canavanine- and thialysine-resistant, because they carry the can1Δ and lyp1Δ mutations. Medium lacking leucine, adenine, arginine, and lysine and containing both canavanine and thialysine selects for these cells. A preliminary screen using approximately 100 strains selected from the viable deletion collection was carried out to determine conditions for the chromoduction. We found that pinning the array of double mutants at the density found in a standard 384 well plate was necessary to ensure efficient mating of the donor strain to the array.
The screen was done in duplicate, and chromoductants from the duplicate arrays were streaked side-by-side on a single plate with limiting adenine for scoring sectoring patterns (see Table S1). This process was completed in less than three months by eight individuals, demonstrating the feasibility of including a chromoduction step in the SGA procedure to transfer a single chromosome or plasmid into the double mutant array. If the phenotype of chromoductants could be scored directly on selective medium, then the entire procedure could be accomplished with robots.
Chromosome loss rates were determined by fluctuation analysis using the colony isolation method [83]. Red colonies were tested for leucine and tryptophan auxotrophies to distinguish chromosome losses from gene conversions or mitotic recombination events; leucine auxotrophy and nourseothricin-resistance were used in fluctuations involving the ΔL-6ORIΔ fragment. The presence of origin deletions was confirmed by PCR. Loss rates were calculated using the method of Lea and Coulson [84].
Genomic DNA was prepared from log-phase cultures as described [85], digested with either NdeI, ClaI+EcoRV, or FspI+SphI+ClaI, subjected to BND-cellulose (Sigma) chromatography, electrophoresed on neutral-neutral 2D gels, blotted, and hybridized as described [86]. The probe for ARS301 was the1.3-kb EcoRI-XhoI fragment from p78_4.6; the probe for ARS302/ARS303/ARS320 was the 1.9-kb EcoRI-HindIII fragment from p78_5.2; the probe for ARS304 was the 3.5-kb PshAI-BamHI fragment from D10B; the probe for ARS314 was the 1.8-kb HindIII fragment from pH 1.8 [55], [87]. These fragments were labeled with [α-32P] dATP (Perkin Elmer) using the Megaprime DNA-labeling system (GE Healthcare). Images were acquired on a Molecular Dynamics Typhoon 9410, and the exposure was adjusted using ImageQuant 5.2 software. Quantitations of bubble-shaped and Y-shaped replication intermediates were determined using the polygon tool and the line tool of ImageQuant 5.2.
Colonies were photographed after ∼5 days of growth at 30°C on limiting adenine plates. Images were acquired as TIFF files with a Nikon D-100 camera fitted with an AF Micro-Nikkor 60 mm f/2.8 D lens. Images were cropped and adjusted for color balance and brightness/contrast in Photoshop.cs v8.0 (Adobe Systems).
|
10.1371/journal.pgen.1007517 | Sequence features governing aggregation or degradation of prion-like proteins | Enhanced protein aggregation and/or impaired clearance of aggregates can lead to neurodegenerative disorders such as Alzheimer’s Disease, Huntington’s Disease, and prion diseases. Therefore, many protein quality control factors specialize in recognizing and degrading aggregation-prone proteins. Prions, which generally result from self-propagating protein aggregates, must therefore evade or outcompete these quality control systems in order to form and propagate in a cellular context. We developed a genetic screen in yeast that allowed us to explore the sequence features that promote degradation versus aggregation of a model glutamine/asparagine (Q/N)-rich prion domain from the yeast prion protein, Sup35, and two model glycine (G)-rich prion-like domains from the human proteins hnRNPA1 and hnRNPA2. Unexpectedly, we found that aggregation propensity and degradation propensity could be uncoupled in multiple ways. First, only a subset of classically aggregation-promoting amino acids elicited a strong degradation response in the G-rich prion-like domains. Specifically, large aliphatic residues enhanced degradation of the prion-like domains, whereas aromatic residues promoted prion aggregation without enhancing degradation. Second, the degradation-promoting effect of aliphatic residues was suppressed in the context of the Q/N-rich prion domain, and instead led to a dose-dependent increase in the frequency of spontaneous prion formation. Degradation suppression correlated with Q/N content of the surrounding prion domain, potentially indicating an underappreciated activity for these residues in yeast prion domains. Collectively, these results provide key insights into how certain aggregation-prone proteins may evade protein quality control degradation systems.
| Protein aggregation is associated with a variety of diseases, including Alzheimer’s disease and Amyotrophic Lateral Sclerosis. Cells possess a number of factors that can recognize aggregation-prone protein features and prevent aggregation. One common way this is achieved is through the pre-emptive degradation of aggregation-prone proteins. While considerable progress has been made in understanding how the amino acid sequence of a protein relates to intrinsic aggregation propensity, little is known about how aggregation-prone proteins avoid intracellular anti-aggregation systems. We used a genetic screen in yeast to define sequence features of aggregation-prone domains that lead to degradation or prion aggregation as it occurs in the context of eukaryotic protein quality control factors. Unexpectedly, we found that only a subset of aggregation-promoting amino acids could effectively stimulate degradation of an aggregation-prone domain. Furthermore, this degradation-promoting effect could be suppressed by classical prion domain features. Our results highlight the complex interplay between pre-emptive protein degradation and protein aggregation, and implicate the unusual composition of yeast prion domains in preventing their degradation.
| Protein misfolding disorders involve the conversion of native proteins into non-native, deleterious forms. Some misfolded proteins form highly ordered amyloid aggregates, stabilized by intermolecular cross-β sheets. Once formed, these aggregates can convert remaining soluble proteins to the aggregated form via a templated misfolding mechanism [1]. Harmful aggregates must be prevented, sequestered, disassembled, or degraded by cells to prevent disruption of essential cellular functions. Enhanced protein aggregation or impaired clearance of aggregates can lead to neurodegenerative disorders such as Alzheimer’s Disease, Parkinson’s Disease, Amyotrophic Lateral Sclerosis (ALS), and Huntington’s Disease (for review, see [2–9]).
Prion diseases represent a unique sub-class of protein misfolding disorders in which protein aggregates are infectious. Prions can arise de novo through protein misfolding events that convert native proteins into the infectious form, or may be acquired through environmental encounter with the infectious form [10]. Although first described in mammals, a number of prion proteins were later found to occur in budding yeast [11, 12].
Saccharomyces cerevisiae has been used extensively as a model organism to study prions [11, 13]. Discovery and characterization of the first two yeast prion proteins, Ure2 and Sup35, revealed that both proteins contain remarkably glutamine/asparagine (Q/N) rich prion domains [12, 14, 15]. The prion domains also contain relatively few charged and hydrophobic residues. Scrambling experiments demonstrated that the ability of Ure2 and Sup35 to form prions is largely dependent on the amino acid composition of the prion domains, rather than the primary amino acid sequence [16, 17]. Methods for scanning the yeast proteome for additional proteins with similar compositional features resulted in successful identification of new yeast prions [18–20]. To date, nine yeast proteins have been demonstrated to form aggregation-mediated prions [12, 18, 21–27]. The majority of these proteins also contain prion domains with high Q/N content and low charged/hydrophobic content.
Examination of the human proteome with more sophisticated composition-based search algorithms revealed a number of human proteins with “prion-like domains” (PrLDs), defined as domains that compositionally resemble yeast prion domains [5, 28]. Many of the top candidates (including TDP-43 [29, 30], FUS [31, 32], EWSR1 [33, 34], TAF15 [34–36], hnRNPA1 [37], hnRNPA2B1 [37], and TIA1 [38]) have been implicated in protein misfolding disorders. In addition to containing PrLDs, aggregates formed by these proteins are thought to spread throughout an individual in an infectious prion-like manner along a neuroanatomical path that parallels the progression of pathological symptoms [3, 39]. Furthermore, the PrLDs from hnRNPA1 and hnRNPA2B1 are able to support prion activity when substituted in place of the portion of the Sup35 prion domain that is responsible for nucleating prion activity [37, 40].
Although composition-based algorithms have been reasonably effective at identifying candidate yeast prion proteins and potential disease-associated human PrLDs, these algorithms are less effective at predicting the aggregation propensity of these domains or the effects of mutations [41]. One limitation of these methods is that while they assess the frequency with which amino acids occur in prion domains, this frequency may not reflect the importance of each amino acid in prion formation. To address this knowledge gap, we previously used a quantitative mutagenesis method to score the prion propensity of each amino acid in the context of a Q/N-rich prion domain [42]. Interestingly, although the yeast prions tend to be strikingly Q/N-rich, both glutamine and asparagine were found to have neutral prion propensity scores [42, 43]. Instead, many of the non-aromatic hydrophobic amino acids (I, M, and V) and the aromatic amino acids (F, W, and Y) were observed to have strong prion-promoting effects, implicating these amino acids as key nucleators of prion aggregation [42, 44, 45]. We then used these prion-propensity scores to create PAPA, a prion prediction algorithm optimized for yeast prion domains [46, 47]. PAPA is reasonably effective at predicting the prion propensity of Q/N-rich domains, as well as the effects of mutations on prion activity [45, 47, 48].
However, although the composition of the human PrLDs resembles the composition of yeast prion domains, the human PrLDs tend to be less Q/N-rich and contain a higher percentage of serine and glycine (for review, see [41]). Therefore, it is likely that prediction methods developed for yeast prion domains may not be optimized for human PrLDs. To understand how amino acid context (i.e. the starting composition) of PrLDs affects amino acid prion propensities, we sought to determine the prion propensity of each amino acid in the context of two model glycine (G) rich human PrLDs from hnRNPA1 and A2. As in the context of Q/N-rich yeast prion domains, we found that aromatic amino acids were strongly prion-promoting in the context of these G-rich PrLDs. However, contrary to their effect in Q/N-rich yeast prion domains, the non-aromatic hydrophobic amino acids were not strongly prion-promoting; instead, they served as a signal for targeted degradation of the G-rich PrLDs. This suggests that aromatic amino acids may have the unique capacity to increase the aggregation propensity of prion or prion-like domains while avoiding efficient detection by protein degradation systems. Furthermore, Q/N residues strongly inhibited degradation of the G-rich PrLDs, suggesting that they may help prevent degradation of prion and prion-like domains. Indeed, many of the same sequences that led to degradation in the context of the G-rich PrLD had no effect on turnover of a Q/N-rich prion domain. These results broaden our understanding of the proteostatic regulation of aggregation-prone proteins, and shed light on the role of Q/N residues within prion domains.
The core PrLDs from the human RNA-binding proteins hnRNPA1 and hnRNPA2 were chosen as model substrates to examine the sequence requirements for aggregation within G-rich PrLDs. Both proteins contain a C-terminal G-rich PrLD. Mutations in these domains cause ALS and multisystem proteinopathy in humans, increase their aggregation propensity in vitro, and cause muscle degeneration when the proteins are expressed in Drosophila [37, 48, 49].
We previously used a yeast prion system to examine the effect of mutations on the aggregation propensity of the hnRNPA1 and A2 PrLDs [37]. The yeast prion protein Sup35 contains three functionally distinct domains: an N-terminal prion domain that is necessary and sufficient for formation of prion aggregates; a C-terminal functional domain, which is involved in translation termination; and a highly charged middle domain [15, 50, 51]. The first 40 amino acids of the prion domain, referred to as the nucleation domain (ND), are very Q/N-rich and are responsible for nucleating prion aggregates [40]. We therefore replaced the Sup35 ND with the core PrLD from hnRNPA1 and hnRNPA2 to test whether these PrLDs could support prion activity [37]. These fusion proteins allowed us to use well-established Sup35 prion detection assays to probe the relationship between amino acid sequence and aggregation activity for the hnRNPA1 and hnRNPA2 PrLDs.
Formation of [PSI+], the prion form of Sup35, can be assayed by monitoring nonsense suppression of the ade2-1 allele in the presence of tRNA suppressor SUP16 [52]. ade2-1 mutants are unable to grow on medium lacking adenine (SC-ade), and grow red on medium containing limited adenine (YPD) due to accumulation of a pigment derived from the substrate of the Ade2 enzyme; [PSI+] formation results in a low level of read-through of the ade2-1 premature stop codon, allowing for growth on SC-ade, and formation of white colonies on YPD. The fusion proteins showed a number of hallmarks of mutation-dependent prion activity [37, 48], including; 1) spontaneous formation of ADE+ colonies, and an increase in ADE+ colony formation upon PrLD overexpression; 2) curability of the ADE+ phenotype by 4mM GuHCl, a treatment that cures [PSI+] [53]; 3) transmission of the phenotype by cytoduction; and 4) the formation of microscopically-visible foci in ADE+ cells and the absence of foci in ade- cells. Furthermore, the in vitro amyloid propensity, the formation of visible foci, and the frequency of appearance of the ADE+ phenotype could be influenced in a predictable manner with rationally-designed mutations derived from an established prion propensity scale.
These studies provide strong evidence that the fusion proteins form prions. However, some Sup35 mutants with modified prion domains can show similar nonsense suppression that is not due to Sup35 prion formation [54]. Therefore, we took additional steps to confirm the prion activity of the fusion proteins. Prion maintenance requires continuous expression of the prion protein. To provide additional evidence that the hnRNP-Sup35 fusions form canonical aggregation-mediated prions in yeast, we induced prion formation by overexpressing the A2 D290V PrLD in cells expressing hnRNPA2-Sup35(D290V) as the sole copy of Sup35 [37]. Two independent [PRION+] colonies were transformed with a Sup35 plasmid lacking the prion domain (Sup35MC) and passaged in the absence of selection for the A2-Sup35 plasmid. In both cases, the prions formed by the A2-Sup35 fusions were cured upon loss of the A2-Sup35 expressing vector, and only rarely spontaneously reappeared upon re-introduction of the vector, indicating that the prion phenotype required expression of the A2 PrLD to be maintained (Fig 1A). Since the hnRNP-Sup35 fusions contain a portion of the native Sup35 prion domain, we also examined whether the prions formed by the A2 PrLD were transferable to wild-type Sup35, which could suggest that the remainder of the Sup35 prion domain (rather than the A2 PrLD) was predominantly responsible for prion activity. Co-expression of wild-type Sup35 with A2 D290V suppressed the prion phenotype. After passaging these cells in the absence of selection for the A2-Sup35 plasmid, 23 out of 24 isolates that had not maintained the A2-Sup35 plasmid showed a [prion-] phenotype (Fig 1B), while the final isolate exhibited an atypical yellow phenotype with small colonies that resembled neither a [PRION+] or a [prion-] phenotype (potentially indicating contamination). This suggests that prions formed by the A2-Sup35 fusions were not sufficient to structurally convert Sup35 to a prion state. Collectively, these results indicate that the hnRNP-Sup35 fusions form canonical, PrLD-dependent prions.
We previously developed a method to quantitatively score the effects of mutations on Sup35 prion activity [42, 44]. We replaced 8-amino acid segments of the prion domain with a random sequence, generating libraries of mutants. Each mutant was expressed as the sole copy of Sup35 in the cell. Randomly mutagenized libraries were plated onto medium lacking adenine to select for mutants that maintained the ability to form [PSI+]. This method was applied to various regions of wild-type and scrambled Sup35, including the Sup35 nucleation domain [42, 44]. Therefore, to examine how the sequence requirements for aggregation differ between Q/N-rich and G-rich PrLDs, we repeated this method, mutating the hnRNPA1-Sup35 and hnRNPA2-Sup35 fusions (herein referred to as A1-Sup35 and A2-Sup35 respectively; Fig 2A). As targets for mutagenesis, we selected segments with a mixture of predicted aggregation-promoting, aggregation-inhibiting, and neutral amino acids, near the site corresponding to a region previously mutagenized in Sup35 (Fig 2B).
Spontaneous [PSI+] formation is typically a stochastic and very rare event, occurring at a rate of less than 10−6 per generation [55]. By contrast, mutations that reduce Sup35 activity without causing prion aggregation will result in a constitutive ADE+ phenotype. Thus, to detect rare prion formation events from among a library of mutants, it is necessary to first eliminate mutants that have a constitutive ADE+ phenotype (Fig 2A). In previous screens with wild-type or scrambled Sup35, such constitutive ADE+ mutants were relatively rare, comprising ~5% of screened isolates [42, 44]. Unexpectedly, for the mutagenized A2-Sup35 and A1-Sup35 fusions, approximately 30–40% of the isolates were able to grow in the absence of adenine. These ADE+ isolates were not cured by treatment with 4mM GuHCl, suggesting that the growth on SC-ade resulted from non-prion-based inactivation of the hnRNP-Sup35 fusion proteins. As observed for [PRION+] isolates, replacement of the A2-SUP35 plasmid with a plasmid expressing Sup35MC results in loss of the ADE+ phenotype. However, in contrast to [PRION+] strains, when plasmids expressing A1- or A2-Sup35 mutants were isolated from representative strains with a constitutive ADE+ phenotype and shuffled back into the parent strain, the ADE+ phenotype spontaneously re-appeared (S1 Fig). This indicates that the phenotype results from loss of activity of the A2-Sup35 fusion protein, not from mutations in other cellular proteins or from classical PrLD-dependent prion propagation. Finally, while prion formation was associated with increased levels of insoluble A2-Sup35 protein, the constitutive ADE+ mutants did not contain substantial amounts of insoluble A2-Sup35 protein (S2 Fig).
Therefore, we sought to determine the basis of Sup35 inactivation among these isolates. We sequenced the mutagenized region of the A1/A2-SUP35 gene from randomly selected ade- and ADE+ isolates to determine whether specific sequence features were correlated with the ADE+ phenotype. For each amino acid, an odds ratio was calculated (Eq 1), representing the degree of over- or under-representation of the amino acid among ADE+ isolates (Table 1). For both libraries, each of the non-aromatic hydrophobic amino acids (I, L, M, and V) were over-represented among ADE+ isolates, while glutamine, asparagine, and each of the charged amino acids (D, E, K, and R) were under-represented (Table 1; Fig 3). Individually, not all of these biases reached the standard threshold of statistical significance (p < 0.05; Table 1). Grouping amino acids of similar physical properties can increase statistical significance by effectively increasing sample sizes. When considered as a group, the biases for hydrophobic amino acids, against charged amino acids, and against Q/N were each statistically significant in both libraries (P<0.01 in all cases; Table 1).
One possible explanation for the ADE+ phenotype is that the hnRNP-Sup35 fusions could be poorly expressed or rapidly degraded, causing a decrease in steady state levels of the fusion proteins. To test this possibility, four representative A2-Sup35 isolates that exemplified the amino acid biases among the ADE+ library were selected for comparison with randomly selected isolates from the ade- library. The ADE+ and ade- phenotypes originally observed for these isolates were confirmed by spotting onto SC-ade, YPD, and YPAD (Fig 4A). Previous studies suggest that an ADE+ phenotype is observed when steady-state Sup35 levels drop below about 40% of wild-type [56]. In synthetic complete medium, all four ADE+ isolates had steady-state A2-Sup35 levels that were less than 40% of wild-type, while three of four isolates from the ade- library had steady state A2-Sup35 levels above this threshold (S3A Fig). When cells were shifted to medium lacking adenine, A2-Sup35 levels dropped for all eight strains, but showed lowest levels for the four ADE+ isolates. Furthermore, as a group, steady state protein levels for ADE+ isolates were significantly lower (p < 0.001) than the grouped protein levels for ade- isolates in both synthetic complete and adenine-deficient synthetic complete media.
Exposed hydrophobic patches are known in some cases to trigger protein degradation [57, 58]. Therefore, we hypothesized that the lower average expression levels seen among the ADE+ isolates might be due to increased degradation. Cycloheximide (CHX) globally inhibits translation by preventing translocation of the ribosome along mRNA, providing a convenient tool to assay protein turnover [59]. After treatment with CHX, the fusion proteins within ADE+ isolates were rapidly degraded (Fig 4A). Three of the four ADE+ isolates contained little or no detectable A2-Sup35 by 2.5 hours after addition of CHX, while the fourth showed a substantial decrease in A2-Sup35 levels over the 5 hour timecourse (Fig 4A). By contrast, A2-Sup35 levels remained relatively stable or decreased only slightly over a period of 5 hours after addition of CHX for all of the ade- isolates, as well as for the wild-type A2-Sup35 fusion (Fig 4A). These results suggest that hydrophobic amino acids trigger degradation of the A2-Sup35 fusions.
Interestingly, random mutagenesis of the Sup35 prion domain yielded very few isolates with the degradation phenotype in the initial screen [44], suggesting that the Sup35 prion domain can buffer the effects of degradation-promoting peptides. Indeed, when the degradation-promoting 8-amino acid sequences from the A2-Sup35 library were substituted into the corresponding region of the Sup35 prion domain, each of the proteins resulted in phenotypically ade- cells (Fig 4B), and maintained steady-state Sup35 levels well above the 40% of wild-type (S3B Fig). Furthermore, none of the peptides accelerated the degradation rate of Sup35 over 5 hours (Fig 4B). Therefore, while the A2 PrLD is susceptible to the degradation-promoting effects of hydrophobic amino acids, the Sup35 prion domain can mask these effects and resist degradation.
The ubiquitin-proteasome system is one of the main protein recycling pathways in eukaryotic cells. MG-132, a commonly used proteasome inhibitor, is effective in yeast lacking the pleiotropic drug resistance 5 gene (pdr5Δ). To assess whether degradation of the A2-Sup35 proteins occurs via the proteasome, PDR5 was deleted from the genome, and the turnover of the A2-Sup35 proteins was assessed in the presence or absence of MG-132. Pre-treatment with MG-132 for 1 hour prior to addition of CHX resulted in nearly complete stabilization of the degradation-prone A2-Sup35 fusions over the 5 hour timecourse (Fig 5). This result suggests that the ADE+ phenotype is due to enhanced turnover of the A2-Sup35 fusion proteins via the ubiquitin-proteasome system.
Since degradation-promoting sequences failed to cause degradation of Sup35 (Fig 4B), we reasoned that our previous dataset from random mutagenesis of Sup35 [44] would contain some peptide sequences that did not cause degradation in the context of the Sup35 prion domain, but would promote degradation of the A2-Sup35 fusion protein. To identify potential degradation-promoting sequences, each peptide from the library was scored by summing the log-odds ratios from Table 1 for the eight amino acids in the mutagenized region. Three sequences predicted to promote degradation (i.e., sequences enriched in non-aromatic hydrophobic residues, with few charged or Q/N residues) were selected from the dataset.
When substituted into A2-Sup35, all three predicted degradation-promoting peptides led to enhanced turnover of A2-Sup35 and characteristic degradation phenotypes, albeit to varying degrees (Fig 6A). All three strains appeared light pink on YPD, and growth on SC-ade correlated qualitatively with the degree of degradation conferred by each peptide. Additionally, two sequences predicted to have no effect on A2-Sup35 turnover (i.e., sequences enriched in charged and polar residues) were chosen from the same dataset as controls. When substituted into A2-Sup35, neither peptide enhanced degradation, and both strains displayed the associated ade- phenotypes (Fig 6A). By contrast, four of the five peptides substituted into the Sup35 prion domain had little effect on turnover and resulted in the characteristic ade- phenotype (Fig 6B), while the fifth showed modest degradation and only a weak ADE+ phenotype. These results demonstrate that the compositional biases originally observed in the ADE+ libraries are sufficient to predictively categorize sequences as degradation-promoting or degradation-inhibiting.
The sequences obtained through random mutagenesis are heterogeneous with respect to composition and sequence. To more rigorously define the minimum number of non-aromatic hydrophobic residues required to accelerate the rate of degradation or prion formation, hydrophobic content was progressively increased in WT A2-Sup35 and WT Sup35. Valine, leucine, and methionine (the hydrophobic residues most over-represented in the A2-Sup35 ADE+ library) were inserted in an alternating fashion adjacent to the region targeted for random mutagenesis (Figs 2B and 7).
As few as two hydrophobic residues were sufficient to slightly increase turnover of A2-Sup35, as indicated by western blot and the characteristic ADE+ phenotype (Fig 7A). Three hydrophobic residues further accelerate A2-Sup35 degradation, and four to seven hydrophobic residues caused almost complete loss of A2-Sup35 by 2.5 hours after the addition of CHX. Two or fewer hydrophobic residues inserted into Sup35 resulted in uniform ade- phenotypes, whereas three or more hydrophobic residues resulted in the appearance of white sectors, which are classical indications of prion formation (Fig 7B). Strikingly, the degree of sectoring increased in a dose-dependent fashion as hydrophobic content increased. Elimination of the [PIN+] prion did not affect the degradation of A2-Sup35 or stability of Sup35 upon insertion of hydrophobic residues (S4 Fig).
To more accurately quantify the frequency of ADE+ colony formation by each mutant, serial dilution of each mutant was plated on SC-ade, starting from a higher density than originally assayed. Fewer than two hydrophobic residues in A2-Sup35 resulted in minor growth only at high cell density, whereas two or more hydrophobic residues resulted in robust growth even at very low cell density (Fig 7C). Treatment with GuHCl did not alter the color phenotype on YPD (Fig 7D), suggesting that the ADE+ growth was not due to prion formation.
By contrast, three or more hydrophobic residues in Sup35 resulted in a progressive increase in the frequency of ADE+ colonies, consistent with the progressive increase in sectoring observed on YPD for these mutants (Fig 7E). Treating the cells with GuHCl reverted the ADE+ phenotype to an ade- phenotype (Fig 7F), confirming that growth on SC-ade was due to the formation of bona fide prions.
These results were not unique to these specific positions within Sup35 and A2-Sup35. We made additional hydrophobic insertions one-quarter and three-quarters of the way through the Sup35 ND and the hnRNPA2 PrLD (positions 10 and 30 for Sup35; positions 11 and 33 for A2; Fig 2B). As with the original hydrophobic insertions (Fig 7A), insertions at both additional positions in the A2 PrLD resulted in increased degradation, although the effects of insertion were weaker at position 33 (Fig 7G). Likewise, Sup35 was far more resistant to the degradation-promoting effects of hydrophobic amino acids at both positions, although modest degradation was observed when six hydrophobic amino acids were inserted at position 10 (Fig 7H).
It is possible that physical interactions between the Sup35ND and the remainder of the Sup35 sequence or with native Sup35 binding partners are responsible for the apparent stability of the Sup35 ND. However, insertion of hydrophobic residues in the A2 PrLD alone fused to GFP resulted in a progressive increase in degradation rate (Fig 8A, top), whereas insertion of hydrophobic residues in the Sup35 ND had no effect on degradation (Fig 8B, top). Nearly identical trends were observed for FLAG-tagged version of the A2-Sup35 and Sup35 NM domains (Fig 8, bottom).
Collectively, these results demonstrate that the Sup35 ND can mask the degradation-promoting effects of hydrophobic residues, and that this effect is not dependent on the remainder of the protein.
The features promoting degradation of the A2 PrLD are consistent with previous studies indicating that the degradation machinery recognizes exposed hydrophobic segments, and that there is a strong correlation between the sequence features that promote aggregation and degradation [57, 58]. We were interested in whether this correlation is absolute, or whether there are sequence features that can promote aggregation of the G-rich PrLDs without promoting degradation. Our A1- and A2-Sup35 fusions provide a useful system for comparing the sequence requirements for degradation versus aggregation. To determine whether specific sequence features could promote prion aggregation without triggering degradation, isolates with an initial ade- phenotype were plated onto medium lacking adenine to screen for the ability to spontaneously form prions (Fig 2A). [PRION+] isolates were confirmed by curing with GuHCl, and the mutagenized A1/A2-SUP35 gene in each was sequenced. Sequences from each library were pooled, and the prion propensity scores for each amino acid were determined, as described previously ([42, 44]; Eq 4).
Interestingly, while both non-aromatic and aromatic hydrophobic residues were strongly prion-promoting within the Q/N-rich Sup35 ND [42, 44], only aromatic amino acids were significantly over-represented among [PRION+] isolates for the A2-Sup35 and A1-Sup35 libraries; non-aromatic hydrophobic residues were approximately equally represented among [PRION+] and ade- isolates (Table 2). Furthermore, Q/N residues were significantly under-represented among A2-Sup35 [PRION+] isolates, although their effects were mixed among A1-Sup35 [PRION+] isolates. Together, these results suggest that a hitherto unappreciated property of aromatic amino acids is the unique ability to promote protein aggregation of prion and prion-like domains, while avoiding detection by the degradation machinery.
Indeed, while there is a statistically significant (P = 0.008 by Spearman rank analysis) correlation between the prion propensity (as scored by PAPA) of each amino acid and its propensity to promote degradation (Fig 9), there are five amino acids which have substantially lower degradation propensities than would be predicted by their prion propensities: the three aromatic amino acids, glutamine, and asparagine. Strikingly, these amino acids are all overrepresented among yeast prion proteins. While both aromatic and non-aromatic hydrophobic amino acids strongly promote prion formation [42], candidate prion domains with prion activity tend to contain more aromatic residues and fewer aliphatic residues than candidate prion domains with no detectable prion activity [44]. Likewise, although serine, glycine, threonine, glutamine, and asparagine each promote intrinsic disorder and have similar prion propensities [42], Q/N residues are far more common among yeast prion domains.
Collectively, these results suggest a possible explanation for the amino acid biases observed among yeast prion domains. Many components of protein quality control systems act specifically to antagonize protein aggregation. Therefore, proteins that form observable protein aggregates must possess mechanisms to avoid or outcompete antagonistic proteostasis machinery. Yeast prion domains tend to favor amino acids that promote aggregation while being poorly recognized by the degradation machinery.
These results may also provide an explanation for Sup35’s resistance to degradation. Q/N residues were among the lowest scoring amino acids in the degradation libraries. The human PrLDs and the Sup35 ND differ most notably in their Q/N content; the Sup35 ND contains a much higher percentage of Q/N-residues, while the A1 and A2 core PrLDs are more G-rich. This suggests the simple hypothesis that the high Q/N-content of the Sup35 ND may protect highly aggregation-prone features from recognition by components of the proteostasis machinery. To test this hypothesis, two of the degradation-prone members of the A2 library and their Sup35 counterparts were chosen as initial substrates for mutagenesis. To examine the relationship between Q/N content and degradation, we mutated some or all of the Q/N’s in the Sup35 nucleation domain to G’s (Fig 10A). Similarly, we mutated some or all of the G’s in the A2 PrLDs to Q/N.
The rate of degradation of Sup35 correlated with Q/N-content in a dose-dependent manner. Partial substitution of Q/N-residues for G’s significantly increased the turnover rate of each Sup35 derivative and resulted in the emergence of the ADE+ phenotype (Fig 10B; S5 Fig). Substitution of the remaining Q/N’s for G’s further enhanced the rate of Sup35 degradation. Partial or full substitution of G’s for Q/N’s in the A2 PrLD resulted in a modest, albeit statistically significant increase in stability for one of the two mutagenized PrLDs. However, no stabilizing effect was observed for the second mutagenized PrLD, suggesting that other sequence features of the A2 PrLD besides Q/N content must contribute to its sensitivity to degradation. Therefore, in addition to their role in prion formation, Q/N residues help the Sup35 prion domain resist degradation by intracellular anti-aggregation systems.
Protein misfolding is a selective challenge faced by all cellular life. Misfolded proteins can result in proteotoxicity, either through loss-of-function of the native protein or through a toxic gain-of-function of the misfolded species. To address these selective challenges, eukaryotic cells possess extensive proteostasis machinery, which constitutively act to procure and maintain pools of natively folded proteins. The proteostasis machinery broadly consists of three main systems: 1) the protein chaperone network, which aids in nascent protein folding as well as the re-folding of partially or fully denatured proteins, 2) the ubiquitin-proteasome system, and 3) the autophagy system, which together aid in the destruction of aged, terminally misfolded, or aggregated proteins (for review, see [60, 61]).
Despite the constant surveillance of protein quality control systems, numerous diseases result from misfolding and aggregation of proteins. Additionally, a variety of proteins form functional aggregates that are involved in the regulation of various cellular processes [62–64]. Therefore, understanding how the proteostasis machinery detects misfolded proteins, and how some aggregation-prone proteins evade this detection, may provide insight into both functional and pathogenic aggregation.
One way through which proteostasis network components achieve specificity for misfolded proteins is by recognizing patches of solvent-exposed hydrophobicity [65–73]. Hydrophobic patches are generally buried in the interior of folded proteins [74], so exposed hydrophobicity can act as a signal of protein misfolding. Additionally, there is a strong correlation between hydrophobicity and aggregation propensity [75], so recognizing exposed hydrophobicity would seem to be an effective mechanism to recognize aggregation-prone misfolded proteins.
One well-characterized example that uses this mechanism is the yeast E3 ubiquitin ligase San1, a nuclear protein involved in the ubiquitin-proteasome degradation system [76]. San1 is a largely disordered protein that is particularly adept at targeting toxic misfolded proteins for degradation [77], primarily by recognizing exposed hydrophobic residues in substrates [58]. Interestingly, San1 recognition of these substrates tends to correlate with their insolubility, demonstrating the effectiveness of targeting hydrophobicity to prevent protein aggregation [57]. It should be noted that degradation of the A1- and A2-Sup35 fusions was independent of San1 (S6 Fig), so additional work will be required to identify the cellular factors responsible for recognition and degradation of these PrLDs. Identifying these factors and studying the mechanisms by which they recognize aggregation-prone proteins may help explain mechanistically how aromatic amino acids can promote aggregation without triggering PQC degradation.
Although our data is generally consistent with the idea that exposed hydrophobic amino acids promote recognition by the proteostasis machinery, our results provide some additional unexpected insights. First, in contrast to what has been proposed for San1, we show that aggregation propensity and recognition by the proteostasis machinery can be uncoupled in a composition-dependent manner: aromatic amino acids within the G-rich hnRNP PrLDs increase aggregation propensity without substantially enhancing recognition by the proteostasis machinery. Second, the ability of the proteostasis machinery to recognize hydrophobic patches was highly context dependent: the Q/N-rich Sup35 ND had the inherent capacity to mask otherwise degradation-promoting amino acids. These results highlight an important point related to the proteostasis of prion and prion-like domains. While it is sometimes useful to broadly categorize certain amino acids as “aggregation-promoting” or “degradation-promoting”, the effects of these amino acids may vary from protein to protein depending on the larger sequence context within which they are found, and on the interactions between these domains and cellular proteostasis factors. Since both the short sequence features and the surrounding context play such important roles in aggregation and degradation, future examination of the extent to which these heuristics apply to other aggregation-prone proteins and in other organisms would be interesting.
While PolyQ regions reportedly resist degradation by the proteasome [78] (although this too remains quite controversial [79, 80]; for review, see also [81, 82]), Sup35 has a roughly average half-life in vivo [83] suggesting that it is not inherently unusually resistant to degradation. Some evidence indicates that certain fragments of the Q/N-rich Sup35 prion domain exhibit a high rate of turnover [84], and the Sup35 prion domain can be proteolytically cleaved [85], indicating that the degradation and proteolytic systems are not incapable of processing the Sup35 prion domain in vivo. However, our results illuminate a principle fundamentally distinct from inherent stability–namely, sequences capable of potently inducing degradation in the G-rich PrLDs are, in some way, protected from the proteostasis machinery by surrounding Q/N-rich regions. Therefore, Q/N residues may potentiate the aggregation of prion domains, in part, by protecting aggregation-prone features from the proteostasis machinery. Although increased degradation and increased aggregation are not necessarily alternatives [84, 86], our results suggest that the composition of the Sup35 prion domain allows it to resist degradation, while maintaining the ability to form prions.
While the exact mechanism by which Sup35 resists degradation is unclear, high Q/N-content appears to play an important role. The Sup35 prion domain and the A1/A2 PrLDs are each predicted to be intrinsically disordered. However, the Sup35 prion domain is thought to form a collapsed but disordered structure [87], which may hide hydrophobic patches from the proteostasis machinery. High Q/N content may help mask hydrophobic patches by promoting a collapsed but disordered structure, or by shielding hydrophobic amino acids within these structures. Alternatively, rather than preventing the initial recognition of hydrophobic patches by the proteostasis machinery, Q/N residues may inhibit a downstream step in the subsequent events leading to degradation. Interestingly, the Q/N content of Sup35 is relatively well-conserved across independent Saccharomyces cerevisiae strains and between different yeast species [88, 89]. Although high Q/N content within the N-domain may be maintained by selection for multiple reasons, it is possible that the stabilizing effects of Q/N at least contribute to the observed compositional conservation.
Other features of the Sup35 prion domain besides Q/N content also seem well-suited to avoid detection by the degradation machinery, potentially explaining why increasing the Q/N content of the A2 PrLD was not sufficient to fully stabilize the PrLD. We previously showed that six amino acids are highly prion-promoting: F, Y, W, I, V, and M [42]. The Sup35 prion domain contains 23 of these highly prion-promoting amino acids, yet all except the initiating methionine are aromatic. Additionally, the prion-promoting amino acids are well-dispersed. There is only one position where two occur adjacent to each other, and almost all have adjacent Q/N residues. Thus, the Sup35 prion domain possesses many features that promote aggregation, yet avoids multiple features that can contribute to degradation. Furthermore, these biases are not unique to the Sup35 prion domain; most other yeast prion domains are also Q/N-rich, and tend to favor aromatic amino acids over non-aromatic prion-promoting amino acids [44].
Numerous labs have made extensive progress in defining how the amino acid sequence of a protein affects its intrinsic aggregation propensity. However, our results highlight that intrinsic aggregation propensity is only a small piece of the puzzle. A more complete understanding of functional and pathogenic protein aggregation requires a clearer view of how amino acid sequence affects interactions with other cellular proteins. Our results provide one unexpected piece to this puzzle, demonstrating that specific sequence features can promote protein aggregation, while simultaneously hiding from the proteostasis machinery.
Standard yeast media and methods were used as previously described [90], except that YPD plates contained 0.5% yeast extract rather than the standard 1%. YPAD for all experiments contained the standard 1% yeast extract, as well as 0.02% adenine hemisulfate. Prion curing assays were performed for individual ADE+ isolates by streaking onto YPD with and without 4mM GuHCl, then re-streaking to YPD to test for loss of the ADE+ phenotype. In all experiments, yeast were grown at 30°C. The yeast strains used in this study were YER826/pER589 (α kar1-1 SUQ5 ade2-1 his3 leu2 trp1 ura3 sup35::KanMx) and YER1161 (α kar1-1 SUQ5 ade2-1 his3 leu2 trp1 ura3 sup35::KanMx pdr5::HIS3). pER589 expresses a truncated version of Sup35 lacking the prion domain (Sup35MC) as the sole copy of Sup35 in the cell. This plasmid was subsequently replaced by plasmid shuffling in order to assay activity of the full-length, randomly mutagenized hnRNP-Sup35 fusions.
The A1-Sup35 and A2-Sup35 fusion libraries were generated in a manner similar to MacLea et al. [44]. Briefly, the N-terminal end and C-terminal end of each gene were amplified from a plasmid containing either the A1-Sup35 fusion or the A2-Sup35 fusion (pER595 for hnRNPA1 and pER697 for hnRNPA2; [37]). Oligonucleotides (from Integrated DNA Technologies) were used to re-amplify the respective products and incorporate a 24-nucleotide degenerate region in which each of the four nucleotides has a 25% probability of occurring at the first two positions of each codon, while C, G, and T each have a 33% probability of occurring at the final position of each codon. The N-terminal and mutagenized C-terminal products, which contain complementary segments, were mixed and re-amplified. The final PCR products were co-transformed with BamHI/HindIII-cut pJ526 into YER826 and plated on synthetic complete media lacking leucine (SC-Leu) to select for cells containing a recombined plasmid. Individual colonies were then picked and stamped onto media containing 5-Fluoroorotic acid (5-FOA) to select for loss of pER589.
After 5-FOA treatment, cells were transferred to YPAD, YPD, and SC-ade. After three days at 30°C, isolates for which more than 5 colonies appeared on SC-ade were identified and placed in a category (ADE+ library) separate from those with fewer than 5 colonies (ade-). Randomly selected representative isolates from both groups were sequenced to generate each library. The odds ratio for the ADE+ phenotype (ORA) for each amino acid was determined as follows:
ORA=[fD1-fD]/[fN1-fN]
(1)
where fD represents the per residue frequency of the amino acid among the isolates that were able to grow on SC-ade, and fN represents the per residue frequency of the amino acid among the naïve isolates (i.e., those that were unable to grow on SC-ade). Final degradation propensity scores for each amino acid (DPaa) were determined as follows:
DPaa=ln(ORA)
(2)
In addition, prion isolates were identified as previously described [42, 44] (Fig 1) and sequenced to generate the prion library. Briefly, the isolates that were initially unable to grow on SC-ade were pooled from the solid YPAD media and re-plated on SC-ade at approximately 106 and 105 cells per plate. After 3–5 days at 30°C, individual colonies were streaked onto YPD and YPD plus 4mM GuHCl to assay for prion loss. Odds ratios for prion activity (ORP) for each amino acid were determined as follows:
ORP=[fP1-fP]/[fN1-fN]
(3)
where fP represents the per residue frequency of the amino acid among the prion-forming isolates. Final prion propensity scores (PPaa) were determined as follows:
PPaa=ln(ORP)
(4)
Cells were diluted to an optical density of 0.75 in liquid YPAD media and incubated with shaking at 30°C for 1hr before treatment with CHX, or DMSO for untreated cells. Where applicable, MG-132 was added to a final concentration of 10μg/mL 1 hr prior to addition of CHX. After the treatment period, the optical densities of all cultures were measured. 10mL of the least-dense culture for each strain were harvested. Based on the optical densities, the approximate number of cells harvested for each of the remaining cultures was normalized to the least-dense culture within each unique strain. Cells were pelleted by centrifugation at 3,000rpm for 5 minutes at 4°C. Cell pellets were lysed as previously described [57]. 30μL of prepared lysate were loaded onto a 12% polyacrylamide gel, transferred to a PVDF membrane, and probed with an appropriate antibody. Primary antibodies (all monoclonal) used in this study were: an anti-Sup35C (BE4 [91], kindly made available by Susan Liebman), an anti-GFP antibody (Santa Cruz Biotechnology), and an anti-FLAG antibody (Sigma). Blots were quantified using Image Studio Version 5.2. Background-subtracted intensities for all quantified blots can be found in S1 Table.
As with the degradation assays, yeast cultures were diluted to an optical density of 0.75 in liquid YPAD media and incubated with shaking at 30°C for 1hr before normalizing to the least-dense culture and harvesting. Cell pellets were re-suspended in 200μL of chilled non-denaturing lysis buffer (100mM TrisHCl, pH7.5, 200mM NaCl, 1mM EDTA, 5% Glycerol, 0.1% Triton-X 100, and Bond-Breaker TCEP solution (Thermo Fisher Scientific) and ProBlock Gold yeast protease inhibitor cocktail (Gold Biotechnology) to manufacturer recommendations; adapted from [92]), transferred to a round-bottom 2mL tube, and vortexed with a single large glass bead (~3mm diameter) on maximum speed for 10 minutes. Lysates were centrifuged gently (700 x g for 5 minutes at 4°C) to pellet unlysed cells and large cellular debris. 50μL of total lysate sample was mixed with 50μL of denaturing buffer (1% SDS, 8M urea, 10mM MOPS pH6.8, 10mM EDTA pH8.0, 0.01% bromophenol blue, and ProBlock Gold yeast protease inhibitor cocktail (Gold Biotechnology) to manufacturer recommendations). 100μL of remaining lysate was centrifuged at 16.3k rpm for 15 minutes at 4°C to pellet protein aggregates. The supernatant was removed and mixed 1:1 with denaturing buffer (soluble sample). The remaining pellet was resuspended in an equal volume of a 1:1 mixture of non-denaturing:denaturing buffer. Samples were boiled for 5 minutes, then centrifuged at 12,900 x g for 5 minutes before loading.
Original strains were transformed with a covering plasmid expressing a version of Sup35 lacking the prion domain (Sup35MC) and a URA3 selectable marker. Transformants were passaged on SC-Ura until loss of the original plasmid expressing the A2-Sup35 fusion. After loss of the plasmid, strains were re-transformed with the original A2-Sup35 fusion plasmid and the URA3 covering plasmid counter-selected on 5-fluoroorotic acid (FOA). Color phenotypes for each strain were compared by streaking onto YPD.
|
10.1371/journal.pgen.1000615 | Phosphofructo-1-Kinase Deficiency Leads to a Severe Cardiac and Hematological Disorder in Addition to Skeletal Muscle Glycogenosis | Mutations in the gene for muscle phosphofructo-1-kinase (PFKM), a key regulatory enzyme of glycolysis, cause Type VII glycogen storage disease (GSDVII). Clinical manifestations of the disease span from the severe infantile form, leading to death during childhood, to the classical form, which presents mainly with exercise intolerance. PFKM deficiency is considered as a skeletal muscle glycogenosis, but the relative contribution of altered glucose metabolism in other tissues to the pathogenesis of the disease is not fully understood. To elucidate this issue, we have generated mice deficient for PFKM (Pfkm−/−). Here, we show that Pfkm−/− mice had high lethality around weaning and reduced lifespan, because of the metabolic alterations. In skeletal muscle, including respiratory muscles, the lack of PFK activity blocked glycolysis and resulted in considerable glycogen storage and low ATP content. Although erythrocytes of Pfkm−/− mice preserved 50% of PFK activity, they showed strong reduction of 2,3-biphosphoglycerate concentrations and hemolysis, which was associated with compensatory reticulocytosis and splenomegaly. As a consequence of these haematological alterations, and of reduced PFK activity in the heart, Pfkm−/− mice developed cardiac hypertrophy with age. Taken together, these alterations resulted in muscle hypoxia and hypervascularization, impaired oxidative metabolism, fiber necrosis, and exercise intolerance. These results indicate that, in GSDVII, marked alterations in muscle bioenergetics and erythrocyte metabolism interact to produce a complex systemic disorder. Therefore, GSDVII is not simply a muscle glycogenosis, and Pfkm−/− mice constitute a unique model of GSDVII which may be useful for the design and assessment of new therapies.
| Type VII glycogen storage disease (GSDVII), or Tarui disease, is a rare genetic disorder characterized by glycogen accumulation in skeletal muscle. The molecular cause is loss of activity of the muscle isoform of phosphofructokinase (PFK), which phosphorylates fructose-6-phosphate to fructose-1,6-bisphosphate, commiting glucose to glycolysis. Entry of fructose-6-phosphate into glycolysis is thus blocked, increasing glycogen synthesis and accumulation. Clinical manifestations of the disease are heterogeneous, ranging from exercise intolerance to early childhood death. To further understand the human pathology, we generated mice lacking muscle PFK. As in human patients, these mice showed severe exercise intolerance, hemolysis, and most died young. Lack of glycolysis in skeletal muscle also causes alterations in bioenergetics and compensatory changes in key metabolic genes. Additionally, although erythrocytes retained 50% of normal PFK activity, their overall functionality was impaired, aggravating the muscle dysfunction. Moreover, marked metabolic alterations in the heart lead to chronic hypertrophy, suggesting that cardiac pathology in GSDVII may be underestimated or misdiagnosed. This study indicates that this disease is more complex than a muscle glycogenosis and that symptoms other than those classically described should be taken into consideration. Finally, this animal model will enable us to develop new therapeutic approaches and better diagnostic tools.
| Phosphofructo-1-kinase (PFK) is a tetrameric enzyme that phosphorylates fructose-6-phosphate to fructose-1,6-bisphosphate, committing glucose to glycolysis. Three PFK isoenzymes, encoded by separate genes, have been identified in mammals: muscle-type (PFKM), liver-type (PFKL), and platelet-type (PFKP), all of which are expressed in a tissue specific manner [1]. Thus, skeletal muscle expresses only PFKM homotetramers, liver mainly PFKL homotetramers, although it can also express M- and P-type subunits, while erythrocytes contain PFKM and PFKL heterotetramers [2],[3]. Several mutations in PFKM cause type VII glycogen storage disease (GSDVII), which is a rare disease described by Tarui (Tarui's disease) [4]. GSDVII is inherited as an autosomal recessive trait and patients show loss of PFK activity in skeletal muscle and also partial deficiency in erythrocytes. Although GSDVII is characterized by accumulation of glycogen in skeletal muscle and hemolysis, there are several subtypes with different clinical features. No genotype-phenotype correlation explaining the phenotypic heterogeneity of the disease has been described [5]. It can be detected as a severe form with onset in infancy with hypotonia, limb weakness, progressive myopathy and respiratory failure leading to death early in the childhood [6],[7]. Neonatal mortality may be responsible for the low number of cases diagnosed. Adult patients with the classical form of the disease develop myopathy with muscle cramps and myoglobinuria when exercised as well as compensated haemolytic anemia.
GSDVII is considered as a muscle glycogenosis. Although, alterations in oxidative metabolism and bioenergetics in skeletal muscle have also been described in human patients, few data on metabolic and fiber structural changes are available. In addition, the contribution of altered glucose metabolism in other tissues to the pathogenesis of the disease is not fully understood and may also lead to misdiagnosis [8]. No therapies are available for GSDVII patients and development of effective treatments requires both understanding the molecular mechanisms that lead to the disease and the development of animal models in which to test new treatments. Inherited PFKM deficiency has only been described in dogs [9],[10]. However, PFKM deficient dogs exhibit mild muscle disease not closely reproducing the human muscle pathology [11]. In the present study, to determine the molecular mechanisms underlying this disease, we have generated mice lacking the muscle isoform of PFK. We found that PFKM deficiency leads to marked alterations in muscle bioenergetics and erythrocyte metabolism that interact to produce the complex pathology characteristic of GSDVII. The availability of the Pfkm−/− mouse model allows the study of GSDVII as a systemic disorder, not simply as muscle glycogenosis.
To generate Pfkm deficient mice, standard gene-targeting methods in mouse embryonic stem cells were used. Homologous recombination of the targeting construct resulted in the deletion of the 5′ promoter region and exon 3, which contains the translation start codon (Figure 1A). The presence of heterozygous and homozygous (Pfkm+/− and Pfkm−/−) mice was confirmed by Southern blot (data not shown) and by PCR (Figure 1B). Pfkm+/− mice were viable and fertile while Pfkm-null mice presented high lethality around weaning (about 60%) and those surviving died early during adulthood, at around 3 to 6 month of age, although few animals survived for more than one year.
Pfkm+/− mice showed 50% lower muscle Pfkm expression and activity (Figure 1C and 1D). However, this lower enzyme activity in Pfkm+/− mice did not alter any metabolic parameter, such as glucose-6-phosphate and glycogen levels (data not shown), indicating that half of normal PFK activity is sufficient to prevent metabolic alterations, as observed in heterozygous humans [12]. No Pfkm mRNA transcript was observed in skeletal muscle of Pfkm−/− mice (Figure 1C), in agreement with the lack of enzyme activity (Figure 1D). This deficiency led to increased glucose-6-phosphate (Figure 1E), intracellular glucose (Figure 1F) and glycogen (Figure 1G) content in skeletal muscle. Considerable glycogen storage was also evidenced by histochemical analysis of Pfkm−/− skeletal muscle (Figure 2A). Furthermore, electron microscopy revealed very high subsarcolemmal and intermyofibrillar accumulation of glycogen, which altered fiber morphology (Figure 2B). In addition, Pfkm−/− mice showed lower serum lactate levels (Figure 1H), suggesting lower flux through glycolysis in skeletal muscle. Nevertheless, these mice were normoglycemic (Pfkm+/+, 115±12 vs. Pfkm−/−, 113±16 mg/dl; (n = 12)). Consistent with this, PFK activity and glucose metabolism were unchanged in the liver (data not shown).
Similar to patients with the classical form of GSDVII, two-month-old Pfkm-null mice were intolerant to exercise. These mice were unable to run for more than 1.5 min. in a treadmill before developing severe muscle cramps, mainly in the rear limbs (Figure 2C). When exercised, Pfkm−/− mice accumulated higher levels of glucose-6-phosphate (Figure 1E), consistent with increased muscle glucose uptake (Figure 1F) and mobilization of muscle glycogen (Figure 1G).
Since the muscles of these mice fail to perform glycolysis, lactate did not rise after exercise (Figure 1H). Furthermore, in Pfkm−/− skeletal muscles, ATP and ADP levels were lower even in the resting state and fell with exercise (Figure 2D). These lower levels of ATP agreed with the presence of muscle cramps after exercise, spontaneous cramps during manipulation, and immediate rigor mortis after death (not shown). Thus, skeletal muscle of Pfkm−/− mice was unable to meet the energy demand required to maintain normal contractile activity.
Despite low ATP levels in Pfkm−/− mice, the expression of key genes in oxidative metabolism and mitochondrial biogenesis was higher than in wild-type mice, such as peroxisome proliferator-activated receptor γ coactivator-1α (PGC-1α), peroxisome proliferator-activated receptor δ (PPARδ) muscle carnitine palmitoyltransferase 1 (M-CPT-1), citrate synthase (CS) and uncoupling protein 2 (UCP2) (Figure 3A). Moreover, succinate dehydrogenase and NADH-tetrazolium reductase activities, markers of oxidative capacity, were also higher (Figure 3B). Up-regulation of the expression of type I and IIa myosin heavy chain (MyHC-I and IIa) oxidative-type fiber proteins, without changes in the glycolytic MyHC-IIb, was also observed (Figure 3C). Consistent with these findings, Pfkm−/− mice showed proliferation of enlarged mitochondria surrounded by glycogen depots (Figure 2B). Increased expression of genes involved in glucose uptake and phosphorylation, glucose transporter 4 (GLUT4) and hexokinase-II (HK), was found in skeletal muscle of Pfkm−/− mice (Figure 3D), which also agreed with increased muscle glucose and glucose-6-phosphate content (Figure 1E and 1F). In addition, the expression of the pentose phosphate pathway transaldolase (TALDO1) and transketolase (TK) genes was higher in skeletal muscle of Pfkm−/− than in wild-type mice (Figure 3E). Therefore, despite an increased compensatory response, oxidative metabolism was unable to overcome the glycolysis blockade in Pfkm−/− mice.
Respiratory skeletal muscles were also severely altered in Pfkm−/− mice. The lack of PFK activity in diaphragm, led to increased glucose-6-phosphate and glycogen content (Figure 4A–4C). High accumulation of glycogen was also observed in diaphragm (Figure 4D) and intercostal muscle (Figure 4E) sections by PAS staining. These metabolic alterations may have contributed to alter the respiratory capacity of Pfkm−/− mice.
Cardiac muscle, which expresses the PFKM, PFKL and PFKP [3], showed lower PFK activity in Pfkm−/− mice (about 20% of wild-type) and higher glucose-6-phosphate and glycogen levels (Figure 5A–5C). In addition, increased glycogen storage was also evident in electron microscopy sections of cardiac muscle (Figure 5D). Two-month-old Pfkm−/− mice showed increased (about 55%) heart weight (Pfkm+/+, 4.4±0.1 mg/g b.w. vs. Pfkm−/−, 6.9±0.3 mg/g b.w.; (n = 5) p<0.01) and developed cardiac hypertrophy and evident cardiomegaly with age (Figure 5E and 5F). Moreover, left ventricle enlargement without interstitial fibrosis was observed after Masson trichromic staining of longitudinal sections of Pfkm−/− hearts (Figure 5G).
The decrease of erythrocyte 2,3-BPG levels increases hemoglobin affinity for oxygen and thus impairs oxygen extraction from hemoglobin [15]. Thus, the inability of oxidative metabolism to compensate for glycolysis blockade in Pfkm−/− skeletal muscle may also be due to decreased availability of oxygen to generate sufficient energy. Furthermore, consistent with decreased oxygen availability and marked hemolysis, skeletal muscle of Pfkm−/− mice showed hypoxia, evidenced by higher expression (6-fold) of the hypoxia induced factor 1α (HIF-1 α) (Figure 7A). Moreover, expression of genes activated by HIF-1α, such as pyruvate kinase M (PK-M), lactate dehydrogenase (LDH), and glucose transporter-1 (GLUT1), were up-regulated in this tissue (Figure 7A). This increase in GLUT1 was also consistent with the observed higher intracellular glucose (Figure 1F). The increase in HIF-1α was also parallel to increased vascular endothelial growth factor (VEGF) expression (Figure 7B). In addition, it has been described in skeletal muscle that PGC1α is induced by a lack of oxygen and that PGC1α powerfully regulates VEGF expression [16], which may have also occurred in Pfkm−/− mice. The increase in VEGF led to hypervascularization, as evidenced by greater immunostaining of the platelet endothelial cell adhesion molecule (PECAM-1), an endothelial cell marker, and collagen IV, a basement membrane marker (Figure 7B). Furthermore, the chronic lower levels of ATP in skeletal muscle of Pfkm−/− mice resulted in multiple sites of muscle fiber degeneration and necrosis, characterized by inflammatory infiltration of mononucleated cells and by phagocytosis of necrotic fibers (Figure 7C). In addition, intense skeletal muscle regenerative activity was evidenced by wide distribution of centrally-located nuclei fibers in Pfkm−/− mice (Figure 7D). Thus, severe muscle fiber alterations, in addition to glycogen accumulation, result from PFKM deficiency.
In this study we show that mice with targeted ablation of the muscle isoform of PFK develop myopathic and hemolytic features similar to those noted in type VII glycogenosis in humans. The early lethality observed in Pfkm−/− mice also resembled the most severe variant of the disease, which presents in infancy and rapidly proceeds to a progressive myopathy and death [6]. Importantly, the full range of phenotypic changes we have observed in our model may impact on diagnosis and detection of human patients since phenotypic heterogeneity is common. In addition, future treatment strategies will need to consider the full extent of pathogenesis to optimize effectivity and safety.
The increased glycogen and glucose-6-phosphate in skeletal muscle observed in Pfkm−/− mice is the classic hallmark described in biopsies of human patients with GSDVII. Suppression of glycolysis impaired the use of glycogen as a fuel leading to increased storage. Moreover, blood glucose cannot be metabolized by the glycolytic pathway causing glucose-6-phosphate accumulation in skeletal muscle. Allosteric activation of glycogen synthase by glucose-6-phosphate may have contributed to increase glycogen storage [17]. Skeletal muscle uses glucose, either blood- or glycogen- derived, as the major fuel during muscular activity. The impairment of the principal catabolic pathway in skeletal muscle of Pfkm−/− mice led to energetic deprivation, which resulted in failure to perform exercise. Similarly, PFKM deficient patients show severe alterations in muscle bioenergetics leading to muscle weakness and exercise intolerance [18],[19]. Ineffective utilization of glycogen in patients with type V glycogen storage (GSDV) or McArdle's disease also leads to impairment of exercise capability. GSDV results from deficiency of the muscle isoform of glycogen phosphorylase, which leads to blockade of glycogen breakdown and to high glycogen storage in skeletal muscle [20]. However, GSDV patients show exercise tolerance after carbohydrate infusion since they can metabolize circulating glucose because glycolytic flux is preserved [21]. In contrast, in GSDVII patients, glucose infusion induces exertional fatigue attributed to an insulin-mediated decreased availability of blood free fatty acids and ketone bodies [22].
Muscle fibers of Pfkm−/− mice failed to generate enough ATP to maintain contractile activity, and mice developed muscle cramps early during the exercise test and with manipulation. In addition, even in rested state, Pfkm−/− mice showed low levels of ATP in the skeletal muscle, which is known to lead to muscle weakness and mitochondrial myopathy in other animal models [23],[24]. Physiological situations involving energy deprivation in skeletal muscle, like exercise and fasting, lead to adaptive changes towards the oxidation of fat as a fuel [25]. In skeletal muscle of Pfkm−/− mice, increased expression of oxidative marker genes and proliferation of enlarged mitochondria revealed an attempt to overcome glycolysis deficiency by shifting substrate metabolism toward a higher reliance on oxidative metabolism. Factors involved in this adaptation included PGC-1α, PPARδ and muscle CPT-1, which are responsible for mitochondrial biogenesis, oxidative phosphorylation and fatty acid oxidation [25]. Furthermore, PGC-1α and PPARδ may have been involved in structural changes towards the formation of oxidative muscle fibers by increasing the expression of MyHC-I [26],[27]. Moreover, PGC-1α up-regulation was probably responsible for the increased expression of GLUT-4 and HK-II in skeletal muscle of Pfkm−/− mice [28]. This led to enhanced glucose uptake and phosphorylation, also consistent with the high levels of glucose and glucose-6-phosphate detected in skeletal muscle of Pfkm−/− mice. In addition, the increased expression of transaldolase and transketolase enzymes suggested that glucose could be used through the pentose phosphate pathway in skeletal muscle of Pfkm−/− mice. However, despite these compensatory responses, oxidative metabolism was unable to overcome the glycolysis blockade in Pfkm−/− mice.
Anaplerosis of the tricarboxylic acid (TCA) or Krebs cycle plays a key role in oxidative metabolism in skeletal muscle by providing the TCA cycle with intermediates to permit its continued function. Impaired production of glycolytic substrates could limit oxidative metabolism by reducing concentrations of Krebs cycle intermediates [29],[30]. Blockade of glucose utilization through the glycolysis pathway in skeletal muscle of Pfkm−/− mice may lead to impaired production of the glucose-derived anaplerotic substrates phosphoenolpyruvate and pyruvate. Dysregulation of the TCA cycle intermediates probably impaired oxidative phosphorylation and the ability of skeletal muscle in Pfkm−/− mice to generate an adequate amount of ATP. The significance of the regulation of TCA cycle intermediates in the control of skeletal muscle energy metabolism has clearly been shown in mice overexpressing phosphoenolpyruvate carboxykinase (PEPCK-C). PEPCK-C transgenic mice show increased oxidative capacity in skeletal muscle leading to enhanced exercise performance [31].
GSDVII is also characterized by compensated hemolytic anemia due to reduction in the erythrocyte PFK activity. Pfkm−/− mice clearly underwent hemolysis and compensatory erythropoiesis evidenced by marked reticulocytosis. Since erythrocytes lack mitochondria, glycolysis is essential for their energy metabolism. Consequently, although erythrocytes of Pfkm−/− mice preserve about half of the PFK activity observed in wild-type mice, it was not enough to maintain erythrocyte integrity. Moreover, the kinetic properties of residual L homotetramer may turn it somehow dysfunctional in Pfkm−/− erythrocytes [32]. Removal of defective erythrocytes was probably responsible for the increased spleen size in Pfkm−/− mice. Splenomegaly has broadly been described as a result of hemolysis or hematopoietic stress in several diseases [33],[34]. Thus, increased hematopoiesis may have also contributed to increase spleen size in Pfkm−/− mice. Similar hematological features are found in spontaneous mutant mice with reduced activity of the glycolytic enzyme pyruvate kinase (Pk-1slc) in red blood cells [35].
Lower PFK activity in erythrocytes of Pfkm−/− mice led to lower concentrations of glycolytic intermediates and 2,3-BPG. In turn, low levels of 2,3-BPG increase the oxygen affinity of hemoglobin, reducing oxygen delivery to the tissues and stimulating erythropoiesis. Skeletal muscle requires large amounts of oxygen during intense exercise and alterations in the affinity of hemoglobin for oxygen could impair muscle performance [15]. Consistent with decreased oxygen availability and marked hemolysis, skeletal muscle of Pfkm−/− mice showed features of hypoxia and angiogenesis together with necrosis and intense regenerative activity. This decreased oxygen availability probably contributed to impair the compensatory oxidative metabolism in the skeletal muscle of PFK deficient mice, exacerbating its loss of functionality. In addition, changes in oxygen delivery to tissues may result in lower respiratory and cardiac function in Pfkm−/− mice.
Involvement of respiratory and cardiac muscles in the pathogenesis of GSDVII is not clearly understood. Myopathic alterations in the respiratory muscles are responsible for loss of respiratory function and even death in a wide spectrum of muscle disorders [36],[37] and other glycogen storage diseases [38],[39]. In addition, premature death due to a respiratory failure is a feature of the severe infantile form of GSDVII [6],[40]. The structural and metabolic abnormalities observed in the diaphragm and respiratory muscles of Pfkm−/− mice suggest impaired respiratory function and may have contributed to the lethality observed in these mice. On the other hand, cardiac abnormalities, such as low voltage electrocardiogram, tachycardia, ventricular hypertrophy and atrium enlargement, have only been described in a few patients [41]. Cardiac hypertrophy may result as an adaptive response to increased workload, and prolonged hypertrophy is associated with increased risk for sudden death or progression to heart failure [42]. Although most frequent causes of heart hypertrophy are chronic hypertension, exercise, myocardial infarction or aortic valve stenosis, several reports point to defects in cardiac energetic metabolism underlying heart enlargement [43]. Thus, heart specific ablation of GLUT-4 glucose transporter or deletion of the adenine nucleotide translocator-1 gene lead to heart hypertrophy in mice [24],[44]. Therefore, altered glucose metabolism in the heart of Pfkm−/− mice may have led to deficient energy production in cardiomyocyte and compensatory chronic heart hypertrophy, which probably increased mortality in these mice. These results suggest that the cardiac pathology in GSDVII may probably be underestimated or misdiagnosed [41]. In addition, this study indicates that symptoms other than classically described may be taken into consideration for the diagnostic of the GSDVII.
In summary, these results indicate that the skeletal and cardiac muscle impairments observed in Pfkm−/− mice interact with disturbed erythrocyte metabolism to produce the heterogeneous and complex pathology characteristic of type VII glycogen storage disease. The availability of this murine model of GSDVII allows determination of the role of such metabolic alterations in different tissues and organs together with their interactions, and, importantly, allows the study of GSDVII as a systemic disorder, not simply as a muscle glycogenosis. Moreover, Pfkm−/− mice constitute a unique model of GSDVII, which will most likely be very useful for the design and assessment of new therapeutic interventions for this disease.
Genomic clones for mouse Pfkm were isolated from a mouse 129/SvJ library (Stratagene). To construct the targeting vector, two fragments of the genomic DNA flanking the exon 3 were subcloned at convenient restriction sites in the pPNT vector. Linearized pPTN/pfkm was transfected into 129/SvJ derived embryonic stem cells (ES) (CMTI-1, Speciality Media). Selection was performed with G418 and gancyclovir, and resistant clones were screened for homologous recombination by Southern blot. Targeted ES cells were injected into blastocysts from C57BL/6J mice and transferred into uteri of pseudopregnant females. Chimeric males were mated to C57BL/6J females and the offspring was screened by PCR analysis using both locus-specific and Neo cassette-specific primers: PFK-Fw: 5′-AATGCACTCCGATCTGCTCC-3′; Neo: 5′-CGCCTTCTATCGCCTTCTTG ACGAGTTCTT-3′; PFK-Rev: 5′-GCAAGCAATGCCTAAATCTG-3′. Homozygous mutant mice were obtained by mating heterozygous littermates. Mice were fed ad libitum with a standard diet (Panlab, Barcelona, Spain) and maintained under a light-dark cycle of 12 h (lights on at 9:00 A.M.). Animals were killed and samples were taken between 9:00 and 10:00 A.M. In the experiments described, male mice, aged 2–3 months were used with littermates as controls. All experimental procedures involving mice were approved by the Ethics Committee in Animal and Human Experimentation of the Universitat Autònoma de Barcelona.
Total RNA was obtained from skeletal muscle samples and analyzed by Northern blot. Northern blots were hybridized to 32P-labeled pfkm cDNA probe labeled following the method of random oligopriming, as described by the manufacturer (Amersham Corp.). For real-time qPCR, 1 µg of RNA samples was used as a template to synthesize cDNA in a volume of 20 µl (Omniscript kit, Qiagen). Oligo-dT was used as a primer for the reaction in the presence of Protector RNase inhibitor (Roche). RT-PCR was performed in a SmartCycler II (Cepheid) using QuantiTect SYBR Green PCR kit (Qiagen). The sequences of the respective sense and antisense oligonucleotide primers were: Primers sequences: PGC-1α: (F) ATACCGCAAAGAGCACGAGAAG and (R) CTCAAGAGCAGCGAAAGCGTCACAG; PPARδ: (F) TCCATCGTCAACAAAGACGGG and (R) ACTTGGGCTCAATGATGTCAC; M-CPT1: (F) GCACACCAGGCAGTAGCTTT and (R) CAGGAGTTGATTCCAGACAGGTA; CS: (F) TGCCCACACAAGCCATTTG and (R) CTGACACGTCTTTGCCAACTT; HIF-1α: (F) AGCCC TAGATGGCTTTGTGA and (R) TATCGAGGCTGTGTCGACTG; PK-M: (F) CGATCTGTGGAGATGCTGAA and (R) AATGGGATCAGATGCAAAGC; LDH: (F) TGTCTCCAGCAAAGACTACTGT and (R) GACTGTACTTGACAAT GTTGGGA; GLUT-1: (F) CAGTTCGGCTATAACACTGGTG and (R) GCCCCCGACAGAGAAGATG; MyHC-I: (F) AGAGGGTGGCAAAGTCACTG and (R) GCCATGTCCTCGATCTTGTC; MyHC-IIa: (F) CGATGA TCTTGCCAGTAATG and (R) TGATAACTGAGATACCAGCG; MyHC-IIb: (F) ACAGACTAAAGTGAAAGCC and (R) CTCTCAACAGAAAGATGGAT; GLUT4: (F) GACGGACACTCCATCTGTTG and (R) CATAGCTCATGGCTGG AACC; HKII: (F) GAAGGGGCTAGGAGCTACCA and (R) CTCGGAG CACACGGAAGTT; TALDO1: (F) GATTCCAGGCCGTGTATCCAC and (R) AATCCCCTCCCAGGTTGATGA; TKT: (F) TGGCATACACAGGCAAATACTT and (R) TCCAGCTTGTAAATTCCAGCAA and 36B4: (F) GGCCCTGCACTCTCGCTTT and (R) TGCCAGGACGCGCTTGT. Data was normalized with 36B4 gene values and analyzed as previously described [45].
To determine PFK activity and the concentration of metabolites mice were anesthetized with a mixture of ketamine (100 mg/kg) and xylacine (10 mg/kg). Afterwards, skeletal muscle was freeze clamped in situ, and kept at −80°C until analysis. Diaphragm and heart were rapidly excised, weighed, frozen in liquid nitrogen and kept at −80°C. Heparinized blood samples were centrifuged, cells collected and frozen. For PFK activity, samples were homogenized in 10 volumes (1 ml/100 mg tissue) of an ice-cold buffer (pH 7.4) containing 20 mM Tris-HCl, 100 mM KCl, 5 mM MgCl2, 5 mM Phosphate Buffer and 30% Glycerol. Samples were centrifugued and PFK activity was determined in the presence of 6 mM fructose 6-phosphate and 18 mM glucose 6-phosphate by spectrophotometric analysis as previously indicated [46]. The concentrations of glycogen, glucose 6-phosphate, glucose and 2,3-BPG were measured in perchloric extracts, which were adjusted to pH 5 with 5 M K2CO3 to determine glycogen and glucose, and to pH 7 for glucose 6-phosphate and 2,3-BPG. Glycogen levels were measured using the α-amyloglucosidase method [47]. Glucose and glucose 6-phosphate concentrations were determined enzymatically [48]. 2,3-BPG was determined using a specific kit (Roche Diagnostics GMBH). The concentration of ATP and ADP was determined as described previously [49],[50]. Serum lactate dehydrogenase activity and total bilirubin and lactate levels were measured in the autoanalyzer Pentra400 (ABX Diagnostics) using specific kits (ABX Diagnostics). Glucose concentration in blood was determined by using a Glucometer Elite (Bayer) following the manufacturer's instructions.
Mice were exercised for 5 min on an enclosed treadmill LE-8708 (Panlab) supplied with an electrified grid at the rear of the belt to provide motivation. The speed of the belt was 30 cm/sec.
To determine the number of retyculocytes, blood samples were stained with new methylene blue, extended on slices, and counted under microscope. Hematopoietic cultures were performed in extracts of bone marrow and spleen. Triplicate assays were done in 35 mm plates. Samples were cultured for 7 days at 37°C, 5% CO2 and 95% relative humidity in MethoCult GF M3434 medium (StemCell Technologies Inc.). Colonies were counted under inverted microscope including CFU-GM, BFU-E, and CFU-Mix.
Skeletal muscle and heart were fixed for 12 to 24 h in formalin, embedded in paraffin and sectioned. To determine muscle morphology, sections were stained with hematoxylin/eosin. Glycogen content was analyzed by Periodic Acid Schiff (PAS) staining. Heart fibrosis was determined by Masson trichrome staining. For histochemical analysis of succinate dehydrogenase (SDH) and NADH-tetrazolium reductase (NADH-TR) activities, gastrocnemius muscle was dissected and frozen in isopentane supercooled with liquid nitrogen. Frozen sections were analyzed as previously indicated [51],[52].
For immunohistochemical detection of VEGF and collagen IV proteins, paraffin sections were incubated overnight at 4°C with rabbit anti-mouse VEGF antibody (Santa Cruz) diluted at 1∶50 and with rabbit anti-mouse collagen IV antibody (Chemicon Inc.) diluted at 1∶100. For immunohistochemical detection of PECAM-1, cryosections were incubated overnight at 4°C with rat anti-mouse PECAM-1 antibody (Pharmingen BDbiosciences) diluted at 1∶100. Afterwards, samples were incubated with the biotinylated secondary antibodies (dilution 1∶200): Rabbit against rat IgG (Vector laboratories) or goat against rabbit IgG (Vector laboratories). The localization of VEGF was determined using streptavidin conjugate Alexa fluor 488 (Molecular Probes), collagen IV using streptavidin conjugate Alexa fluor 568 (Molecular Probes) and PECAM-1 by ABC peroxidase mouse IgG staining kit (Pierce Biotechnology).
Skeletal and cardiac muscle samples were obtained and fixed in 2.5% glutaraldehyde and 2% paraformaldehyde for 2 h at 4°C. After washing in cold cacodylate buffer, the specimens were postfixed in 1% osmium tetroxide, stained in aqueous uranyl acetate, and then dehydrated through a graded ethanol series and embedded in epoxy resin. Ultrathin sections (600–800 Å) from the resin blocks were stained using lead citrate and examined in a transmission electron microscopy (Hitachi H-7000).
All values were expressed as mean±SEM. Two-tailed P values were calculated by unpaired Student's t test. Differences were considered statistically significant at P values less than 0.05.
|
10.1371/journal.pntd.0001041 | A Phase Two Randomised Controlled Double Blind Trial of High Dose
Intravenous Methylprednisolone and Oral Prednisolone versus Intravenous Normal
Saline and Oral Prednisolone in Individuals with Leprosy Type 1 Reactions and/or
Nerve Function Impairment | Leprosy Type 1 reactions are a major cause of nerve damage and the
preventable disability that results. Type 1 reactions are treated with oral
corticosteroids and there are few data to support the optimal dose and
duration of treatment. Type 1 reactions have a Th1 immune profile: cells in
cutaneous and neural lesions expressing interferon-γ and interleukin-12.
Methylprednisolone has been used in other Th1 mediated diseases such as
rheumatoid arthritis in an attempt to switch off the immune response and so
we investigated the efficacy of three days of high dose (1 g) intravenous
methylprednisolone at the start of prednisolone therapy in leprosy Type 1
reactions and nerve function impairment.
Forty-two individuals were randomised to receive methylprednisolone followed
by oral prednisolone (n = 20) or oral prednisolone
alone (n = 22). There were no significant differences
in the rate of adverse events or clinical improvement at the completion of
the study. However individuals treated with methylprednisolone were less
likely than those treated with prednisolone alone to experience
deterioration in sensory function between day 29 and day 113 of the study.
The study also demonstrated that 50% of individuals with Type 1
reactions and/or nerve function impairment required additional prednisolone
despite treatment with 16 weeks of corticosteroids.
The study lends further support to the use of more prolonged courses of
corticosteroid to treat Type 1 reactions and the investigation of risk
factors for the recurrence of Type 1 reaction and nerve function impairment
during and after a corticosteroid treatment.
Controlled-Trials.comISRCTN31894035
| Leprosy is caused by a bacterium and is curable with a combination of antibiotics
known as multi-drug therapy that patients take for six or 12 months. However a
significant proportion of leprosy patients experience inflammation in their skin
and/or nerves, which may occur even after successful completion of multi-drug
therapy. These episodes of inflammation are called leprosy Type 1 reactions.
Type 1 reactions are an important complication of leprosy because they may
result in nerve damage that leads to disability and deformity. Type 1 reactions
require treatment with immunosuppressive agents such as corticosteroids. The
optimum dose and duration of corticosteroid therapy remains unclear. We
conducted a study to see if it would be safe to use a large dose of a
corticosteroid called methylprednisolone for three days at the start of a 16
week corticosteroid treatment regime of prednisolone in patients with leprosy
Type 1 reactions and leprosy patients with nerve damage present for less than
six months. We did this by comparing individuals who were given
methylprednisolone followed by prednisolone and those who received just
prednisolone. In this small study we did not see any significant difference in
the frequency of adverse events due to corticosteroid treatment in the two
groups. We did not demonstrate a significant difference in improvement in
individuals in the methylprednisolone group (who received a larger dose of
corticosteroids) than those in the prednisolone treated group. Overall
approximately 50% of individuals required more prednisolone in addition
to the 16 week course of treatment to prevent further nerve damage or reactions.
This suggests that it would be worthwhile investigating longer treatment courses
with corticosteroids and other immunosuppressive drugs.
| Leprosy is a chronic granulomatous infection principally affecting the skin and
peripheral nerves caused by the obligate intracellular organism
Mycobacterium leprae
[2]. The disease
causes skin lesions and neuropathy. Complications secondary to the neuropathy can
result in deformity and disability. 249 007 new cases of leprosy were diagnosed and
reported to World Health Organization (WHO) in 2008 [3].
Type 1 reactions (T1Rs) are a major cause of nerve function impairment (NFI) in
patients with leprosy and affect up to 30% of susceptible individuals [4]. T1Rs
predominantly affect borderline leprosy[4]. They may be a presenting
feature of leprosy or occur during multi-drug therapy (MDT) or after completion. A
T1R is characterised by acute inflammation in skin lesions or nerves or both. Skin
lesions become acutely inflamed and oedematous and may ulcerate. Oedema of the
hands, feet and face can also be a feature of a T1R.
Leprosy T1Rs are treated with oral corticosteroids. However treatment with a
standardised reducing 12 week course of oral prednisolone (total dose 1.68 g) which
had been used in a previous pilot study in Nepal resulted in 37% of
individuals requiring additional prednisolone [5]. The randomised controlled
treatment trials TRIPOD 2 and TRIPOD 3 that were reported during the design of this
study had both used a 16 week course of oral prednisolone (total dose 2.52 g) [6], [7].
T1Rs appear to be mediated via Th1 type cells and lesions in reaction express the
pro-inflammatory IFNγ, IL12 and the oxygen free radical producer iNOS [8]. The expression
of TNFα protein in the skin and nerves of patients during T1Rs is increased[9]. High
dose intravenous (IV) methylprednisolone (MP) has been used as a standard treatment
in the early phase of an exacerbation of Th1 cytokine mediated relapsing chronic
diseases. These conditions include rheumatoid arthritis (RA) [10] and multiple sclerosis (MS)
[11]. In 18
patients with MS treated with IV MP 1 g for three days there was a significant
suppression of mitogen stimulated IFNγ, TNFα and IL2 production by blood
leucocytes ex vivo after treatment [12]. MP has also been shown to
reduce serum levels of TNFα in RA [13]. Eleven patients given 1 g IV
showed significantly reduced serum levels of TNFα at 4 and 24 hours. In a
comparative study of lymphocyte-suppressive potency between prednisolone and MP in
44 individuals with RA the latter was more effective in those with greater disease
activity as defined by rheumatoid factor titres [14].
We compared three daily infusions of IV high dose MP and oral prednisolone with a 16
week course of oral prednisolone alone. High dose IV MP had not been used previously
in a trial of treatment of leprosy T1R so a Phase 2 trial was needed to confirm
safety before considering whether to proceed to a larger Phase 3 trial of clinical
efficacy.
The aims of the trial were as follows:
A double blind parallel-group randomised controlled trial was designed to compare the
safety and effect of early high dose IV MP followed by oral prednisolone with IV
Normal saline and oral prednisolone. The study was approved by the Nepal Health
Research Council and the Ethics Committee of the London School of Hygiene and
Tropical Medicine (Number 4022).
Participants (aged between 16–65 years and weighing more than 30 kg) were
recruited from the leprosy service of Anandaban Hospital, Kathmandu, Nepal. Two
groups of individuals were eligible for entry into the trial:
Participants with any type of leprosy of the Ridley-Jopling Classification [15] were eligible.
Initially, enrolment into the study required individuals with clinical evidence of a
T1R to have associated nerve function impairment. This was changed nine months after
the start of the trial so that individuals with T1Rs involving the skin only would
also be eligible for enrolment. This was done because only 14 individuals had been
recruited in this period and recruitment had been optimal as determined by case note
review of a random selection of clinic attendees. The change to this eligibility
criterion was approved by the two Ethics committees.
The following individuals were excluded: those unwilling to give consent or return
for follow-up or who had taken systemic steroids within three months of enrolment,
those who had received other immunosuppressant therapy including thalidomide within
three months of enrolment, those with severe active infection such as tuberculosis
or severe intercurrent disease, those with a contraindication to high dose
methylprednisolone such as peptic ulcer disease, diabetes mellitus, glaucoma and
uncontrolled hypertension or known allergy to methylprednisolone. Pregnant women
were excluded and females of child bearing capacity were not recruited unless they
had at least one month of adequate contraception.
The participants were treated with corticosteroids for 112 days. The total duration
of the study was 337 days from entry into the trial. The intervention for the MP
treated individuals was 1 gram MP in Normal saline given as an IV infusion and eight
dummy tablets (Comprehensive Medical Services India, Chennai India) identical in
appearance to prednisolone tablets daily for the first three days of the trial. The
prednisolone treated individuals received 40 mg (eight tablets) of prednisolone and
an identical appearing IV infusion which contained only Normal saline daily for the
first three days of the trial. Thereafter individuals in both groups received the
same reducing course of prednisolone. This course was prednisolone 40 mg daily from
day 4 to day 14 of the study. The amount of prednisolone was then reduced to 35 mg
daily for the next 14 days and then by a further 5 mg daily every 14 days to zero.
An individual allocated to the MP group received a total dose of corticosteroid
equivalent to 6.15 g of prednisolone. Individuals in the prednisolone alone group
received 2.52 g of prednisolone in total.
All individuals enrolled into the study received albendazole 400 mg daily for the
first three days of the trial and famotidine 40 mg daily whilst they were receiving
corticosteroids. The albendazole was given to reduce the risk of hyperinfection with
Strongyloides stercoralis. The famotidine was used to reduce
the risk of peptic ulceration.
The primary outcome measure was the frequency of adverse events in the two treatment
arms. These were assessed by a study physician prior to treatment and then at day 4
(after the three IV infusions) and then days 8, 15, 29, 57, 85, 113, 141, 169, 197,
225, 253, 281, 309 and 337. Adverse events were enquired about and examined for at
each assessment. A standardized form contained a list of adverse events attributable
to corticosteroids which participants were asked if they had experienced. There was
also a free text space available where other symptoms mentioned by the study
participants or identified by the physician could be recorded. Adverse events were
defined as major or minor in accordance with the classification used in the TRIPOD
studies [16].
Major adverse events were defined as psychosis, peptic ulcer, glaucoma, cataract,
diabetes mellitus, severe infections (including tuberculosis), infected neuropathic
ulcers, hypertension and death. Minor adverse events were defined as moon face,
dermatophyte fungal or yeast infections, acne and gastric pain requiring an antacid
(in addition to the famotidine each individual was prescribed whilst on
corticosteroids). Individuals were questioned about the symptoms of nocturia,
polyuria and polydipsia as a method of screening for diabetes mellitus in addition
to urinalysis being performed.
Secondary outcomes measures were:
Peripheral nerve function was assessed clinically. Sensory testing (ST) was performed
using two Semmes-Weinstein monofilaments (SWM) (Sorri-Bauru, Bauru, São
Paulo, Brazil) at designated test sites on the hands and feet as previously reported
[17]. The
sensation in the areas of skin supplied by the ulnar and median nerves was tested
with 2 g and 10 g monofilaments. The area of skin supplied by the posterior tibial
nerve was tested with the 10 g and 300 g monofilaments. Trigeminal nerve sensation
was tested using cotton wool. Voluntary muscle testing (VMT) was assessed using the
modified Medical Research Council grading of power [18]. The facial nerve was tested by
assessing forced eye closure. The median nerve was tested using resisted thumb
abduction, the ulnar nerve by resisted little finger abduction and the radial nerve
by resisted wrist extension. The lateral popliteal nerve was tested by resisted foot
dorsiflexion. ST and VMT assessments were carried out by trained physio-technicians
and if necessary repeated by the study physicians.
Patients with deterioration in nerve function or skin signs were treated with further
prednisolone. This was defined as a sustained deterioration (for a period of at
least two weeks) of nerve function, the development of nerve pain unresponsive to
analgesics, palpable swelling of skin patches or new erythematous and raised skin
patches. Any decline in nerve function which the study doctors believed required
immediate additional prednisolone was also regarded as deterioration. Individuals
who experienced deterioration in skin and/or nerve function whilst receiving a dose
of prednisolone less than 20 mg daily had the dose increased back to 20 mg and
reduced by 5 mg every 14 days to zero. The exception to this was if they had a T1R
involving a facial patch in which case the prednisolone was increased to 40 mg
regardless of the dose of prednisolone the individual was taking. Individuals taking
a dose of prednisolone greater than 20 mg had their dose increased to 40 mg and
tapered by 5 mg every 14 days to zero.
In order to have 80% power to show that MP was not associated with a
significantly greater (α<0.05) rate of major adverse effects it was
calculated that the study would need 201 participants in each group based on a
higher rate of 7%. Using this same assumption but with the TRIPOD data for
all the Nepali participants (major adverse effect rate of 2.4%) then 64
individuals would be needed to be enrolled in each arm.
Eligible individuals were enrolled consecutively. Block randomisation in groups of
four using a table of random numbers generated by Dr Peter Nicholls was used. A
standard envelope system was used for allocation concealment. The envelopes were
pre-packed in London by Dr Claire Watson who had no other involvement with the
study. The participants were randomly allocated to the MP/prednisolone or the
prednisolone alone arm and so had an equal chance of being in either arm of the
study. The allocation procedure was decentralized and operated solely by the chief
pharmacist at Anandaban Hospital who kept a separate record of the allocation. The
pharmacist had no contact with the study participants during their inpatient
stay.
All study participants, physicians, ward staff and other assessors
(physio-technicians) were blinded to the allocation. Only Dr Peter Nicholls had
access to the study data and the randomisation code. The allocation code was
revealed to the other researchers once recruitment, follow-up and data collection
had been completed.
The data were stored in an Access database and analysed using the Statistical Package
for the Social Sciences (SPSS version 16 SPSS Inc., Chicago, Illinois). An intention
to treat analysis was used for calculating the effects of treatment on individuals
in each group.
The trial was registered with Current Controlled Trials Ltd (www.controlled-trials.com) in accordance with the policy of the
International Committee of Medical Journal Editors [1] and was assigned the unique
identifier ISRCTN31894035. The protocol for the trial can be accessed as a
supplementary file Protocol S1 to this publication.
Forty-two patients were enrolled into the trial between 7th December 2005
and 31st December 2007. The final assessment and data entry was completed
on 5th November 2008. The participants flow through the study is
illustrated in the CONSORT flow diagram (Figure 1).
Thirty-three males and nine females were recruited. Twenty-two individuals were
randomised to receive prednisolone only. There were no statistically significant
differences between the groups with respect to gender, age, Ridley-Jopling
classification, or treatment with MDT (Table 1). The two groups did not differ
significantly in terms of the nature of the reaction, the type of NFI at baseline or
the pattern of old (> 6 months duration) NFI.
Eight participants (19%) did not complete the full schedule of follow-up. Five
were randomised to the prednisolone arm and three received MP. Efforts were made to
get these individuals to attend by telephoning or writing to them but without
success. Two of these individuals stopped attending whilst on corticosteroids.
Table 2 shows the number of
individuals who experienced a particular adverse event. Twenty-three participants
experienced at least one adverse event, twelve (54.5%) in the prednisolone
arm and 11 (55%) in the MP arm. Seven individuals experienced more than one
adverse event. There were no statistically significant differences in the number of
individuals experiencing a given adverse event between the two groups of the
study.
Two individuals (one from each arm of the study) experienced a major adverse event.
One was diagnosed with glaucoma and the other with infected neuropathic ulcers. None
of the participants developed hypertension, tuberculosis or diabetes mellitus. The
risk ratio of having an adverse event (of any type; major or minor) given that the
participant received MP was 1.0083 (95% CI: 0.5817 to 1.7480;
p = 0.9764) compared to prednisolone.
Individuals were most likely to experience an adverse event whilst taking the first
course of corticosteroids between days 1 and 112. Figure 2 is a Kaplan-Meier survival curve showing
the cumulative “survival” probability (i.e. not having an adverse event)
for individuals in each group. There was no significant difference between the two
groups (Log Rank (Mantel-Cox) 0.945).
Four individuals had their first adverse event after the initial study intervention
had been completed (post day 112). Two others had a new adverse event after the
intervention period. Two individuals experienced an adverse event, weight gain and
infected neuropathic ulcers respectively, whilst not taking corticosteroids.
The total clinical severity scores, calculated using the validated scale, for each
arm of the study at day 1 (enrolment) and days 4, 29,113 and 337 are shown using
boxplots (fig.3). There was a
downward trend in the total clinical severity scores of both groups. There were no
statistically significant differences between the prednisolone and MP groups at any
time point.
There was no significant difference in the median sensory scores (corrected for
impairment >6 months) of individuals in the two groups at baseline. Both groups
showed a downward trend in the sensory scores during treatment but there were no
significant differences at any of the pre-specified time points. The Kaplan-Meier
survival analyses of deterioration in sensory score during the study to days 29, 113
and 337 (fig.4) demonstrate that
there is no difference between the groups at day 29 but at day 113 there was a
significant difference in the probability of deterioration in sensation between
individuals in the two arms of the study (p = 0.046). Patients
in the prednisolone alone group were more likely to experience deterioration in
sensation between day 30 and day 113. This effect is not maintained at the end of
the study follow-up period at day 337. The motor scores of the two groups at
baseline are not significantly different. They showed a downward trend during the
course of the study. There are no significant differences between the scores of the
group at any of the time points. There were no significant differences between the
groups in the probability of an individual experiencing deterioration in motor
function at days 29, 113 or 337.
Figure 5 shows events when
additional steroid was prescribed and censoring individuals who were unavailable for
further assessment or who received prednisolone either inappropriately or for ENL.
There was no significant difference in the probability of being prescribed
additional prednisolone between the two groups (Log Rank (Mantel Cox)
p = 0.126). The amount of additional prednisolone required by
individuals randomised to either treatment group did not differ significantly. The
mean amount of additional prednisolone prescribed during the study was 1252.5 mg
(SD±1862.0) for the MP group and 1432.7 mg (SD±1245.9) for the
prednisolone alone group (p = 0.718).
Twenty individuals (47.6%) required additional prednisolone because they
experienced a deterioration of nerve function (n = 11) or a
recurrence of a T1R (n = 6) or both
(n = 3). Two individuals received additional prednisolone
inappropriately and two developed ENL requiring prednisolone. Five of the 20
individuals (appropriately prescribed additional prednisolone for a trial
indication) required prednisolone before day 112, the last day of the intervention
period. The median time to requiring additional prednisolone for these individuals
was 61 days (range = 14–105) after enrolment when
individuals were receiving prednisolone 20 mg daily. The other 75% had
finished the prednisolone before experiencing a deterioration requiring further
treatment. The median number of days between finishing the study intervention (day
112) and requiring additional prednisolone was 63 days
(range = 2–224).
Analysing the additional corticosteroid requirement by Ridley-Jopling classification
fifty-two percent (12 of 23) of individuals with borderline tuberculoid (BT)
leprosy, 67% (two of three) of individuals with borderline borderline (BB)
leprosy, 38% (five of 13) of those with borderline lepromatous leprosy (LL)
and 50% (one of two) of lepromatous leprosy patients required additional
prednisolone for a trial indication (those with ENL were excluded). Two of the BT
patients had positive slit-skin smears. The median time from enrolment to the
deterioration requiring additional prednisolone was 152 days for BT patients, 138
days for BB patients, 125 days for BL patients and 313 days for those with LL. There
were no significant differences in the proportion of individuals with a particular
Ridley Jopling classification or the time to requiring additional prednisolone.
In this small, study the occurrence and timing of minor and major adverse events did
not differ significantly between the prednisolone and the MP treated groups. The
study was underpowered and limited the ability to detect significant differences of
less than 30% between the groups. Twenty-one (50%) individuals
experienced at least one minor adverse event and two (4.8%) a major adverse
outcome. In the TRIPOD trials 8.4% (14/167) of the prednisolone treated
Nepali cohorts experienced a minor adverse event[16]. This was not significantly
different from the placebo treated group. The individuals in these groups were
treated with either 1.96 g or 2.52 g of prednisolone depending on which of the three
trials they were enrolled into.
The two major adverse events that occurred during the study were glaucoma and
infected neuropathic ulcers but these were probably not due to the trial
medications. One individual developed glaucoma at day 305. He developed ENL at day
111. ENL like corticosteroid therapy is a recognised cause of secondary glaucoma. He
required continuous oral prednisolone (receiving a total additional dose of 2.87 g
of prednisolone between days 111 and 305) despite treatment of his ENL with high
dose (300 mg daily) clofazimine. The majority of individuals who develop ENL require
long term treatment and many become corticosteroid dependent [19]. There were no cases of
glaucoma in any of the TRIPOD participants. Infected neuropathic ulcers affected one
individual treated with MP. This occurred 58 days after this man completed the trial
intervention. Two individuals in the TRIPOD studies (one from the prednisolone
treated group) developed infected ulcers. It is not reported whether the
prednisolone treated person was taking the drug at the time the infection was
diagnosed.
The symptoms of nocturia, polyuria and polydipsia were reported by four (9.5%)
of individuals. The two individuals who had glycosuria did not complain of these
symptoms. Their glycosuria was not persistent and therefore not considered to be
clinically significant. The two individuals were both receiving additional
prednisolone at the time but neither had received MP. There were no individuals in
the study diagnosed with diabetes mellitus. The TRIPOD 1 study reported one
individual from the prednisolone treated group who developed glycosuria. This was
considered a major adverse event in that study but the authors did not comment
whether this patient was diagnosed with diabetes mellitus [20]. Three individuals in the
steroid treated groups of the three TRIPOD studies developed diabetes mellitus
compared with one in the placebo groups but this difference was not significant
[16].
The size of the study limited our ability to detect rare adverse events however a
much higher rate of acne and moon face was recorded than the TRIPOD studies. Another
factor that might have reduced our estimation of adverse events is the duration of
follow-up which may have been too short, however most studies have assumed that
adverse events will occur during the treatment phase predominantly. We were also
unable to examine the effect of our interventions on bone density which may be
significantly affected by corticosteroid therapy in the doses and durations commonly
used to manage leprosy T1R and NFI. The findings would support the view that MP, in
the doses used in the study, is safe.
MP did not appear to have a larger therapeutic effect than prednisolone alone on the
symptoms and signs of leprosy T1Rs and NFI in this study. The use of a validated
scale to measure leprosy T1Rs and NFI allows the comparison of the two groups in
this study. There were no significant differences in the total severity score or the
sensory or motor scores between the prednisolone and MP treated groups at any of the
pre-defined time points. However there was a trend towards improvement in sensory
and motor scores during the study. Participants in the prednisolone treated group
were significantly more likely to have experienced deterioration in sensory function
than the MP treated group by the end of the intervention (day 113). However this
difference was not sustained to the end of the study. This effect may have occurred
by chance as it was not reproduced in the skin or in motor function. The number of
participants contributing to all of the survival analyses towards the end of the
study is small and the results therefore less reliable. This phenomenon of
deterioration after stopping corticosteroids is similar to the results of the TRIPOD
1 study of prednisolone given to patients as prophylaxis to prevent the occurrence
of reactions and NFI. It demonstrated a protective effect of prednisolone compared
with placebo during the 16 weeks of treatment which was lost by 48 weeks. The higher
dose may have a greater effect whilst an individual is receiving corticosteroids but
not once they are no longer taking the drug.
Forty-five per cent of the MP group and 50% of the prednisolone alone group
were prescribed additional prednisolone. Of the 20 individuals who required
additional prednisolone 12 (60%) did not do so until at least 28 days after
completing the trial intervention. The clinical nature of the deterioration (skin or
nerves or both) did not differ significantly between those who experienced it whilst
receiving the study intervention and those who experienced deterioration after
completing it (χ2 = 0.292). The delay in
deterioration in the majority of individuals requiring additional prednisolone is
similar to that seen in the TRIPOD 1 study[20].
After the start of this trial data suggesting that more prolonged courses of
prednisolone may be more effective in treating T1Rs were published. The requirement
for extra prednisolone was used as the outcome measure in the multi-centre double
blind randomised controlled trial of three different prednisolone regimens conducted
in India [21]. The
proportion of individuals requiring additional prednisolone in the three groups was
24%, 31% and 46% respectively. Individuals who received
prednisolone for 20 weeks were significantly less likely to require additional
steroid than those treated for 12. However this does not necessarily reflect
clinical improvement. The decision to use additional prednisolone was left to the
individual clinician's judgement at each of the six centres. It is not clear
how consistency was ensured between individual physicians or at different stages of
the trial. The protocol of the MP study was stringent in treating NFI.
“Mild” deterioration in NFI and NFI of short duration were both treated.
Any sustained (as little as one week) deterioration in monofilament testing at even
a single test site was an indication for additional prednisolone and so a lower
threshold for defining deterioration is likely to have been employed in the current
study. This may in part account for the high proportion of individuals who received
additional prednisolone. It is likely that some of the change labelled as
deterioration was due to test response variability. In the TRIPOD 2 cohort
27% of prednisolone treated individuals with mild sensory impairment
experienced deterioration necessitating additional prednisolone. A group with mild
isolated sensory impairment would be expected to require less additional
prednisolone than a group that included severe nerve impairment both sensory and
motor and marked skin involvement.
The results of this small study should be interpreted with caution but it would
appear that given the available data MP does not result in an increase in the number
or severity of adverse events in individuals with leprosy in Nepal. However close
detailed adverse event recording would still be required in any future studies of MP
in this setting. The establishment of registries of corticosteroid treated patients
at specialised centres could facilitate the collection of reliable adverse event
data without the need to resort to more costly randomised controlled trials.
The clinical outcome of patients in the two arms of this study was not significantly
different in terms of the validated clinical severity scale. The MP treated group
had significantly less deterioration in sensory function during the 112 days of
corticosteroid therapy but this was not maintained to the end of the 337 day
follow-up period. This may be a reflection of the small numbers in the study,
particularly towards the end of follow-up. A much larger study would be required to
examine this potential effect further. However given the high proportion of
individuals (who received MP) requiring additional prednisolone and the data
published by Rao and colleagues[21] we do not think further clinical trials of high dose IV
MP are warranted at present. Any future studies must also take into account the
greater cost of giving intravenous treatment and its acceptability to patients.
This study has highlighted that corticosteroid treatment for T1R and NFI is
sub-optimal even when given in large doses for 16 weeks. The majority of patients
who experienced a “re-reaction” required additional prednisolone after
the 16 week corticosteroid intervention had ended. It adds further support to the
argument that treatment should be given for longer durations. Investigating risk
factors for requiring additional prednisolone and the differences between those who
have deterioration in symptoms whilst taking corticosteroids and those whose
deterioration occurs later (or not at all) might enable clinicians to identify those
individuals who might benefit from prolonged corticosteroid treatment at the outset.
At present there is convincing evidence for corticosteroid regimes of at least 20
weeks [21] but some
would argue for 24 weeks [22] and others even longer [23]. The development of more
prolonged treatment protocols would require further monitoring of adverse events and
in particular the long term sequelae of corticosteroid therapy. However studies with
adequate power using improvement in nerve function as the primary outcome of the
effect of corticosteroids and other agents need to be conducted.
|
10.1371/journal.pgen.1006748 | Dysregulation of INF2-mediated mitochondrial fission in SPOP-mutated prostate cancer | Next-generation sequencing of the exome and genome of prostate cancers has identified numerous genetic alternations. SPOP (Speckle-type POZ Protein) was one of the most frequently mutated genes in primary prostate cancer, suggesting SPOP is a potential driver of prostate cancer development and progression. However, how SPOP mutations contribute to prostate cancer pathogenesis remains poorly understood. SPOP acts as an adaptor protein of the CUL3-RBX1 E3 ubiquitin ligase complex that generally recruits substrates for ubiquitination and subsequent degradation. ER-localized isoform of the formin protein inverted formin 2 (INF2) mediates actin polymerization at ER-mitochondria intersections and facilitates DRP1 recruitment to mitochondria, which is a critical step in mitochondrial fission. Here, we revealed that SPOP recognizes a Ser/Thr (S/T)-rich motif in the C-terminal region of INF2 and triggers atypical polyubiquitination of INF2. These ubiquitination modifications do not lead to INF2 instability, but rather reduces INF2 localization in ER and mitochondrially associated DRP1 puncta formation, therefore abrogates its ability to facilitate mitochondrial fission. INF2 mutant escaping from SPOP-mediated ubiquitination is more potent in prompting mitochondrial fission. Moreover, prostate cancer-associated SPOP mutants increase INF2 localization in ER and promote mitochondrial fission, probably through a dominant-negative effect to inhibit endogenous SPOP. Moreover, INF2 is important for SPOP inactivation-induced prostate cancer cell migration and invasion. These findings reveal novel molecular events underlying the regulation of INF2 function and localization, and provided insights in understanding the relationship between SPOP mutations and dysregulation of mitochondrial dynamics in prostate cancer.
| Prostate cancer is the leading cause of global cancer-related death. The development of improved diagnoses and novel therapies has been confounded by significant patient heterogeneity. During recent years, significant progress has been made in identifying the molecular alterations in prostate cancer using next-generation sequencing. SPOP gene was frequently altered by somatic point mutations in a distinct molecular subclass of prostate cancer, although the precise role that SPOP mutation plays in the development of prostate cancer is unclear. Mitochondria are highly motile organelles that undergo constant fission and fusion. Unbalanced mitochondrial fission and fusion events are associated with mitochondrial dysfunction and frequently linked to human cancer. Here, we are the first to report that SPOP mutations are associated with dysregulation of mitochondrial dynamics in prostate cancer and this finding may have potential clinical implications in prostate cancer treatment.
| Large-scale exome/genome sequencing studies have recently revealed that recurrent mutations in the SPOP gene occur in up to 15% of prostate cancers [1–4]. Interestingly, the SPOP mutant subset of prostate cancers had some notable molecular features, including mutual exclusivity with ERG gene rearrangement, elevated levels of DNA methylation, homogeneous gene expression patterns, frequent deletion of CHD1 and overexpression of SPINK1 mRNA, supporting the concept that SPOP mutation tumors represent a distinct molecular subclass of prostate cancer [4] SPOP is one of the adaptor proteins of the CUL3-RBX1 E3 ubiquitin ligase complexes. It selectively recruits substrates via its N-terminal MATH domain, whereas its BTB and BACK domains mediate oligomerization and interaction with CUL3 [5]. SPOP has been linked to the ubiquitination and degradation of several substrates, including the steroid receptor coactivator 3 (SRC-3), androgen receptor (AR), DEK, ERG, SENP7 and several others [6–11]. All prostate cancer-associated SPOP mutations identified so far affect evolutionarily conserved residues in the MATH domain, suggesting that these mutations may alter the interaction of SPOP with its substrates [1–4]. Inactivation of SPOP by knockdown or overexpression of prostate cancer-associated SPOP mutants leads to increased prostate cancer cell proliferation, migration and invasion, implying SPOP is a tumor suppressor [2,8–10]. However, limited numbers of SPOP substrates have been identified and functionally explored.
Mitochondria are highly motile organelles that undergo constant fission and fusion, and are actively transported to specific subcellular locations [12]. Unbalanced mitochondrial fission and fusion events are associated with mitochondrial dysfunction and frequently linked to the pathogenesis of many human diseases, including cancer [12,13]. The majority of studies that have explored mitochondrial morphology in tumor cells support a pro-tumorigenic role for mitochondrial fission and tumor suppressor role for mitochondrial fusion [14]. Mitochondrial fragmentation has been observed in various types of tumor cells [15–17]. Inhibition of mitochondrial fission decreases cell proliferation, migration and invasion in various cancer models including lung, colon, breast, thyroid cancer and glioblastoma[16–20]. While cancer is a disease characterized by multiple genetic aberrations, little is known about whether cancer-associated mutations can directly affect mitochondrial dynamics, and how this impacts upon tumor phenotypes.
Inverted formin 2 (INF2) is a unique vertebrate formin protein that accelerates both actin polymerization and depolymerization [21]. In mammalian cells, INF2 can be expressed as two C-terminal splice variants: the prenylated (CAAX) isoform, which is tightly bound to endoplasmic reticulum (ER) [22], and the nonCAAX isoform, which is cytoplasmic [23]. Recent studies have persuasively showed in mammalian cells that actin polymerization mediated by ER-localized INF2 CAAX isoform is required for mitochondrial fission [24]. By contrast, the cellular function of the nonCAAX isoform of INF2 has been less characterized. Suppression of INF2-nonCAAX isoform in cells causes Golgi dispersal, suggesting INF2 might be involved in maintenance of Golgi architecture [23]. Mutations in INF2 are linked to two human genetic diseases: focal and segmental glomerulosclerosis (FSGS), a degenerative kidney disease [25], and Charcot-Marie-Tooth disease (CMTD), a neurological disorder [26]. However, little is known about how INF2 protein is physiologically regulated.
In this study, we demonstrate that SPOP suppresses mitochondrial fission by promoting atypical ubiquitination and relocalization of ER-localized INF2. Moreover, this effect is abrogated by the prostate cancer-associated SPOP mutations. Thus, our results provide a functional link between SPOP mutations and dysregulation of mitochondrial dynamics in prostate cancer.
To identify molecular mediators of the tumor suppressive function of SPOP, we performed a yeast two-hybrid screen in a human fetal brain cDNA library using the full length SPOP as bait. Among the positive clones identified, 4 clones were INF2 fragments. Considering INF2 is an important regulator of actin polymerization and mitochondrial fission, we explored whether INF2 is an authentic SPOP substrate and its function is dysregulated in SPOP-mutated prostate cancer. We first examined whether SPOP interacts with INF2 in cells. To do this, FLAG-INF2, and Myc-SPOP were co-expressed in 293T cells. Cell lysates were subsequently prepared for co-immunoprecipitation (co-IP) with anti-FLAG antibody. As shown in Fig 1A, Myc-SPOP was immunoprecipitated by FLAG-INF2, suggesting an interaction between SPOP and INF2 proteins. Similar results were also obtained in the reciprocal co-IP experiment in which FLAG-SPOP was able to immunoprecipitate Myc-INF2(Fig 1B). FLAG-SPOP was able to immunoprecipitate endogenous INF2, and two known SPOP substrates (AR and DEK) in LNCaP cells (Fig 1C). Next, we decided to extend our analysis by investigating whether endogenous SPOP and INF2 can interact with each other in prostate cancer cells. Immunoprecipitation using anti-INF2 antibody was performed using cell lysates prepared from LNCaP cells. As shown in Fig 1D, INF2 was able to immunoprecipitate SPOP and a known interactor IQGAP1, suggesting that SPOP can interact with INF2 protein at endogenous level.
SPOP contains two structural domains: a substrate-binding MATH domain at the N-terminus and a CUL3-binding BTB domain at the C-terminus. To determine which domain may mediate its interaction with INF2, we generated two deletion mutants of SPOP (SPOP-ΔBTB and ΔMATH), corresponding to the deletion of these two domains respectively (Fig 1E). Co-IP assay was performed to examine the binding of INF2 with the full length SPOP (SPOP-WT) and the two deletion mutants. As shown in Fig 1F, SPOP-WT and SPOP-ΔBTB, but not SPOP-ΔMATH interacted with INF2. Therefore, our findings demonstrate that SPOP binds INF2 via the MATH domain.
We then explored whether SPOP can promote the ubiquitination and degradation of INF2. Unexpectedly, overexpression of wild-type SPOP or its mutants (SPOP-ΔBTB, ΔMATH) did not alter the protein level of ectopically co-expressed INF2 (Fig 2A). Moreover, we found that ectopic expression of SPOP in LNCaP or DU145 prostate cancer cells did not alter the protein level of endogenous INF2 (Fig 2B). In contrast, SPOP overexpression in LNCaP cells (AR positive) decreased the expression of endogenous AR, a known SPOP substrate (Fig 2B) [7,27]. Consistent with these findings, depletion of endogenous SPOP by two independent shRNAs did not alter INF2 protein level in both LNCaP and DU145 (AR-negative) cells, but elevated AR protein level in LNCaP cells (Fig 2C). Thus, these results demonstrate that SPOP does not affect INF2 protein level. To determine whether SPOP regulates INF2 polyubiquitination, HA-Ub and FLAG-INF2 were co-expressed in 293T cells with increasing doses of SPOP-WT or its mutants (SPOP-ΔBTB, ΔMATH). As shown in Fig 2D, INF2 protein was robustly polyubiquitinated by co-expression of SPOP-WT, but not SPOP-ΔBTB or ΔMATH, in a dose-dependent manner. Accordingly, depletion of SPOP in LNCaP cells decreased the ubiquitination of endogenous INF2 (Fig 2E). Since the INF2 construct used in above analysis is the CAAX isoform, we examined whether SPOP can ubiquitinate the nonCAAX isoform. As shown Fig 2F, INF2 nonCAAX isoform was also robustly polyubiquitinated by SPOP. Taken together, our data suggest that SPOP can promote INF2 ubiquitination, but not degradation.
We then examined the linkage specificity of SPOP-mediated INF2 ubiquitination. In vivo ubiquitination assay was performed using a panel of ubiquitin mutants containing a single K/R mutation on each of the seven lysines in the ubiquitin sequence, potentially involved in the formation of polyUb chains. We also included a lysine-free ubiquitin mutant (K-ALL-R), in which all of the lysines were replaced with arginines. As shown in Fig 2G, expression of the K-ALL-R mutant abolished SPOP-mediated INF2 ubiquitination, excluding the possibility that SPOP promotes multiple mono-ubiquitination of INF2. Expression of K6R or K11R mutant marginally altered the amount of ubiquitinated INF2 (Fig 2G), suggesting that K6 and K11 are largely dispensable for SPOP-mediated INF2 ubiquitination. By contrast, a significant reduction of INF2 ubiquitination is instead observed when other ubiquitin mutants, including K27R, K29R, K33R, K48R and K63R, were used (Fig 2G). We next used a reciprocal series of mutants, where all the seven lysines in ubiquitin were converted to arginine residues, except one (one-K-Only mutants). As shown in Fig 2H, expression of K27O, K29O, K33O, K48O or K63O mutants completely abolished SPOP-mediated INF2 ubiquitination. Therefore, these data indicate that SPOP catalyzes synthesis of mixed-linkage polyUb chains on INF2, and K27, K29, K33, K48 and K63 residues in Ub are all essentially involved.
Having established that SPOP promotes atypical ubiquitination of INF2, we set out to identify the ubiquitin attachment sites on INF2. We co-expressed the FLAG-INF2, Myc-SPOP and HA-Ub constructs in 293T cells, and the immunoprecipitated ubiquitin-INF2 conjugates were analyzed by liquid chromatography tandem mass spectrometry (LC-MS/MS). It revealed that INF2 was ubiquitinated at least at 7 lysine residues (Fig 2I). Interestingly, 5 of 7 ubiquitin attachment sites are localized in a short stretch of sequence (amino acids 612–682) within the FH2 domain of INF2 (Fig 2J). To evaluate whether this region is important for INF2 ubiquitination, we constructed a series of INF2 deletion mutants and performed in vivo ubiquitination assay. While these deletion mutants were capable of binding to SPOP in a manner similar to the full length INF2 (Fig 2K), the N2 and ΔInter mutants, which lack the 612–682 aa region, were much less ubiquitinated by SPOP (Fig 2l). These data suggest that the lysine residues located in the 612–682 aa of INF2 serve as the predominant ubiquitin attachment sites.
Previous studies reported that one or several SBC motifs (Φ-π-S-S/T-S/T; Φ: nonpolar residues, π: polar residues) are present in known SPOP substrates [6–11,28]. We examined the protein sequence of INF2 that is required for SPOP-binding. To this end, we first deduced the minimal interacting region from the four INF2 fragments obtained in yeast two-hybrid screen. We found INF2 (1024~1249 aa) corresponds to the smallest region necessary for SPOP interaction (Fig 3A). Next we performed a protein motif search in the C-terminal region of INF2 and discovered a perfectly matched SBC motif (Fig 3A). Moreover, this motif is very similar to the SBC motifs present in MacroH2A, DAXX and DEK (Fig 3B). To examine whether this potential motif is actually required for SPOP-INF2 interaction, we generated an INF2 mutant in which the motif sequence was deleted. 293T cells were co-transfected with SPOP and wild-type INF2 or ΔSBC mutant. Co-IP assay demonstrated that SPOP only bound to the wild-type INF2, but not the ΔSBC mutant although they were expressed at comparable levels (Fig 3C), suggesting that the SBC motif of INF2 was required for SPOP binding. In vivo ubiquitination assay demonstrated that deletion of the SBC motif totally abolished SPOP-mediated INF2 ubiquitination (Fig 3D). Collectively, we have identified a conserved SBC motif present in INF2 that is indispensable for SPOP-INF2 interaction.
All the SPOP mutations detected thus far in prostate cancers exclusively occur in the MATH domain, which is responsible for substrate binding (Fig 4A). We postulated that prostate cancer-associated mutants of SPOP may be defective in mediating INF2 polyubiquitination. To test this, we generated a series of Myc-tagged prostate cancer-associated mutants of SPOP, including Y87C, Y87N, F102C, S119N, F125V, K129E, W131G, W131C, F133L, F133V and K134N, and examined their interactions with INF2 by co-IP assays. As shown in Fig 4B, mutations of the residues at the MATH domain abrogated the ability of SPOP to interact with INF2. Moreover, in vivo ubiquitination assay indicated that prostate cancer-associated SPOP mutants largely lost the capacity to promote INF2 polyubiquitination (Fig 4C).
Previous study showed that only one copy of SPOP allele is mutated in prostate cancer and SPOP mutants exert their tumor-promoting function in a dominant-negative manner to inhibit the wild-type SPOP [2]. We hypothesized that prostate cancer-associated mutations of SPOP might disrupt the interaction between wild-type SPOP and INF2. Indeed, we found that co-expression of SPOP mutants (Y87N, F125V or F133L) reduced the interaction between wild-type SPOP and INF2 (Fig 4D). Moreover, co-expression of SPOP mutants suppressed wild-type SPOP-induced INF2 ubiquitination (Fig 4E). Taken together, our findings suggest that INF2 ubiquitination may be dysregulated by oncogenic prostate cancer-associated SPOP mutants.
INF2 (CAAX isoform) is ER-anchored and INF2-mediated actin assembly is specially triggered at ER-mitochondrial intersections to ensure mitochondrial division [24]. Previous study showed that INF2-CAAX isoform was ER membrane-bound, but a pools of INF2 was cytosolic. [22]. SPOP was originally named as speckle-type POZ protein since ectopically expressed SPOP in COS-7 cells primarily exhibited a discrete speckled pattern in the nucleus [29]. Through quantitative analysis, we found that SPOP was localized exclusively in the nucleus as speckles in approximately 70% cells, but in both the cytoplasm and nucleus in the rest 30% cells, indicating that SPOP shuttles between cytoplasm and nucleus in a proportion of cells (S1 Fig). Thus, we hypothesized that SPOP-INF2 interaction occurs in the cytoplasm and SPOP-mediated atypical ubiquitination may regulate the subcellular localization of INF2. To test this hypothesis, we co-expressed GFP-tagged INF2 (CAAX isoform) and mApple-tagged Sec61β (an ER marker) in cells. We found that these two proteins were perfectly co-localized (Fig 5A), confirming that INF2 CAAX isoform is ER-localized. However, in approximately 30% cells that HA-SPOP was localized in both the cytoplasm and nucleus, INF2 was primarily present as speckles in the cytoplasm and co-localized with SPOP, but not Sec61β (Fig 5A). In contrast, in the rest 70% cells that HA-SPOP was localized exclusively in the nucleus, INF2 was still co-localized Sec61β (Fig 5A). These results suggest that SPOP can inhibit the ER localization of INF2, but this activity strictly depends its cytoplasmic localization. Moreover, deletion of the SBC motif (ΔSBC) or the region containing main ubiquitination sites (ΔInter) in INF2 did not alter its localization in ER (S2 Fig), but SPOP-induced speckle pattern of INF2 in the cytoplasm was not observed (S2 Fig), suggesting that SPOP-INF2 interaction and SPOP-induced INF2 ubiquitination are both required for INF2 localization outside of ER. Next, we investigated the impact of prostate cancer-associated mutants of SPOP on INF2 localization. To this end, we focused on three hotspot mutations Y87N, F125V and F133L. Interestingly, these mutants were exclusively localized as nuclear speckles in nearly 100% cells (S1 Fig), implying that cytoplasmic retention ability of SPOP may be impaired by prostate cancer-associated mutations. Accordingly, we found that ectopic expression of SPOP mutants had no obvious effect on the ER localization of INF2 by immunofluorescence analysis (Fig 5A). We used ER fractionation methods as a second method to corroborate the immunofluorescence analysis. As shown in Fig 5B, overexpression of wild-type SPOP, but not the prostate cancer-associated mutants of SPOP, reduced the protein amounts of GFP-INF2 in ER fractions.
Lastly, we investigated whether SPOP would affect the localization of endogenous INF2. As shown in Fig 5C, in a proportion of SPOP-WT-transfected cells, endogenous INF2 was present in cytoplasmic speckles and co-localized with SPOP, but this effect was not observed in cells expressing SPOP mutants. ER fractionation experiments demonstrated that stably overexpression of wild-type SPOP reduced the protein amounts of endogenous INF2 in ER fractions (Fig 5D). In contrast, overexpression of SPOP mutants moderately increased the protein amounts of endogenous INF2 in ER fractions (Fig 5D), probably those acting through a dominant-negative effect to inhibit endogenous SPOP.
Taken together, our data suggests that wild-type SPOP, but not prostate cancer-associated mutants, can promote INF2 disassociation from ER.
Considering that actin polymerization between mitochondria and INF2-enriched ER membranes is a critical step in mitochondrial fission [24], we reasoned that SPOP might suppress mitochondrial fission by inhibiting INF2 localization in ER. To test this, DU145 cells were infected with lentivirus expressing wild-type SPOP or prostate cancer-associated SPOP mutants. The mitochondrial morphology was monitored by Mitotracker Red dye. As shown in Fig 6A and 6B, stably overexpression of HA-SPOP resulted in significant increases in mitochondrial average length, accompanying with endogenous INF2 speckles in cytoplasm. However, this effect was only observed in approximately 30% cells that HA-SPOP was localized in both the cytoplasm and nucleus, but not in those cells that HA-SPOP was exclusively localized in nucleus (Fig 6A and 6B). These data suggest that SPOP-mediated suppression of mitochondrial fission strictly depends on its cytoplasmic localization. In contrast, the prostate cancer-associated SPOP mutants (SPOP-Y87N, F125V and F133L) lost the capacity to suppress mitochondrial fission monitored by immunofluorescence (Fig 6A). Statistical analysis showed stably overexpression SPOP mutants even resulted in moderate decreases in mitochondrial average length probably those acting through a dominant-negative effect to inhibit endogenous SPOP (Fig 6B). Previous study reported that the constitutive active mutant INF2-A149D can decreased mitochondrial length [24]. We found that the INF2-A149D-ΔSBC mutant, which can escape from SPOP-mediated ubiquitination, is more potent in decreasing mitochondria average length than INF2-A149D (Fig 6C). Consistent with these findings, depletion of SPOP in DU145 cells resulted in decreases in mitochondria average length (Fig 6D). Moreover, Co-depletion of DRP1 and SPOP by shRNAs reversed the effect of SPOP single depletion on mitochondria size (Fig 6D). Thus, SPOP inactivation-induced mitochondrial fission occurs upstream of DRP1.
Taken together, our data suggest that wild-type SPOP, but not prostate cancer-associated mutants, can suppress INF2-mediated mitochondrial fission.
Our above data indicated that SPOP regulates INF2-mediated mitochondrial fission strictly depends on its cytoplasmic localization, but the nuclear-cytoplasmic shuttling mechanism of SPOP was still poorly understood. It is clear that import of large proteins is generally mediated by nuclear localization signals (NLS), which contain basic amino acids [30]. SPOP contains an evolutionarily conserved NLS sequence at its C-terminus (S3A Fig). We found that SPOP lacking the NLS sequence (SPOP-ΔNLS) accumulated exclusively in the cytoplasm as puncta pattern and perfectly co-localized with GFP-INF2 (S3B Fig). In contrast, two prostate cancer-associated SPOP mutants lacking the NLS sequence (SPOP-F125V-ΔNLS, SPOP-F133L-ΔNLS) accumulated exclusively in the cytoplasm as puncta pattern similar as SPOP-ΔNLS, but these mutants did not co-localize with GFP-INF2, possibly due to impaired interaction with INF2 (S3B Fig). Moreover, SPOP-ΔNLS cannot alter the ER localization of INF2-ΔSBC and INF2-ΔInter mutants (S3C Fig), suggesting that SPOP-INF2 interaction and SPOP-induced INF2 ubiquitination are required for INF2 localization outside of ER. Proteins containing classic NLS are known to be transported into the nucleus by forming complexes with shuttling carriers, such as Karyopherin-alpha and-beta (KPNA and KPNB) [30]. Our yeast two-hybrid screen identified several clones corresponding to KPNA5 (importin subunit alpha-6). Indeed, deletion of the NLS sequence totally abolished the interaction between SPOP(WT, F125V, F133L) and overexpressed or endogenous KPNA5 (S3D Fig), suggesting that KPNA5 might participate in nuclear transport of wild-type and prostate cancer-associated SPOP mutants.
Not surprisingly, we found that SPOP-ΔNLS was able to immunoprecipitate more endogenous INF2 than SPOP-WT(S3E Fig), and SPOP-ΔNLS was more effective to promote INF2 ubiquitination than SPOP-WT(S3F Fig). It has been reported that INF2 promotes mitochondrial fission controls mitochondrial assembly of DRP1 [24]. DRP1 localized to cytoplasm and to mitochondrially associated puncta in cells. Depletion of INF2 reduced mitochondrially associated puncta, in addition to causing mitochondrial elongation[24]. We observed that SPOP-ΔNLS overexpression reduced DRP1 puncta associated with mitochondria and increased mitochondria length more efficient than SPOP-WT (Fig 7A, 7B and 7C). The levels of DRP1 in purified mitochondrial fractions from SPOP-ΔNLS overexpressing cells were also lower than those from SPOP-WT overexpressing cells (S3G Fig). In contrast, overexpression of SPOP-ΔBTB or ΔMATH mutant had no impact on DRP1 puncta, and mitochondria length (Fig 7A, 7B and 7C). It is not surprising since INF2 protein cannot be polyubiquitinated by SPOP-ΔBTB or ΔMATH.
Taken together, our data confirmed that cytoplasmic retention of SPOP is required for its regulation of mitochondrial fission.
To determine the biological importance of SPOP regulation of INF2-mediated mitochondrial fission, we first used two independent shRNAs (#1 targets total INF2, #2 targets INF2 CAAX isoform only) to knock down INF2 expression. Consistent with previous studies, INF2 depletion in LNCaP or DU145 cells resulted in a significant increase in mitochondrial average length (S4 Fig). However, this change was not associated with major change in mitochondrial function, as the basal mitochondrial reactive oxygen species (ROS) production (Fig 8A), oxygen consumption rate (OCR) (Fig 8B), and membrane potential (Fig 8C) were not significantly altered following INF2 depletion. Moreover, we found that INF2 depletion marginally affected the cell cycle progression (Fig 8D) or overall cell growth (Fig 8E). These results led us to explore other cancer cell phenotypes affected by INF2 depletion.
Recently, emerging evidence supports a role for mitochondrial dynamics in tumor cell migration and invasion in various cancer models [16–20]. Indeed, we found that depletion of INF2 in DU145 cells markedly decreased cell migration and invasion (Fig 8F and 8G). In contrast, depletion of SPOP enhanced cell migration and invasion (Fig 8F and 8G). More importantly, co-depletion of SPOP and INF2 reduced cell migration and invasion compared with depletion of SPOP only (Fig 8F and 8G). Similar results were obtained when we used SPOP-F133L mutant overexpression to replace knockdown of SPOP by shRNA (S5A and S5B Fig). Previous studies demonstrated that INF2 functions upstream of DRP1[24]. We found that treatment with DU145 cells with DRP1 selective inhibitor Mdivi-1 or knockdown of DRP1 significantly reduced SPOP depletion-enhanced cell migration and invasion (Fig 8H and 8I; S5C and S5D Fig). Similar effects were observed in another prostate cancer cells LNCaP (S6 Fig). Together, our data suggests that SPOP suppresses prostate cell migration and invasion, at least in part, by regulating INF2-mediated mitochondrial fission.
Although SPOP mutation is now recognized as a distinct molecular feature in a subtype of prostate cancer, the underlying mechanisms remain poorly understood [4]. Previous studies showed that SPOP inactivation increased cell proliferation primarily in AR-positive prostate cancer cells, but increased prostate cell migration and invasion in an AR-independent manner [2,9,10]. These effects were partly dependent on stabilization of SPOP substrates such as AR and ERG [9,10]. ERG up-regulation leads to transactivation of its target genes, including ADAMTS1, CXCR4, OPN and MMP9, all of which play important roles in promoting cell migration and invasion [9,10]. In this study, we revealed that the ER-localized isoform of the INF2 is ubiquitinated and regulated by SPOP. SPOP inactivation-induced prostate cancer cell migration and invasion is partly mediated by INF2 and mitochondrial fission (Fig 9). In the past few years, there is accumulating evidence that mitochondrial fission and fusion play active roles in regulation of cell movement, migration and invasion [14,31]. For example, there are higher levels of DRP1 and less Mfn1 (a GTPase for mitochondrial fusion) in the metastatic breast cancer cells compared with non-metastatic breast cancer cells.18 Silencing DRP1 or overexpression of Mfn1 results in mitochondrial fusion, and significantly suppresses migration and invasion abilities of breast cancer cells [18]. Similar effect has been detected in glioblastoma and lung or thyroid cancer[17,19,20]. The possible mechanism for mitochondrial fission-enhanced cell movement is that mitochondria are usually trafficked to sites of high-energy demand, and in migrating cells, mitochondria are more frequently located at their leading edge where demands high energy [14,32]. Our study, for the first time, links prostate cancer-associated SPOP mutations to mitochondrial dynamics-related cell migration and invasion. Interestingly, a recent study demonstrated that aberrant activation of MAPK signaling by K-Ras (G12V) mutation in pancreatic cancer activates DRP1 via ERK-mediated phosphorylation, and DRP1-meditated mitochondrial fission is crucial for Ras-driven transformation [33]. Similarly, BRAF (V600E), the most common mutation in melanoma, correlates with DRP1 phosphorylation in melanoma tumor tissues, whereas MAPK inhibition reverses DRP1-mediated mitochondrial fission, and sensitizes cells to mitochondrial-targeting drugs [34]. Therefore, cancer-associated mutations may promote mitochondrial fission through multiple signaling pathways in different tumors. It also should be noted that SPOP can ubiquitinate the cytoplasmic INF2 non-CAAX isoform similar as the ER-localized INF2 CAAX isoform (Fig 2F). A recent study revealed that a proportion of cytoplasmic INF2 was localized in focal adhesion (FA) and the protruding edge of migrating cells [35]. We cannot rule out the possibility that SPOP-mediated ubiquitination of INF2 non-CAAX isoform also affects cell migration and invasion.
Ubiquitination has critical functions in nearly all aspects of biological processes. Although ubiquitination is traditionally thought to only target proteins for degradation, recent studies suggest additional roles of ubiquitination in nonproteolytic functions involved in protein function regulation [36]. It is well known that K48-linked polyUb ubiquitin chains are sufficient to target substrates to the 26S proteasome for degradation and that K63-linked polyUb ubiquitin chains have been demonstrated to regulate a variety of nonproteolytic cellular functions, though the roles of other atypical ubiquitin linkages through M1, K6, K11, K27, K29 or K33 or mixed linkages within the same chain remain poorly understood [36]. Previous studies demonstrated that the mitochondrial ubiquitin ligase MITOL regulates mitochondrial-ER membrane bridges through K63-linked ubiquitination of mitochondrial Mfn2 (a GTPase for mitochondrial fission), suggesting that atypical ubiquitination plays roles in mitochondrial dynamics [37]. In this study, we demonstrated SPOP catalyzes synthesis of mixed-linkage polyUb chains (K27, K29, K33, K48 and K63) on INF2, which does not trigger INF2 degradation. Instead, these forms of ubiquitination cause INF2 dissociation from ER and impair its ability to promote mitochondrial fission. Until now, the majorities of known SPOP substrates are ubiquitinated and degraded by SPOP. But a previous study showed that SPOP is able to ubiquitinate the PcG protein BMI1 and the histone variant MacroH2A. These ubiquitinations do not affect the overall stability of BMI1 or MacroH2A, but facilitates PcG-mediated transcriptional repression and deposition of MacroH2A during stable X chromosome inactivation process [38]. These data and others reinforce a notion that SPOP can promote both degradative or non-degradative ubiquitination towards different substrates. Moreover, it is also possible that unknown deubiquitinase(s) might exist to recycle INF2 from cytoplasmic speckles to ER.
Another interesting aspect of our work that needs further investigation is the potential molecular mechanism that accounts for nuclear-cytoplasmic shuttling of SPOP. One study demonstrated that hypoxia condition promotes SPOP cytoplasmic accumulation in clear cell renal cell carcinoma (ccRCC) cells [39]. However, our preliminary results found that hypoxia treatment did not affect SPOP localization at least, in prostate cancer cells (S7 Fig). Considering that SPOP-mediated suppression of mitochondrial fission is strictly dependent on its cytoplasmic localization, elucidation the molecular mechanisms of cytoplasmic accumulation of SPOP is an important direction to pursue in the future. Moreover, three prostate cancer-associated SPOP mutants (Y87N, F125V and F133L) nearly lost their cytoplasmic localization compared with wild-type SPOP. It is possible that these mutations impair the capacity of SPOP to interact with proteins which facilitate cytoplasmic retention of SPOP. Our results showed that deletion of the NLS sequence forced prostate cancer-associated SPOP mutants to localize in cytosol as puncta, but these mutants cannot alter the ER localization of INF2 like SPOP-WT (Fig 7B). So it is possible that the direct interaction with some cytoplasmic substrates of SPOP, including but not limited to INF2, may cause a pool of SPOP to accumulate in cytosol by blocking access to Importin proteins. Prostate cancer-associated SPOP mutants lost the capacity to interact with its cytoplasmic binding partner, and localized exclusively in the nucleus. Taken together, our data suggest SPOP might exert its tumor-suppressive roles both in nucleus and cytoplasm.
293T, HeLa cells and prostate cancer cell lines (LNCaP, DU145, PC-3) were obtained from the American Type Culture Collection (ATCC). 293T and HeLa cells were maintained in DMEM with 10% (v/v) FBS. LNCaP and DU145 cells were maintained in DMEM with 10% (v/v) FBS. All cells were grown at 37°C with 5% CO2.
Expression vectors for SPOP-WT or mutants are described previously. FLAG-INF2-CAAX was obtained from Dr. Miguel Angel Alonso (Universidad Autónoma de Madrid). INF2 mutants were generated by KOD-Plus-Mutagenesis Kit (TOYOBO) following the manufacturer’s instructions.
For WB detection of ER-localized INF2 from HeLa cells, the microsomal fraction from approximately 5×106 HeLa cells was prepared using an ER extraction kit (ER0100, Sigma-Aldrich). The mitochondrial fraction was prepared using a Mitochondria Isolation Kit (MitoISO1, Sigma-Aldrich).
The pLKO.3G GFP-shRNA plasmids were purchased from Addgene. The shRNA sequence of sh-SPOP#1: 5’-GGAGAACGCUGCAGAAAUU-3’; sh-SPOP#2: 5’-ATAAGTCCAATAACGACAGGC-3’; shINF2-#1: 5’- CCCUCUGUGGUCAACUACU-3’; shINF2-#2 (target to CAAX isoform only): 5’-ACAAAGAAACTGTGTGTGTGA-3’;23 shDRP1: 5’-GCCAGCUAGAUAUUAACAACAAGAA-3’. shControl: 5’- ACAGACUUCGGAGUACCUG-3’. Viruses were collected from the medium 48 hr after transfection. For knockdown experiments, cells were infected with the collected viruses over 48 hr in the presence of polybrene, followed by GFP sorting for 3–4 days. pTsin- lentivirus vectors were used for overexpression of HA(FLAG)-SPOP-WT or mutants.
The following antibodies were used: SPOP (ab137537; Abcam), SPOP (16750-1-AP; proteintech), INF2 (20466-1-AP; proteintech), AR (SC-816; Santa Cruz), DEK (16448-1-AP, Proteintech), IQGAP1(ab133490; Abcam), DRP1(8570S; CST), KPNA5(A7731; Abcam), COX4 (Abcam; ab14744), Ubiquitin (6652–1; epitomics), Myc (9E10; Sigma), FLAG (M2; Sigma), HA (MM5-101R; Convance), Actin (AC-74; Sigma). Mdivi-1 was purchased from Selleckchem. MitoSOX Red dye was purchased from Invitrogen.
Ubiquitinated INF2 was prepared by transfecting FALG-INF2, HA-Ub and Myc-SPOP in 293T cells (5x100 mm dish). After 48 hr, the cells were lysed in RIPA buffer and the transfected INF2 was immunopurified from cell lysates with anti-Flag M2 agarose beads (Sigma) before being resolved by 7.5% SDS-PAGE. After Coomassie blue staining, the band corresponding to ubiquitinated INF2 was excised. The liquid chromatography tandem mass spectrometry analysis was carried out at the Proteomics Center of our institute.
For cell cycle analysis, cells were washed 48 h post-treatment with PBS and fixed in 70% ethanol overnight. The cells were washed again with PBS, stained with propidium iodide and analyzed by flow cytometry.
Cell proliferation rate was determined using Cell Counting Kit-8 (CCK-8) according to the manufacturer’s protocol (Dojindo Laboratories, Japan). Briefly, the cells were seeded onto 96-well plates at a density of 1,000 cells per well. During a 2 to 8-d culture periods, 10 μl of the CCK-8 solution was added to cell culture, and incubated for 2 hr. The resulting color was assayed at 450 nm using a microplate absorbance reader (Bio-Rad). Each assay was carried out in triplicate.
Cell migration and invasion were determined by Transwell (Costar) migration and invasion assays. LNCaP cells were precultured in serum-free medium for 48 hr. For migration assay, 3x104 cells were seeded in serum-free medium in the upper chamber, and the lower chamber was filled with RPMI1640 containing 5% FBS. After 48 h, the non-migrating cells on the upper chambers were carefully removed with a cotton swab, and migrated cells underside of the filter stained and counted in nine different fields. Matrigel invasion assays were performed using Transwell inserts (Costar) coated with Matrigel (BD Biosciences)/fibronectin ((BD Biosciences).
OCR was measured using a Seahorse XF24 Extracellular Flux Analyzer with the XF Cell Mito Stress Test Kit. Cells were seeded at 8 x 104 cells per well in 100μl DMEM containing 10% FBS and allowed to attach for 2 hr. 150μl DMEM-10% FBS was added per well and cells incubated overnight in 5% CO2 humidified incubator. Prior to assay run, cells were changed into assay media, unbuffered DMEM pH 7.4 and subjected to sequential injections of Oligomycin (1 μM), FCCP (0.3 μM), rotenone (1 μM) and antimycin A (0.75 μM). Spare respiratory capacity was calculated by dividing the OCR response to FCCP by the basal respiration, having subtracted the non-mitochondrial respiration previously. All values were normalized to cell number per wells setup in parallel.
Cells were seeded for 24hr and treated as indicated. TMRE (50 nM) or MitoSOX Red (5 μM) was added to the media, and the plates were incubated at 37°C in the dark for 30 min. Then cells were trypsinized and analyzed by flow cytometry.
For immunofluorescence, cells were plated on chamber slides, fixed with 4% paraformaldehyde at room temperature for 30 min. After washing with PBS, cells were permeabilized with 0.1% Triton X-100 in PBS for 15 min. Cells were then washed with PBS, blocked with 0.5% BSA in PBS for 1hr, and incubated with primary antibodies in PBS for at 4°C for overnight. After washing with PBS, fluorescence-labelled secondary antibodies were applied and DAPI was counterstained for 1hr at room temperature. Cells were visualized and imaged using a confocal microscope (LSM710, Zeiss). The analytic method of mitochondrial length was described previously.23
|
10.1371/journal.ppat.1002897 | Vector-Borne Transmission Imposes a Severe Bottleneck on an RNA Virus Population | RNA viruses typically occur in genetically diverse populations due to their error-prone genome replication. Genetic diversity is thought to be important in allowing RNA viruses to explore sequence space, facilitating adaptation to changing environments and hosts. Some arboviruses that infect both a mosquito vector and a mammalian host are known to experience population bottlenecks in their vectors, which may constrain their genetic diversity and could potentially lead to extinction events via Muller's ratchet. To examine this potential challenge of bottlenecks for arbovirus perpetuation, we studied Venezuelan equine encephalitis virus (VEEV) enzootic subtype IE and its natural vector Culex (Melanoconion) taeniopus, as an example of a virus-vector interaction with a long evolutionary history. Using a mixture of marked VEEV clones to infect C. taeniopus and real-time RT-PCR to track these clones during mosquito infection and dissemination, we observed severe bottleneck events that resulted in a significant drop in the number of clones present. At higher initial doses, the midgut was readily infected and there was a severe bottleneck at the midgut escape. Following a lower initial dose, the major bottleneck occurred at initial midgut infection. A second, less severe bottleneck was identified at the salivary gland infection stage following intrathoracic inoculation. Our results suggest that VEEV consistently encounters bottlenecks during infection, dissemination and transmission by its natural enzootic vector. The potential impacts of these bottlenecks on viral fitness and transmission, and the viral mechanisms that prevent genetic drift leading to extinction, deserve further study.
| The ability of arboviruses to perpetuate in nature given that they must infect two disparate hosts (the mosquito vector and the vertebrate host) remains a mystery. We studied how viral genetic diversity is impacted by the dual host transmission cycle. Our studies of an enzootic cycle using Venezuelan equine encephalitis virus (VEEV) and its natural mosquito, Culex taeniopus, revealed the stages of infection that result in a viral population bottleneck. Using a set of marked VEEV clones and repeated sampling at various time points following C. taeniopus infection, we determined the number of clones in various mosquito tissues culminating in transmission. Bottlenecks were identified but the stage of occurrence was dependent on the dose that initiated infection. Understanding the points at which mosquito-borne viruses are constrained will shed light on the ways in which virus diversity varies, leading to selection of mutants that may result in host range changes or alterations in virulence.
| RNA virus replication is characterized by a high frequency of mutation, which leads to genetically diverse populations. This diversity is thought to enable RNA viruses to effectively survive within the host (i.e. escape or evade immune responses), to be transmitted, and to potentially adapt to new hosts or vectors. While generating diversity may enhance viral survival, a slight rise above the natural mutation rate can be detrimental, and too little variation has been shown to decrease RNA viral spread and pathogenesis [1], [2]. Thus, RNA viruses must optimize their mutation rate so that enough mutations are generated to enable sufficient diversity for survival and adaptation, yet without producing too many deleterious mutations that can lead to error catastrophe and extinction.
The within-population diversity of RNA viruses is a by-product of their viral RNA-dependent RNA-polymerases (RdRp), as most viruses lack a proofreading domain in this enzyme. This low fidelity leads to a high error frequency for replication of all RNA viruses, which varies between 10−3 and 10−5 mis-incorporations per nucleotide copied. Genetic diversity acts as a critical determinant of viral evolution by facilitating positive selection (when a mutation confers a fitness advantage and thus produces more progeny), or by genetic drift (fixation of random mutations when populations are small). An extreme example of the latter is termed a bottleneck, which refers to a severe reduction in the population during infection, spread or transmission. Bottlenecks can lead to Muller's Ratchet; because reversion rates are low, asexual populations of organisms that periodically undergo population bottlenecks should tend to accumulate deleterious mutations, unless sex or recombination intervene to allow efficient restoration of the wild-type sequence [3], [4]. The deleterious effect of artificial bottlenecks (i.e. plaque-to-plaque passages) has been demonstrated for many viruses, including the alphavirus Eastern equine encephalitis virus (EEEV) [5], [6], [7], [8], [9], [10], [11], [12]. In addition, the limited oral susceptibility of many mosquito vectors to arboviruses (arthropod-borne viruses) may cause bottlenecks at the stage of initial midgut infection during natural transmission cycles.
Bottlenecks have been identified in many viral systems both in vitro and in vivo. Studies with Foot-and-mouth disease virus (FMDV) by Domingo et al. [13] have shown that repeated bottleneck events result in reduced viral fitness. Interestingly FMDV cannot compensate for this reduced fitness, emphasizing the long-term deleterious effects of bottlenecks on virus populations. Additional studies looking at more natural systems have identified bottlenecks when viruses spread between different tissues within a host [14], [15], [16], [17]. In particular much of what we understand about bottlenecks in natural systems comes from studies with plant viruses, which have identified bottlenecks during viral infection and cell-to-cell movement in various plant species [14], [17], [18]. Subsequent studies to identify bottlenecks during insect transmission identified a bottleneck during aphid transmission of cucumber mosaic virus [19], and a separate study quantified the amount of potato virus Y transmitted by the insect vector (∼0.5–3 viral particles), thus confirming the presence of a significant bottleneck during infection of and transmission by insect vectors [20].
Bottlenecks during the transmission cycle could have profound effects on arbovirus evolution, especially on adaptive changes. Experimental studies have demonstrated that the 2-host transmission cycle constrains the ability of another alphavirus, Venezuelan equine encephalitis virus (VEEV) to adapt to new laboratory hosts, presumably due to fitness tradeoffs for efficient infection of mosquitoes and vertebrates. Releasing VEEV from the 2-host cycle via serial passages in a single host facilitates adaptive evolution [21]. This finding has also been confirmed in other arboviruses [22], [23], [24], [25], [26]. Furthermore, bottlenecks during the natural transmission cycle could also limit adaptive evolution if they reduce population sizes to levels where selection cannot function efficiently. Previous work has defined three main infectious transitions or stages during mosquito infection: (1) midgut infection, when virions initially infect digestive epithelial cells, (2) midgut escape, when the virus must enter the hemocoel to infect secondary target organs and tissues, and (3) salivary gland infection, a requirement for oral transmission [27], [28]. Considering that midgut infection and escape often severely constrain the transmission process, they may represent bottlenecks and therefore limit genetic variation after oral exposure of mosquito vectors [21]. In addition, transmission of small amounts of virus in saliva may represent an additional bottleneck during arbovirus transmission [29].
Using virus-like replicon particles, Smith et al. [30] demonstrated that VEEV infects only a few cells in the midgut epithelium of the epizootic vector, Aedes (Ochlerotatus) taeniorhynchus. This natural bottleneck may reduce the number of virions that initiate infection within the mosquito, thus reducing genetic diversity in the virus population that eventually spreads to the salivary glands. The presence of a bottleneck in the mosquito vector has also been demonstrated for West Nile virus (WNV) in Culex quinquefasciatus [31], using similar methods. More recently studies with the mosquito vector C. pipiens have also demonstrated the presence of bottlenecks during WNV infection [32].
To further assess the presence and possible importance of bottlenecks on virus evolution and transmission, we used VEEV as a model arbovirus. VEEV emerges periodically to cause major epidemics and equine epizootics, a process that is mediated by adaptive mutations in the envelope glycoproteins that allow enhanced infection of epizootic mosquito vectors or equine amplification hosts [33], [34]. Thus, VEEV epitomizes the ability of RNA viruses to emerge and cause disease via adaptive mutations that lead to host range changes. The VEE antigenic complex of alphaviruses comprises 6 subtypes; of these, ID and IE and subtypes II–VI are enzootic strains that circulate continuously between Culex (Melanoconion) spp. mosquitoes and rodents, typically Sigmodon hispidus (cotton rats), Proechimys spp. (spiny rats) and Oligoryzmys spp. (rice rats) among others. For VEEV, the mosquito C. (Melanoconion) taeniopus has been implicated as the enzootic vector of subtype IE strains [35]. This mosquito is highly susceptible to oral infection, and even low levels of viremia in a rodent host can lead to mosquito infection (mosquitoes can be infected after ingesting <5 pfu) and subsequent transmission [36]. This high degree of susceptibility is believed to reflect a long-term evolutionary relationship between the vector and virus in its enzootic cycle.
We therefore hypothesized that the long association of this enzootic vector with VEEV subtype IE transmission has resulted in the high degree of vector infection efficiency. In contrast, a midgut infection bottleneck identified in the epizootic vector A. taeniorhynchus, which has only a transient role in VEEV transmission during epidemics, occurs after infection with VEEV subtype IC. In theory the long-term evolutionary relationship of VEEV and its enzootic vector might have limited the presence of bottlenecks during the enzootic cycle. To test this hypothesis, we conducted experimental infections using the enzootic vector to indirectly quantify the sizes of potential bottlenecks during vector infection. We used a mixture of genetically marked clones to follow the VEEV population from artificial bloodmeals through transmission to surrogate rodent hosts using a technique previously described [14], [37].
To estimate VEEV populations bottlenecks during infection of the enzootic vector, 10 individually marked VEEV clones were created within the backbone of the enzootic subtype IE strain 68U201 infectious clone. Each clone had 6 synonymous mutations in contiguous codons, except for 68U201-007, which had 5. All mutations were introduced into the nsP2 c-terminus that exhibits high sequence diversity and was therefore assumed to be tolerant of synonymous mutations, and were within 150 nt of each other. Each marked virus and the wild-type (wt; 68U201) were rescued and tested for fitness using several methods to ensure that the markers were relatively neutral: 1) standard replication curves performed in Vero cells and CD-1 mice showed statistically indistinguishable kinetics within one log10 of the wt strain 68U201 at all time points sampled (Fig. S1); 2) upon subcutaneous infection of mice, all 10 clones generated viremia titers of >5 log10 pfu/ml on day one post infection, indicating that all would be transmissible to mosquitoes (Fig. S2); All mice exhibited comparable weight loss to those infected with the wt (marked clone range: 31–42%, median: 37.5%; wt: 37%), and all died between days 5–8 after infection (data not shown); 3) Infection of a mouse with an equal mixture of all clones showed the presence of all 10 on days 2–5 days; 4) mosquitoes inoculated intrathoracically (IT) with each clone became infected as determined by cytopathic effect (CPE) assays 8 days post inoculation (data not shown), 5) mosquitoes injected IT with equal mixtures of all clones showed equal replication of each 8 days post infection (Fig. S4); and. 6) the survival of clones in mosquitoes following bottlenecks reflected a random process and no particular clones appeared more likely than others to disseminate to the hemocoel or salivary glands, as confirmed by statistical analysis (see below). In total, these data strongly indicate little or no difference in fitness among the marked clones, confirming the usefulness of the presumably neutral markers to assess stochastic viral population events following bottlenecks.
The 10 VEEV clones were validated with the probes and primer sets for real-time RT-PCR. All probes were able to detect the correct clone and did not cross-react with any of the other clones. The probes for clones 005 and 007 gave weaker signals, resulting in ca. 100-fold less sensitivity of these assays compared to the others. We therefore removed these 2 clones from the analysis. To ensure that the elimination of these clones from the analyses would not confound interpretation of our data due to the presence of unaccounted virus, we assayed by real-time RT-PCR a subset of mosquitoes and identified both 005 and 007 in some midguts and bodies that also contained most or all of the other clones, consistent with a lack of sampling bias when clones 005 and 007 were not assayed. Furthermore, to ensure that these clones did not infect or disseminate better than expected, 2 samples negative by real-time RT-PCR for these 2 clones were amplified by standard RT-PCR and deep sequenced. The lack of detectable clone 005 and 007 mutation peaks in the sequences indicated that they were not present. Moreover because these 2 clones were not observed in the brains of the mice infected during transmission experiments or in the saliva of mosquitoes for which those clones had not previously been identified by real-time RT-PCR, we were confident that the removal of these 2 clones does not impact the outcomes of mixed infections or our analyses.
Three cohorts of C. taeniopus were allowed to engorge immediately following tail vein, intravenous injection of mice with a 200 µl volume containing 6 log10 pfu/ml of each marked VEEV clone. Mice were bled before and after mosquito exposure to estimate oral doses. Titers before and after the feed were 6.7 log10 pfu/ml (±0) and 6.2 log10 pfu/ml (±0.18), or approximately 100-fold higher than would be expected in natural infections, but were designed to give 100% infection of the mosquitoes. Three mosquitoes were sampled daily including bodies as a measure of initial infection, legs/wings as a measure of disseminated infection, and saliva as a measure of transmission potential. RNA was extracted from CPE-positive samples and subjected to real-time RT-PCR to identify the presence of each clone (Table S2). The mean number of clones present in each sample was calculated and the results are presented in Fig. 1. We identified 6–8 clones (mean 7.6) in the bodies of the mosquitoes from days 1–3. There was a decrease in the number of clones present in the bodies at day 4 (mean 3.6) when most blood had been digested or excreted, but clone numbers remained relatively consistent during the remainder of mosquito infections. The legs and wings were consistently positive for viral genomes beginning on day 4 of extrinsic incubation, congruent with earlier time estimates for VEEV dissemination into the hemocoel of C. taeniopus [38]. The legs and wings contained between 1–4 clones (mean 1.6) during days 4–14. The presence of VEEV in a single legs/wings sample that was positive on day 1 could be due to a leaky midgut, a phenomenon previously identified during VEEV infections of C. taeniopus as well as infections of other mosquitoes by other alphaviruses [38], [39]. Clone content in saliva was also relatively consistently from day 4 (1–3 clones, mean 1.2), although the number of VEEV-positive saliva samples was inconsistent. Because the deposition of saliva into the FBS within capillary tubes could not be visually observed, a lack of salivation could be responsible for some CPE-negatives. While mosquitoes tend to deposit more virus into capillary tubes than into live hosts [29], the salivation technique is intermittent in its success and therefore can underestimate the number of mosquitoes with infectious saliva.
Over 14 days of infection, the number of clones present in mosquito bodies was significantly higher than that present in the legs and wings (p<0.01, by one-way ANOVA with Tukey-Kramer post-test) or in the saliva (p<0.01), indicating that escape from the midgut limited genetic diversity of VEEV populations. However, there was no significant difference between the number of clones in the saliva and the legs/wings (p>0.05) on days 1–14, indicating that the infection of the salivary glands did not detectably constrain VEEV diversity potentially transmitted by C. taeniopus.
To determine whether a bottleneck occurred during midgut infection, two additional cohorts of C. taeniopus were fed with the mixture of clones via an artificially viremic mouse, and midguts were dissected and assayed to assess viral diversity. Artificial viremia titers were 5.7 log10 pfu/ml (±0) for the first, high dose cohort and 4.9 log10 pfu/ml (±0.47) for the second, low dose cohort. Midguts and bodies were sampled on day 1, midguts, bodies, legs/wings and saliva on day 4, and bodies, legs/wings and saliva on days 8, 12 and 21, with 4 mosquitoes sampled each time point for the high dose cohort and 9 mosquitoes sampled each time point for the low dose cohort. The results of infection and RT-PCR assays are shown in Fig. 2. For the high dose cohort (Fig. 2A), the number of clones present in the midgut was equivalent to that seen in the mosquito carcasses during days 1–4 in the previous experiment (mean 7.6; range 7–8). The midguts were washed to remove residual bloodmeal, so the presence of the large number of clones was likely due to the presence of marked viruses either bound to the midgut epithelium or replicating in the midgut prior to escape into the hemocoel. The bodies, which were positive from day 4 onwards, contained a slightly higher number of clones (mean 2.8; range 2–5) compared to the legs and wings (mean 2.0, range 1–5), which were also positive by day 4. As before, the mean numbers of clones present in the legs and wings and the saliva showed no significant difference.
For the low dose cohort (Fig. 2B) there was a reduced mean number of clones in the midgut compared to the first two experiments (mean 3.7, range 1–6), suggesting that the lower titer of the bloodmeal limited the number of clones that infected initially. Surprisingly, the mosquito bodies from this low dose cohort were positive by day one rather than not before day 4, as in the high dose cohort. This outcome was unexpected because a higher oral dose is expected to lead to a faster midgut replication and escape into the hemocoel. However, there was still a significant reduction in the number of clones present outside the midgut in the mosquitoes sampled on days 4, 12 and 21: midguts vs. bodies (p = 0.004), bodies vs. legs/wings (p = 0.006) or saliva (p<0.001) and bodies vs. legs/wings (p = 0.03), respectively.
Mosquitoes from both cohorts were allowed to feed on naïve mice at day 21 of the extrinsic incubation period. For cohort one, 10 mice were each exposed to an individual mosquito, and each mosquito was allowed the opportunity to probe and/or engorge for one hour. Mosquitoes were processed immediately after exposure to mice, except that saliva could not be collected from the mosquitoes that engorged. Mice were monitored daily and bled on days one and 3, and the heart, brain, lungs and spleen were sampled on day 6 post infection if the animals showed signs of disease. Of the 10 animals presented to mosquitoes, only 3 showed signs of disease, and their CPE and RT-PCR results are shown in Fig. 3A. Preliminary results from mouse infections with an equal mixture of all 10 clones showed that the clones could be identified in the brain at 5–6 days post infection even if they were not identified in the serum due to sensitivity limitations of the assay (data not shown). Therefore, the number of clones in the brain was used as a surrogate to identify the clones that were transmitted to the mouse. The same 2 clones present in the brain of mouse one were present in the transmitting mosquito legs/wings (representing the hemocoel), indicating no detectible bottleneck during transmission. Similarly, mouse 2 contained only one clone in its brain, which was also the only one present in the legs/wings of the transmitting mosquito. In contrast, mouse 3 exhibited a major bottleneck during transmission; four clones were present in the legs/wings of the transmitting mosquito, yet only one of these was found in the corresponding mouse brain. This bottleneck could have occurred during salivary glands infection, deposition into the saliva, or transmission to the mouse. For the low dose cohort, 15 mice were presented to mosquitoes and again, only 3 developed signs of disease (Fig. 3B). Again there was no difference between the number of clones found in the legs/wings or saliva and the number of clones found in the mouse brain, suggesting no major bottleneck during transmission.
To investigate whether the severity of the bottlenecks at the midgut infection and escape levels masked a subsequent bottleneck during salivary gland infection, mosquitoes were intrathoracically (IT) injected with ∼1–2 µl of a 5 log10 pfu/ml VEEV suspension containing all clones to bypass midgut infection and sampled for the presence of virus as described above. Again, because of the limited sensitivity of the assays for clones 005 and 007, these were excluded from the analysis. Figure 4 shows that there was a significantly larger (p<0.001) mean number of clones present in the legs/wings compared to the saliva. These results suggest that, although a salivary gland infection bottleneck was not observable following oral infection, the IT infection that resulted in a greater clone diversity within the hemocoel allowed this bottleneck to be observed.
To determine the relative quantities of the VEEV clones present at each stage of mosquito infection, we produced standard curves for the real-time RT-PCR assays to estimate infectious titers as previously described [19]. A representative sample of the results is shown in Fig. 5 and 6, and all the results are found in Figs. S3–5. For the high dose cohort, the titers of all 8 clones present in midguts were similar as seen in Fig. 5A. However, the generally smaller numbers of clones in the bodies varied widely in titer, representing major and minor subpopulations. As all clones were present in roughly equal quantities in the midgut, and in the bodies in various ratios, this suggests that there was an equal probability of all 8 clones disseminating from the midgut. No particular clone appeared at high titer consistently in the bodies (Fig. S3), suggesting that little or no selection occurred and that VEEV escape from the midgut was a stochastic process. In all but one case, the clones present in the legs and wings were the same ones present in the highest quantity in the body (see Fig. 5B). Similarly the clones present in the saliva were in general the same ones present at the highest titer in the legs/wings of the mosquito (see Fig. S3). Transmission showed a similar pattern, with the clone/s present in the highest quantities in the legs and wings appearing in the mice. Interestingly, this pattern was seen during progression through all mosquito organs except the midgut (Fig. 5C). After IT inoculation, all 8 clones were present at roughly equal titers within the legs and wings, and the number of clones present within the saliva did not correlate with the relative titers of clones in the legs and wings.
For the low dose mosquito cohort, the clones detected in the midgut were not present in equal quantities (Fig. 6A), suggesting more stochastic variation during midgut infection. However, there was still a reduction in the number of clones disseminating in the body. As observed previously for the transition into the hemocoel (legs/wings), the clones present in the largest quantities invariably disseminated. This was recapitulated in the transition to the legs/wings on day 8 (Fig. 6B), where the 2 clones present in the largest quantities in the body were detected in the leg/wings of the same mosquito. For the low dose transmission experiment, for the first time a clone present in the mouse serum was not seen in the brain. We assume that this animal exhibited a bottleneck such that only 2 out of 3 clones entered the brain. Interestingly, the mosquito transmission of the clones was less consistent than observed in the high dose cohort, suggesting that a greater stochastic element in the initial midgut infection extends to more random dissemination of the clones at various points during the transmission cycle.
Using FST statistics, we estimated the number of viral particles initiating infection (N) of the midgut and escaping to initiate infection of the hemocoel. For the high dose cohort, there was no significant bottleneck during infection of the midgut with N estimated to be 1218 (±1318) infectious viral particles. However, for dissemination into the hemocoel, sampled from the legs and wings, the bottleneck N was estimated at 50.9 (±154). In comparison, we observed a strong initial midgut infection bottleneck for the low dose cohort, as the average number of infecting virions was estimated to be 1.9 (±2.6), suggesting that very few infectious virions initiated midgut infection. The average N for dissemination after low dose infection was 1.0 (±1.7), suggesting another strong bottleneck.
Neutrality of the clones for infection of the midgut was determined using the ChiSquare test. The number of clones infecting the midguts was compared to the expected number under the assumption that the clonal markers were neutral. For infection of the midgut at a high dose there was no difference between the observed versus expected (χ2 = 0.88, DF = 7, P = 0.997). However, for the low dose cohort there was a greater but still insignificant difference between the observed and expected (χ2 = 12.1, DF = 7, P = 0.099). This difference for the low dose cohort is likely due to the presence of the bottleneck at the entry into the midgut and therefore the small number of clones analyzed and the resulting high variance. In addition, the presence of all marked clones in roughly equal quantities (mean = 3.42 ±0.74 log10pfu/ml according to the qRT-PCR results) again supports neutrality of the markers.
It has been previously postulated that infection of and transmission by mosquito vectors may present bottlenecks for arbovirus populations. To evaluate the potential effects of such bottlenecks during the enzootic transmission cycle of an arbovirus, we used C. taeniopus and VEEV subtype IE. The highly efficient infection of this vector by this VEEV strain is believed to reflect a long-term virus-vector evolutionary association. Using a set of marked virus clones, we assessed how often and to what degree bottlenecks affected the number of infectious viral particles transitioning through the mosquito and transmitted to a vertebrate.
Following oral exposure, we identified a major and significant bottleneck when VEEV escapes from the midgut into the hemocoel. Using artificial, intrathoracic infection, we identified a second minor bottleneck at the entry into the salivary glands. This bottleneck was not identified during oral infection, probably because the number of clones in the hemocoel was reduced so severely that a slight drop in the number of clones in the saliva was not apparent. We also examined bottlenecks after oral infection of C. taeniopus with 2 different doses differing by about 10-fold, the lower of which probably more closely represents a natural infection. At this lower titer, all 8 clones did not generally infect the midgut. Also, given the lower titer of the bloodmeal it is possible that the blood the mosquito ingested did not contain all the clones as the titer ingested would have been 1.9–2.33 log10 pfu/mosquito. Using FST statistics we estimated the size of the founder populations for both the high and low dose cohorts. During the high dose there was no major bottleneck upon midgut infection, and all the clones were present in nearly equal quantities. In contrast to the low dose infection, the most severe bottleneck after a high dose oral infection occurred during escape from the midgut into the hemocoel. However, when mosquitoes were infected with a low oral dose, there was a major bottleneck during initial midgut infection, with only approximately 2 infectious virions initiating infection. Following low dose oral infection, there was also a small bottleneck upon dissemination into the hemocoel, with only one clone on average sampled in the legs/wings. Interestingly, given that some clones were present in tissues that had not been positive in assays of the different tissues from the same mosquito, it is likely that some of these clones were present, but at such low quantities that they were not detectable by our methods. Thus, the number of infectious particles initiating the midgut infection following the low dose may be underestimated by our methods.
Experimental infections with rodents collected from VEE-endemic areas in Chiapas, Mexico have indicated peak viremia titers of ca 3–4 log10 pfu/ml for VEEV subtype IE. Our results with comparable oral mosquito doses suggest that a major bottleneck during natural mosquito infection occurs at the stage of the initial infection of the midgut. A further bottleneck occurs during dissemination from the midgut into the hemocoel, which leads to infection of the salivary glands.
When mosquitoes transmitted VEEV to mice, no consistent bottleneck could be identified. However, the possibility of a transmission bottleneck cannot be ruled out, especially given the small number of clones that reached the salivary glands after dissemination into the hemocoel and the small number of samples we tested. Further experiments to evaluate potential bottlenecks in the vertebrate host, as suggested by the inconsistency between clone populations in the brain vs. serum of mice, will be required. However, for this study we focused primarily on the number of clones present in the mouse as an indication of the viral population size transmitted from the mosquito. Pfeiffer and Kierkegaard [37] observed a bottleneck when poliovirus crossed the blood-brain barrier. A comparable bottleneck may occur when VEEV enters the brain. However, given that for arboviruses transmission occurs via bloodmeals, bottlenecks at the blood brain barrier are less important than those that affect viremia for continued transmission.
Using a real-time RT-PCR assay and standard curves, we estimated the titers of the individual clones present in mosquito samples for the oral transmission experiments. Clones infecting the midgut were present in similar quantities in the high dose cohort, confirming similar fitness levels for replication in this organ, but not for the low dose cohort. However, after escape from the midgut, the clones were generally present in differing quantities, with no consistency in the relative amounts, indicating that the synonymous genetic markers were indeed neutral; this finding was consistent for both cohorts. With only a few exceptions, the clones that were present at highest titers in the hemocoel transitioned to the salivary glands. This suggests that the clone present at the highest titer in the hemocoel has the best chance of becoming the dominant population in the salivary glands, as expected based on the stochastic nature of genetic drift following bottlenecks.
Genetic diversity is thought to be critical for the survival of RNA viruses, and the presence of a mutant swarm, or intra-host variation, allows them to explore a wider variety of sequence space and therefore adapt to new hosts and new selective pressures [13], [40]. Previous studies with viruses artificially engineered to replicate with a higher-fidelity polymerase demonstrated reduced spread, pathogenesis and fitness in a host when compared with the wt [1], [37], [41]; thus a high level of diversity is imperative for viruses to maintain fitness and robust infections. Given our results suggesting that the mosquito vector is a source of major bottlenecks during the VEEV transmission cycle, the genetic stability of this and other arboviruses and their persistence in their ecological niches is remarkable. Random sampling of viral RNA genomes during population bottlenecks may shift the viral sequence away from the original, fit, or master sequence, thus creating a founder effect. The severe bottleneck at the level of VEEV infection of, or escape from, the midgut could thus have serious consequences for the virus due to the resultant loss in viral genome diversity. Further studies will be needed to determine if and how VEEV is able to restore adequate levels of diversity following these bottlenecks.
Previous experiments with an epizootic subtype IC strain of VEEV and its vector, A. taeniorhynchus [30], demonstrated that the midguts of these mosquitoes have only a few cells that are susceptible to initial infection, and thus the entry into the midgut represents a severe bottleneck at infection. Our results with an enzootic VEEV strain and vector underscore potential differences between the evolution of enzootic and epizootic VEEV strains. The presence of all 8 clones at equal quantities in midguts of the enzootic vector, C. taeniopus at the high dose, suggests that a larger proportion of its midgut cells are susceptible to infection. Recent studies using VEEV replicon particles also indicate that most if not all C. taeniopus midgut cells are susceptible to infection [42]. Since these prior studies used replicon particles to determine the number of cells initially infected, it was not possible to determine whether a midgut escape barrier occurred. Interestingly, for the low dose cohort, the large proportion of susceptible midgut cells did not appear to overcome the bottleneck, as only a very limited number of clones still infected the midgut. Thus, the potential bottleneck and the severity and timing of the initial bottleneck appear to be dependent principally on the titer of the bloodmeal ingested by the mosquito. This has important implications for further experiments to determine the effects on arbovirus evolution of bottlenecks in this and other experimental systems.
A potential conundrum arising from our results and others is that repeated bottlenecks should result in quasispecies constrictions and fitness declines due to Mullers Ratchet [43], as has been demonstrated for viruses and in particular an arbovirus (EEEV) [12]. However, evidence from other alphavirus studies as well as work performed with flaviviruses such as West Nile virus, indicates that arbovirus diversity is maintained even throughout a mosquito infection [44], and most arboviruses are highly stable both genetically and phenotypically in nature [45] and during laboratory passages [46]. Similar work with VEEV is underway in our laboratory. If arboviral diversity is maintained, 3 hypotheses are suggested: 1) the presence of a bottleneck may be compensated by further viral replication and recovery of diversity within the mosquito; 2) bottlenecks are less severe than our VEEV data suggest and there is sufficient diversity retained for maintenance of population fitness, or; 3) many mosquito-rodent VEEV lineages do decline in fitness due to bottlenecks and become extinct, but the large number of such lineages in enzootic or epidemic habitats ensures that some fit lineages remain.
Levels of genetic diversity within natural alphavirus populations have received very little attention, but eastern equine encephalitis virus is known to maintain diversity comparable to other RNA virus populations during natural infections of birds and mosquitoes [47], supporting hypothesis one and possibly 2. Titers of VEEV present in C. taeniopus suggest opportunities for the restoration of diversity following a bottleneck event such as midgut escape. The total VEEV titer in the mosquito legs and wings is approximately 6 log10 pfu. This suggests that, if only a few virions are responsible for initiation of the hemocoel infection as our data imply, VEEV genetic diversity would be partially restored through the large number of replication cycles, albeit to lower fitness levels if direct reversion is inefficient in restoring fitness as exemplified by Mullers ratchet [43], i.e. restoration of sequence diversity may not necessarily be accompanied by restoration of the high fitness master sequence. The genomic sequencing of different alphavirus isolates collected from the same time and place generally reveal minor differences in consensus sequences, consistent with bottleneck-mediated drift among different transmission lineages and possibly supporting hypothesis 3. Thus the complex interplay of genetic diversity and selection are not really understood, nor has this rate been examined using deep sequencing, which would show the number of mutations generated in the absence of selection. Thus it is unclear to what extent viral diversity could be restored and further experiments using next-generation sequencing will need to be performed to determine the effect of bottlenecks on viral diversity.
In summary, there must be a complicated interplay between the various evolutionary processes to give rise to the viral populations observed in nature. Even within the mosquito, different arbovirus mutations may confer different advantages in critical organs such as the midgut and the salivary glands. In addition, the influence of vertebrate hosts on the maintenance of arbovirus diversity has yet to be determined. Further research to improve understanding of arbovirus evolution will increase insights into the processes that can lead to the emergence of new variants with devastating impacts on human health.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee of the University of Texas Medical Branch.
Vero (African green monkey kidney) and baby hamster kidney (BHK) cells were obtained from the American Type Culture Collection (Bethesda, MD) and maintained in Dulbecco's minimal essential medium (DMEM) (Gibco, Carlsbad, CA) supplemented with 5% fetal bovine serum (FBS), penicillin and streptomycin (100 U/ml). Viruses were rescued from an infectious cDNA clone derived from enzootic, subtype IE VEEV strain 68U201 as described previously and without further passage [30]. The parental 68U201 virus (genomic sequence in GenBank accession no. U34999) was isolated from a sentinel hamster in a sylvatic Guatemalan focus of VEEV in 1968 and passaged once in infant mice and twice in BHK cells prior to cloning [48].
All mosquito and vertebrate tissues were resuspended in DMEM supplemented with 10% FBS Penicillin/Streptomycin (for mosquito tissues fungizone (Sigma-Aldrich)) was also added) and homogenized at 26 hz for 5 minutes, then subjected to centrifugation at 3820× g for 10 minutes. Saliva samples were subjected to centrifugation at 663× g for 10 minutes prior to processing. All samples were tested for the presence of virus by a cytopathic effect assay (CPE). Positive samples were stored at −80°C for subsequent analysis. For saliva samples, supernatants positive for CPE were used in a real-time RT-PCR assay, as the inconsistency in the amount of virus expectorated from the mosquito [29] would have resulted in some samples being below the limit of detection and thus the passaged supernatant was utilized.
Virus suspensions were placed into either TRIZOL (Invitrogen, Carlsbad, CA) in a 1∶4 dilution in order to extract total RNA using the manufacturers protocol, or into Buffer AVL (Qiagen, Valencia, CA) and RNA extracted using the column method as per the manufacturers protocol. Real time RT-PCR was carried out using the ABI 7900HT Fast Real-Time PCR system (ABI, Carlsbad, CA). Each reaction was performed using the TaqMan RNA-to-CT 1-Step kit (ABI) as per the manufacturers instructions in a 10 ul reaction. A list of the primers and the corresponding probes can be found in Table S2. Each probe had a corresponding primer set that was designed to anneal flanking the polymorphic region of each variant. Each sample was tested for each variant individually and each well was run in duplicate. Positive and negative controls were run on each plate and all 10 clones were included as controls to ensure no cross-detection of the other clones by an individual probe. Additionally, we used serial dilutions with titers from 106−101 pfu/ml of the individual clones to create standard curves (data not shown).
Paired T-Tests were used to determine the change in the number of clones from Bodies vs Legs and wings, Bodies vs Saliva and Legs and wings vs Saliva for individual days. Comparison of the differences over the entire experiment was determined by a one-way ANOVA followed by a Tukey-Kramer post-hoc test. The neutrality of the markers was estimated from the data using the ChiSquare contingency table. The size of the bottlenecks was estimated using the FST statistic as described in Monsion et al (2008) [51]. Briefly, Fst was estimated using the equation(1)Where HT is the average proportion of clones throughout the entire experiment and Hs is the clone proportion within a population. For our experiments each tissue within a mosquito was counted as a population. Using the FST statistic we were able to estimate the number of clones as a founder population in the mosquito using a second equation,(2)where F'ST is the initial population and FST the second population, for this experiment the midgut or body and the corresponding legs/wings from the same mosquito respectively. The average N was calculated plus standard deviations.
|
10.1371/journal.pntd.0003055 | Emergence of Coxiella burnetii in Ruminants on Reunion Island? Prevalence and Risk Factors | Q fever is a widespread zoonosis that is caused by Coxiella burnetii (C. burnetii), and ruminants are identified as the main sources of human infections. Some human cases have been described, but very limited information was available about Q fever in ruminants on Reunion Island, a tropical island in the Indian Ocean. A cross-sectional study was undertaken from March 2011 to August 2012 to assess the Q fever prevalence and to identify the major risk factors of C. burnetii infection in ruminants. A total of 516 ruminants (245 cattle, 137 sheep and 134 goats) belonging to 71 farms and localized in different ecosystems of the island were randomly selected. Samples of blood, vaginal mucus and milk were concomitantly collected from females, and a questionnaire was submitted to the farmers. Ticks from positively detected farms were also collected. The overall seropositivity was 11.8% in cattle, 1.4% in sheep and 13.4% in goats. C. burnetii DNA was detected by PCR in 0.81%, 4.4% and 20.1% in cow, sheep and goat vaginal swabs, respectively. C. burnetii shedding in milk was observed in 1% of cows, 0% in sheep and 4.7% in goats. None of the ticks were detected to be positive for C. burnetii. C. burnetii infection increased when the farm was exposed to prevailing winds and when there were no specific precautions for a visitor before entering the farm, and they decreased when a proper quarantine was set up for any introduction of a new ruminant and when the animals returned to the farm at night. MLVA genotyping confirmed the role of these risk factors in infection.
| Q fever is a disease that could be transmitted from animals (cattle, sheep and goats) to humans and caused by a bacterium called Coxiella burnetii (C. burnetii). Some human cases exhibiting characteristic clinical signs of that disease have been detected on Reunion Island, a tropical island in the Indian Ocean, but to date, we did not know if these animals could be seen as potential sources of the disease. Thus, a study was undertaken from March 2011 to August 2012 to detect the presence of that bacterium in these animals and to understand how they could get infected themselves. A total of 516 ruminants (245 cattle, 137 sheep and 134 goats) belonging to 71 farms and localized in different environments of the island were selected. Samples of blood, vaginal mucus and milk were concomitantly collected from females, and a questionnaire was submitted to the farmers. Ticks from positively detected farms were also collected. We observed 11.8% of cattle, 1.4% of sheep and 13.4% of goats had already been in contact with the bacterium. Coxiella burnetii was also directly detected in some vaginal and milk samples. None of the ticks were detected to be positive for C. burnetii. We found that the ruminants could be infected when their farm was exposed to prevailing winds because the bacterium can be transported by the wind, and when there were no specific precautions for visitors before entering the farm, because they could act as mechanical carriers of Coxiella. Conversely, keeping new animals under surveillance for some days to detect any signs of the disease before they enter the farm or keeping the animals in the barn at night limit the risk of infection.
| Q fever is a widespread zoonosis that is caused by Coxiella burnetii (C. burnetii), an obligate intracellular bacterium [1]–[4]. The reservoir includes mammals, birds and arthropods, mainly ticks [5]. Ruminants (sheep, goats and cattle) are identified as the main sources of human infections [6], [7]. Humans are infected mainly by inhalation of an aerosol contaminated with parturient products from the urines or feces of infected animals [8].
The risk of transmission of C. burnetii is dependent on the prevalence of shedder ruminants and on the level of shedding. C. burnetii is shed by ruminants mainly by birth products, but it may be shed via the vaginal mucus, milk, feces, urine and semen [9]. To control the spread of C. burnetii among animals as well as from animals to humans, the detection of shedders of C. burnetii and the knowledge of the prevalence of the infection are imperative. The risk of zoonosis also depends on the level of C. burnetii in the products of the infected animals.
Serological tests (complement fixation, indirect immunofluorescence and enzyme-linked immunosorbent assays (ELISA)) are classically used in epidemiological studies to detect carriers of antibodies against C. burnetii. Serological tests indicate previous exposure [10] to C. burnetii and are not appropriate for the identification of shedder ruminants, especially because seronegatives are present among them [11], [12]. This lack of sensibility in this technique is lower using ELISA [13].
Isolation of C. burnetii is not performed for epidemiological investigation because it is difficult, time consuming and requires confined level L3 laboratories. Conventional polymerase chain reaction presents a very useful method for the detection of C. burnetii DNA [9], [14]. The real-time PCR assays are now recognized as the most convenient tools because these tests have excellent sensitivity, specificity and permits investigators to obtain quantifiable information. Real-time PCR is adapted to large scale studies because this technique can be semi-automated, thus reducing the risk of sample contamination and permitting gained time.
Reunion Island is a French overseas department that has a population of approximately 800,000 inhabitants. Reunion Island is a hotspot in the Earth's crust located in the Indian Ocean, east of Madagascar, approximately 200 km south-west of Mauritius, the nearest island. The island is 63 km long and 45 km wide and covers an area of 2,512 km2. Cities are concentrated on the surrounding coastal lowlands. The climate is tropical and humid, with two main seasons: a hot rainy season from December to March, and a dry and cold season from April to November. The eastern coast (the “windward” coast) experiences rainfall of approximately 2,000 mm per year, whereas the western coast (the “leeward” coast) has an annual rainfall of less than 2,000 mm. The domesticated animal populations on the island comprise approximately 40,000 cattle, 30,000 goats and 2,000 sheep. To date, no information was available about Q fever in humans and animals.
The present study aimed to provide epidemiological information about Q fever in the animal population of Reunion Island using available diagnostic tools and appropriate samples. The data will be used to appreciate the prevalence of C. burnetii infection in the three main domestic ruminant species: cattle, sheep and goats at both the animal and herd levels, as well as to identify the major risk factors of infection.
The research protocol was implemented with the approval of the Direction of Agriculture, Food and Forestry (DAAF) from the French Ministry of Agriculture, under the European animal welfare regulation (project license number 102498). No endangered or protected species were involved in the survey. All the farmers gave their permission to be included in the study and for the samples. The animals were sampled without suffering.
The animals were considered positive when at least one sample (blood, swab or milk sample) tested positive by either serology or PCR. The serological and PCR data were analyzed using a generalized linear mixed model (glmmML library, R software), where the individual health status was the binomial response, and the variables from the questionnaire were the explicative factors.
All of the explicative variables were categorical. The number of categories per variable was limited, such that frequencies of categories were only >10%. These variables were selected from a preliminary step aimed at lowering the chance of obtaining results affected by multicollinearity in the dataset [18]. All bilateral relationships between these variables were evaluated (χ2). A two-stage procedure was used to assess the relationship between explanatory variables and the health status of the animals. Logistic regression was used according to the method described by Hosmer and Lemeshow [19]. In the first stage, a univariate analysis was performed to relate Q fever positivity to each explanatory variable. Only factors associated (Pearson χ2-test, P<0.25) with Q fever positivity were offered to a full model for multivariable analysis [20]. The second stage involved a logistic multiple-regression model. The contribution of each factor to the model was tested with a likelihood-ratio χ2 through a stepwise procedure (backward and forward). At the same time, the simpler models were compared to the full model by the Akaike information criterion [21]. This process was continued automatically until a model was obtained with all factors significant at P<0.05 (two-sided). Goodness-of-fit of the final model was assessed using Pearson χ2, Deviance and the Hosmer–Lemeshow tests [19].
The evaluation of the specificity of the real time PCR assays was reported by Klee et al. (2006) [22]. In the present study, all negative samples from INRA were negative, confirming the specificity of the assays.
All of the samples that were found positive by INRA were confirmed positive. The detection threshold determined from dilution series of synthetic DNA showed that this PCR allowed the detection of 24 samples of 500 copies of the genome/mL and one sample of 250 C. burnetii particles/ml sample, the number of IS1111 elements in the genome being determined to be close to 20 for the Nine Mile strains [22].
Coefficients and Ct averages of intra- and inter-assays were 0.46% and 1.4%, 26.54 and 26.63, respectively.
Bacterial load in vaginal samples by ml of transport medium ranged from 50,600 to 255,000 for cattle, from 82,400 to 314,000 for sheep and from 112,000 to 385,000 for goats.
The typability of the two loci was 85.1% (40 of 47 positive PCR for goats). We obtained nine genotypes among the 40 amplified DNA samples (Table 1).
The overall seropositivity was 11.8% (95% CI 7.8 – 15.9) in cattle, 1.4% (95% CI 0 – 3.5) in sheep and 13.4% (95% CI 8.2–25.6) in goats. C. burnetii DNA was detected by PCR in 0.81% (95% CI 0–1.9) of cow vaginal swabs, 4.4% (95% CI 0.9 – 7.8) of ewe vaginal swabs and 20.1% (95% CI 13.3 – 26.9) of goat vaginal swabs. C. burnetii shedding in milk was observed in 1% (95% CI 0.2 –1.8) of cows, 0% in sheep and 4.7% (95% CI 0 – 11.2) in goats. Twenty-one out of 46 (95% CI 32 – 60) cattle farms were found to be positive either in serology or PCR, 50% (95% CI 33 – 67) of sheep farms and 41% (95% CI 18 – 64) of goat farms. All of these farms were spread throughout the island (figure 1). The within-herd prevalence in the positive farms ranged from 20% to 40% in cattle farms and from 30% to 90% in small ruminant farms. None of the ticks collected were detected to be positive for C. burnetii.
After variable selection (Table 2), the logistic multiple-regression model indicated that the risk of C. burnetii infection was increased when the farm was exposed to prevailing winds (OR = 2,11; 95% CI [1,13; 3,99]) and when there were no specific precautions for a visitor before entering the farm (OR = 3,13; 95%CI [1,57; 6,70]), and decreased when a proper quarantine was set up for any introduction of new ruminant (OR = 0,06; 95%CI [0,01; 0,17]) and when the animals went back to the farm at night (0,53; 95%CI [0,42; 0,64]) (Table 3).
To the best of our knowledge, this is the only documented epidemiological study on Q fever in ruminants in Reunion Island, highlighting that Coxiella burnetii is endemic in this territory.
For our epidemiological survey, we used both serological and PCR techniques to better understand the characteristics of Q fever. Complement fixation technique remains widely used by laboratories in many countries to assess the seroprevalence of C. burnetii infection. This method yields good results for routine diagnosis at the herd level, but multiple studies have concluded (World Organisation for Animal Health 2010) that CFT is less sensitive than ELISA testing. Following international suggestions, ELISA results are deemed reliable for the screening of seroprevalence [23], [24]. However, serological tests (complement fixation or ELISA) only detect antibody-carriers against C. burnetii, demonstrating the previous exposure to the pathogen but not the current shedding of the pathogen [10]. Because we aimed to assess the overall pattern and characteristics of Q fever in Reunion island, the detection of shedders of C. burnetii was important because they are one of the critical points for the control of spreading of the bacteria among animals and from animals to humans [3]. Polymerase chain reaction (PCR) has been used to detect C. burnetii DNA in biological samples. Additionally, we employed a real-time PCR technique that is currently being developed with the aim of providing quantifiable information. The technique allows a priori scaling in the importance of sources of bacterium with regards to the risk of transmission of C. burnetii among animals and from animals to humans. Finally, on the contrary to conventional PCR, real-time PCR can be automated, leading to both a lower risk of sample contamination and a more time-efficient method of detection [25]. Because we found a bacterial load 40 to 310 times higher than the detection threshold, the probability for false negatives remained low. Finally, we observed nine different MLVA genotypes with very good typability compared to that obtained from Roest et al. (53%) [26], possibly due to the number of cycles (60).
In our study, the overall seropositivity was 11.8% in cattle, 1.4% in sheep and 13.4% in goats. These results are much lower than those observed in Europe; for example, ELISA testing showed 38.0% in cattle and 6.0% in sheep for individual seropositivity in Hungary [27]. In Northern Spain, ELISA anti-C. burnetii antibody prevalence was slightly higher in sheep (11.8±2.0%) than in goats (8.7±5.9%) and beef cattle (6.7±2.0%) [28]. Our seroprevalence rates were also lower compared to the results from other tropical countries. The seroprevalence in cattle was estimated to be between 40% and 59.8% in Nigeria, Sudan and Zimbabwe, and only 4% Chad [29]. The seroprevalence in sheep has been reported to vary between countries: 62.5% in Sudan, 22.5% in Egypt and 11% in Chad. Additionally, differences were also observed for seroprevalence in goats: 53% in Sudan, 16.3% in Egypt and 10% in Zimbabwe [29]. Our PCR results were quite surprising, with a low prevalence of C. burnetii in cow and ewe vaginal swabs (0.81% and 4.4%, respectively), but very high prevalence of 21% among goats. Generally, such high rates are observed after a Q fever-related termination of pregnancy as described by Cantas et al. (2011) [30] and Berri et al. (2005) [7]. Indeed, in our study, six small ruminant farms have indicated terminations among pregnant ruminants and our samples were collected within one month after these events. Shedding of C. burnetii in vaginal mucus lasts for one to five weeks [31]. In addition, it has been shown that most of the goats that had aborted or delivered normally in naturally infected herds shed the bacteria [32], [33]. Our findings confirmed these previous results because the within-herd bacterial prevalence in the farms that reported pregnancy terminations was estimated to be between 70% and 90%.
This study demonstrates that the risk of Q fever infection of ruminants increased when farms or grazing pastures are in the way of prevailing winds, confirming the airborne route of transmission for C. burnetii. In contrast to other studies [34], ticks, which were all detected to be negative for C. burnetii, appeared to not be involved in the contamination process. However, the systematic use of deltamethrin may have reduced the tick population and altered their ability to carry C. burnetii. Infection of animals or humans and contamination of the environment with C. burnetii requires transport through the atmosphere. It is assumed that C. burnetii is absorbed or fixed at the aerosol surface and becomes airborne. C. burnetii is resistant to heat and dryness and can survive for more than 150 days in the environment. Most ruminants, especially sheep and goats, spend their days grazing outside in the production areas of the highlands or on the eastern coast, where the highest density of ruminants and farms is met and where the winds are blowing most of the year. Additionally, manure is often used as a fertilizer in market gardening in these areas, potentially contributing to the spread of C. burnetii [35]. It is notable that contaminated aerosols are a major mechanism whereby C. burnetii is transmitted to humans [36]. MLVA genotyping results were in agreement with this risk factor because genotype 5 was observed in five farms located in a 3 km radius of the same area as the eastern windy part of the island [37].
Even if no correlation between pastures exposed to prevailing winds and animals kept at night in their barn was observed, these two variables support the assumption that C. burnetii may be transmitted via airborne route. Indeed, our study also showed that the risk of infection for ruminants was lower when the animals were kept in the barn at night. Generally, older cows that stayed in the cow barn for longer periods of time than young animals are more frequently infected. Hence, the probability of being exposed to the bacterium increases with exposure time [38]. However, in Reunion Island, most of the barns, particularly for sheep and goats, are open spaces sheltered from the wind. In these cases, the probability of infection by droplets and aerosols transmitted by wind is lower.
A lack of precautionary measures for visitors (such as washing hands and changing clothes and boots) before entering the farm was also associated with a higher risk for infection of ruminants with C. burnetii. The visitors, including veterinarians, food factory staff and professional hoof trimmers, may act as mechanical carriers and transfer the pathogen from infected to non-infected herds. This route of transmission has already been highlighted in previously reported articles [39], [40], suggesting that farm personnel often act as mechanical transmitters of contaminated fomites from an infected herd to uninfected ones.
Conversely, the risk of infection for ruminants was decreased when a proper quarantine was set up before any introduction of new animals to the farm. New ruminants are introduced after purchasing or, in the case of goats, when a male is borrowed from another farm to improve the reproductive performances. We should mention that young goats are often used in religious celebrations on Reunion Island. Again, MLVA genotyping results stressed the risk of infection when no quarantine is set up because, in our study, genotype 7 was detected in only in the two farms that purchased live animals from farm A [40]. A recent study reported that purchase of animals increased the risk of introducing C. burnetii infection into cattle herds [23]. This assumption stresses the risk of introduction of the bacteria both biologically and mechanically. Animals that live in close contact can become infected with C. burnetii because bacteria are shed from infected animals by vaginal secretions, placenta, urine or feces. A previous study described the occurrence of pregnancy terminations in goat herds that were exposed to three goats from another herd that reportedly kidded prematurely during a fair [41]. Moreover, when cows were imported into an area of endemic infection, 40% of uninfected cows became C. burnetii-infected within six months [42]. Viable bacteria have been isolated from sperm of seropositive bulls [43].
Our results demonstrate that even with a relative low seroprevalence in ruminants, C. burnetii is circulating consistently in the island. This was particularly evident in goats, where 21% animals were PCR positive. Questions emerged regarding the potential impact of C. burnetii on the general population as well as persons at risk, such as pregnant women. Thus, we have begun another study to assess the consequences of this bacterium on human health.
|
10.1371/journal.pcbi.1003666 | Augmenting Microarray Data with Literature-Based Knowledge to Enhance Gene Regulatory Network Inference | Gene regulatory networks are a crucial aspect of systems biology in describing molecular mechanisms of the cell. Various computational models rely on random gene selection to infer such networks from microarray data. While incorporation of prior knowledge into data analysis has been deemed important, in practice, it has generally been limited to referencing genes in probe sets and using curated knowledge bases. We investigate the impact of augmenting microarray data with semantic relations automatically extracted from the literature, with the view that relations encoding gene/protein interactions eliminate the need for random selection of components in non-exhaustive approaches, producing a more accurate model of cellular behavior. A genetic algorithm is then used to optimize the strength of interactions using microarray data and an artificial neural network fitness function. The result is a directed and weighted network providing the individual contribution of each gene to its target. For testing, we used invasive ductile carcinoma of the breast to query the literature and a microarray set containing gene expression changes in these cells over several time points. Our model demonstrates significantly better fitness than the state-of-the-art model, which relies on an initial random selection of genes. Comparison to the component pathways of the KEGG Pathways in Cancer map reveals that the resulting networks contain both known and novel relationships. The p53 pathway results were manually validated in the literature. 60% of non-KEGG relationships were supported (74% for highly weighted interactions). The method was then applied to yeast data and our model again outperformed the comparison model. Our results demonstrate the advantage of combining gene interactions extracted from the literature in the form of semantic relations with microarray analysis in generating contribution-weighted gene regulatory networks. This methodology can make a significant contribution to understanding the complex interactions involved in cellular behavior and molecular physiology.
| We have developed a methodology that combines standard computational analysis of gene expression data with knowledge in the literature to identify pathways of gene and protein interactions. We extract the knowledge from PubMed citations using a tool (SemRep) that identifies specific relationships between genes or proteins. We string together networks of individual interactions that are found within citations that refer to the target pathways. Upon this skeleton of interactions, we calculate the weight of the interaction with the gene expression data captured over multiple time points using state-of-the-art analysis algorithms. Not surprisingly, this approach of combining prior knowledge into the analysis process significantly improves the performance of the analysis. This work is most significant as an example of how the wealth of textual data related to gene interactions can be incorporated into computational analysis, not solely to identify this type of pathway (a gene regulatory network) but for any type of similar biological problem.
| Gene regulatory networks (GRNs) are DNA-encoded regulatory subsystems in the genome that coordinate input from activator and repressor transcription factors to control various biological functions, including development, cell differentiation, and response to environmental cues. They provide a systems level illustration of physiological function and are composed of modules at varying hierarchical levels [1]. A GRN provides the pathways of gene interactions within the context of location and time [2]. For instance, when a receptor on a particular cell receives a signal and initiates the activation of a transcription factor, the transcription factor increases the expression of the target gene, which in turn alters the production and activation of other pathway components in a cascading manner, changing the behavior of the cell.
The development of microarray technology allowed the discovery of gene regulation to move from individual interactions to thousands of interactions in parallel [3]. However, system-level techniques like microarrays produce large datasets, requiring efficient computational methods to identify significant changes in gene expression and their correlation. One area of systems biology where these computational methods are increasingly applied is for the inference of GRNs [4]–[9]. Although microarray databanks contain a wealth of data in support of the elucidation of GRNs, mammalian datasets are often limited by low or nonexistent replication and too few time points to allow for reliable results. There have been suggestions (e.g. Sîrbu et al. [10]) that the wealth of biological knowledge on possible interactions available in the literature coupled with the limits on available microarray data warrant an attempt to implement an integration of the two to improve reliability of GRN inferencing results.
We propose utilizing knowledge extracted from publications as a network of interactions and then applying microarray data to provide a measure of the quantitative effect of each of the individual interaction components. The qualitative knowledge at the core of our approach is provided by SemRep, a natural language processing system that extracts textual meaning from the biomedical literature in the form of semantic relations called predications (subject-relation-object triples). These predications mostly represent mammalian and specifically human interactions, therefore we use quantitative data provided by analysis of publicly available human breast cancer microarray datasets using a genetic algorithm. We compare the fitness of our model to that of a high-performing model from the literature to determine the performance of our technique as well as another model based on time-delay correlation. We also compare our results to KEGG pathways and find not only included interactions but also interactions not included but validated in the literature. We then modified the components of our method to be compatible with yeast data and again compared with the state-of-the-art model. Our method provides a novel approach to enhancing microarray data analysis with knowledge from the literature, and is the first such approach to incorporate semantic relations. Combining microarray data with semantic relations provides a more accurate model of gene interactions directing the behavior of cells, and consequently an enriched understanding of molecular physiology. This supports the discovery of disease mechanisms, which can then be exploited for diagnostic and therapeutic development in medicine.
Previous computational models used to reconstruct gene regulatory networks from microarray data have employed techniques such as Hidden Markov, Bayesian network, and stochastic differential equation models [11]–[15]. In a Hidden Markov Model, the real states of genes are treated as hidden variables, and the gene expression values are observed. This permits the states of genes at a given time t to be considered as depending only on the previous time point t-1. A Bayesian network is a representation of a joint probability distribution. When applying Bayesian networks to genetic regulatory systems, nodes are identified with genes and their expression levels, edges indicate interactions between genes, and conditional distributions describe these interactions. In the stochastic differential equation, the change of expression of a given target gene over two adjacent time points is equal to the accumulation of the weighted results of the sigmoid function of its regulator genes plus a random error. These computational models all reflect one or more aspects of the nature of genes and gene regulatory networks, but computational complexity limits the dimensionality of the modeled networks, and sensitivity to noise in gene expression measurement largely reduces their accuracy.
Genetic algorithms (GA) [16] have also been widely used for the inference of gene regulatory networks [17]–[20]. Genetic algorithms are inspired by Darwin's theory of evolution. Within this methodology, a population of candidate solutions to a problem is created and then evolves over a specified number of generations using phenomena such as cross-over (swapping components from other candidates) and mutation (internal, random changes). The best candidates are propagated through with incremental changes in the overall structure and the final generation becomes the solution. At each generation, fitness functions are used to determine the fittest candidates. A genetic algorithm is a stochastic algorithm and therefore it is highly likely to find global optima and can easily escape local maxima. Since the genetic algorithm execution technique is not dependent on the error surface, it is capable of solving multi-dimensional, non-differential, non-continuous, and even non-parametrical problems, which is the nature of gene expression data. Keedwell et al. used small random gene subsets evaluated by an artificial neural network (ANN) that is optimized by gradient descent to form the population of the genetic algorithm [21]; Liu and Wu used a differential equation to model the GRNs and genetic programming for optimization [22]; Sîrbu et al. compared various evolutionary algorithms for GRN inferencing and found that Keedwell et al.'s method performs the best overall [10], prompting us to use their method for the basis of our model. They also note that using literature-derived knowledge offers the potential for significant improvement over techniques that probe component genes randomly, and our use of such knowledge forms the basis of our deviation from the Keedwell et al.'s model.
With the rapid rate of growth in the biology literature, text mining is increasingly seen as indispensable in managing and discovering new biological knowledge [23]. An active area of research in biological text mining has been extraction of interactions between biomolecular entities (genes, proteins, etc.) from the research literature. Many systems, adopting various representational means (binary interactions, events, etc.) and using a variety of rule-based and machine learning-based techniques, have been proposed for this task. Early systems that focused only on co-occurrence of entities were soon replaced by systems that relied on shallow parsing and hand-crafted syntactic rules to extract binary interactions [24]–[26, among others]. These methods generally provided high precision at the expense of lower recall, in contrast to co-occurrence based methods. More recently, dependency parsing has become the predominant syntactic tool in extracting biological relations, as evidenced by the BioNLP Shared Task competitions [27]. These competitions have also signaled the increasing focus on events as the representational means for biological relations. Most commonly, dependency relations have formed the basis for syntactic features (shortest paths, dependency n-grams, etc.) for machine learning methods, along with lexical (tokens, n-grams, part-of-speech tags, etc.) and semantic (entity types, hypernyms, etc.) features. Best machine learning approaches have included pipeline models based on support vector machines [28]–[29] as well as model combination techniques [30] and joint inference [31]. Some rule-based systems have reported competitive results in this task, as well [32]. Recently, coreference resolution has also been beneficially integrated into several event extraction systems [33]–[34].
As these text mining methods mature, they are increasingly applied to practical needs of biologists, in tasks such as database curation and pathway generation. For example, two tasks in the recent BioNLP 2013 Shared Task competition [27] (Pathway Curation [35] and Gene Regulation Network in Bacteria [36]) investigated the feasibility of automatically constructing such networks from the literature alone. Given a set of relevant biomolecular entities, the former focused on extracting pathway-relevant events (e.g. gene expression, regulation, binding, regulation), while the latter focused on constructing a gene regulation network for the model bacterium Bacillus subtilis involving these entities. For the latter task, participating groups could either directly construct a regulation network for the provided entities or extract the interactions from which such a network could be derived using a predefined algorithm that the organizers provided. While the results were encouraging, these tasks remain challenging as evidenced by the limited participation and the fact that both tasks presupposed that the entities involved were already known, thus, addressing only a fraction of the problem of network construction from the literature. From an opposing viewpoint, Miwa et al. [37] focused on linking interactions in biological pathways to supporting evidence from the literature; however, their work does not address the task of pathway construction.
There have been several attempts at improving gene regulatory network modeling by incorporating existing knowledge. For example, Steele et al. used the correlation between different gene concept profiles to calculate the probability of edges in GRNs modeled by Bayesian networks [38]. These profiles are determined by the occurrence of terms in the literature, using the Unified Medical Language System (UMLS) for normalization. Gutierrez-Rios et al. used regulatory interactions described in RegulonDB, a database of the regulatory network of Escherichia coli K-12, to establish the network of causal relationships to evaluate the congruence between the literature and whole-genome expression profiles [39]. Additionally, the literature contains examples of efforts to combine literature-derived networks and microarray analysis in contexts other than GRN inferencing. Duarte et al. reconstructed the human metabolomic network depending largely on a manual literature review combined with knowledge extracted from genomic databases and use gene expression analysis to fill in the gap for a subset of network components [40]. Ashley et al. used text mining to derive biological pathways from literature relevant to in-stent restenosis and analyzed which of these were most relevant to the expression profiles identified by microarray analysis of tissue samples from patients with this condition [41]. All of these techniques either use human review to identify asserted interactions or automated approaches to infer them based on co-occurrence of terms. Our approach combines automation, allowing for an exponentially greater survey of the literature, with the identification of assertions in the text by SemRep, moving beyond mere term co-occurrence and increasing the validity of the extracted relations.
Several books, book chapters, and journal reviews are available detailing the pros and cons of various modeling approaches for GRN inferencing [42]–[47]. In addition to a description of various approaches, Karlebach and Shamir [45] also provide a comparative summary of the relative advantages of different types of models for various features. They align GRN models along a spectrum from logical models (e.g. Boolean networks) to continuous models (e.g. linear differential equations) to single-molecule level models (e.g. stochastic simulation models). They identify the logical end of the spectrum as having a decreased detail, less faithfulness to biological reality, lower data quantity needs, and reduced ability to model dynamics, while having greater model size, computational speed, and inferencing ability. Models at the single-molecule level are positioned at the other end of the spectrum, with a higher level of detail, increased faithfulness to biological reality, greater data quantity needs, and increased ability to model dynamics and decreased model size, computational speed, and inferencing ability. Continuous models are positioned in the middle of the spectrum, with moderate levels of the assessed model's characteristics. Although not providing a complete picture of the comparative characteristics of GRN modeling techniques, they give an easily accessible summarization.
The breast cancer microarray data used in our experiments comes from NCBI's Gene expression omnibus (GEO: http://www.ncbi.nlm.nih.gov/geo/) [59]. GEO provides free access to raw and processed microarray and sequencing data submitted by researchers based on their published work. For the yeast study, we used the cdc15 time course data from the Yeast Cell Cycle Analysis Project website (http://genome-www.stanford.edu/cellcycle/) originally used in [60].
For assessment of our methodology, we use the Kyoto Encyclopedia of Genes and Genomes (KEGG: http://www.genome.jp/kegg/) [61], which provides gold standard sets of molecular pathways. KEGG pathways are manually curated networks based on a review of protein-gene and protein-protein interactions described in the literature. It is worth noting that KEGG pathway maps are an abbreviated representation of known interactions, focusing on those considered to be best supported by evidence and most relevant. As a consequence, the interactions in a KEGG pathway form a subset of those in the literature, and it is necessary to assess separately the validity of interactions not in the KEGG pathway. A KEGG pathway can be accessed as a downloadable kgml file or an online map, which have slight differences regarding which genes are included and which interactions are specified. To mediate these differences we use both formats for our assessment and consider whether genes and interactions are included in either version.
We infer gene regulatory networks in three steps, illustrated in Figure 1.
To facilitate the establishment and assessment of the system model, we select the well-studied disease breast cancer as the starting point, due to the availability of relevant citations, microarray data, and established KEGG pathways. The 13 pathways contained in the KEGG Pathways of Cancer (human) map (http://www.genome.jp/kegg-bin/show_pathway?hsa05200) were used to guide the predication network generation and also in evaluating the resulting GRNs. These pathways are the p53 (see Figure 2), Apoptosis, Cell Cycle, PPAR, VEGF, MAPK, Wnt, TGF-beta, mTOR, jak-STAT, ErbB, Focal Adhesion, and Adherens Junction pathways.
In the first step, predication network generation, predications containing gene-gene interactions (INTERACTS_WITH, STIMULATES, and INHIBITS predications) are extracted from the MEDLINE citations that supported manual curation of each KEGG pathway; these citations are listed in the pathway entry on the KEGG website. These predications are augmented with those from additional relevant citations, which we identify by first extracting the Medical Subject Heading (MeSH) terms for each citation identified in KEGG and then manually refining that list to eliminate non-specific terms such as Humans, Male, or Biological Models. The resulting MeSH terms formed the basis of our PubMed searches. The terms that occurred in a significant distribution across the citations were grouped with an “OR” in the query when they were roughly equivalent or part of a set of different subtopics. As an example, the query for p53 citations was “(Tumor Suppressor Protein p53[mh] OR Genes, p53[mh]) AND (Apoptosis[mh] OR Signal Transduction[mh] OR (Phosphorylation[mh] AND (Neoplasms[mh] OR Neoplasm Proteins[mh] OR Tumor Markers[mh]))) AND physiology[sh] AND metabolism[sh].” This procedure was repeated to provide a citation list for each pathway. The number of citations for each pathway is given in Table 1 in ‘Citation’ column. The predications extracted from these citations were then retrieved from SemMedDB. The number of resulting predications for each pathway is shown under the column heading ‘Raw’ in Table 1.
To improve the reliability of our approach, we used three predication filtering mechanisms, explained below. These mechanisms result in a smaller set of predications extracted for each pathway, which forms subnetworks. These subnetworks, together, serve as the predication network. Figure 3 provides an example of a subnetwork created from predications related to p53. This example has been pruned to provide a small network for simplicity.
Argument-predicate distance is the number of intervening noun phrases between an argument and its predicate in the sentence from which they were extracted. It has been shown that smaller argument distance leads to higher precision in extracting interactions [62]. In the example below, both Bax and pro-caspase-3 are potential objects for p53 activity.
Bax has an object distance of 1 since it is the first noun phrase subsequent to the predicate ‘promoted’ and pro-caspase-3 has a distance of 2. Preferring an argument-predicate distance of 1 selects the predication p53 STIMULATES Bax, while eliminating the predication p53 STIMULATES pro-caspase-3. In our experiments, we limit both the predicate-subject distance and the predicate-object distance to 1, which decreased the number of predications to approximately 27% of the initial number of predications on average. The number of predications for each pathway after filtering using argument-predicate distance of 1 is shown in Table 1 (under ‘Dist.’).
After argument-distance filtering is applied, we normalize the set of predications by removing duplicate predications and mapping the subjects and objects to formal gene symbols based on the standard gene name dataset (HUGO, http://www.genenames.org/). If a predication argument cannot be mapped to a formal gene symbol, the predication is pruned. The numbers of predications after duplicate removal and normalization are given in Table 1 (under ‘Uniq.’ and ‘Norm.’, respectively).
Document frequency for a predication is the number of citations in which the predication occurs. Document frequency filtering is based on the hypothesis that the confidence credential of a predication is in direct ratio to its occurrence in documents. In our experiments, we discard all predications that occur in fewer than two articles. The result is shown in Table 1 (‘Freq.’). Note that network filtering discards all relevant predications for the Focal Adhesion pathway, which was not considered in subsequent steps.
In the final step we quantify gene-gene interaction strength in the network generated in previous steps, using a genetic algorithm, depicted in Figure 4. In the initial population of chromosomes, each chromosome contains a candidate set of interactions between all genes in the pathway. The predication network determines the initial gene-gene interactions, while the strength of interaction is initially randomly generated for each of the 2000 chromosomes in the population. Then for each generation, the fittest candidates are replicated into the next generation and the balance of the population is filled through reproduction of the current generation, using random crossover and mutation to introduce novel diversity into the population. We compute the fitness of a chromosome, as compared to the time series microarray data, using an artificial neural network. These procedures are described in more detail in the following subsections.
With the genetic algorithm used to infer the gene regulatory networks, we train the weights of the inbound interactions for every gene separately. A population of chromosomes is created where each chromosome is represented as a matrix of interaction weights between each possible pairing of genes identified from the predications. Each weight, valued between −1 (greatest inhibition) and +1 (greatest stimulation), indicates the strength of interaction. A weight of zero indicates no interaction. The predications in the network define which genes have interactions and whether the interaction is inhibitory or stimulatory, but do not contribute to the determination of the weight. The absence of an interaction is represented in the matrix with a weight of zero but is not altered through subsequent generations. The direction of the interaction is maintained from the subject-object relation in the predication. We randomly generate 2000 chromosomes that contain different weights for the interactions in the predication network. The gene structure in each chromosome is the same, as well as the non-interacting/zero-weighted gene pairs, but the weights representing the strength of inhibition or stimulation of pairs found in the predications are varied. The population of chromosomes is then evaluated with the fitness function in Equation 1, and the fittest 20 percent of the chromosomes are copied into the next generation directly. The rest (80 percent) of the chromosomes for the next generation are generated by crossover at a specific (randomly selected) gene pairing between pairs of randomly selected chromosomes in the current generation with a mutation rate of 0.25 percent. This mutation rate approximates those commonly used to facilitate convergence of the chromosomes toward a fittest result and to avoid excessive intergenerational fluctuation. After evolving for 200 generations (an empirically determined limit), the chromosome with the highest fitness value is selected as the final result.
In our experiments, we use an artificial neural network (ANN) to model the interaction of the genes in each pathway. In this model, the gene expression level of a given gene at a given time point is a function of all other genes at the previous time point (see Figure 5). In a comparative study by Sîrbu et al., this model outperformed 4 other recent GRN models (GA + evolutionary strategy, differential evolution + Akaike's Theoretic criterion, genetic local search, and an iterative algorithm based on GA) in a combined score of 6 performance measures (data fit, parameter quality, noise, sensitivity, specificity, and scalability) [10]. We follow the most common implementation of such a model, using a nonlinear weighted sum. Equation (1) shows the fitness test function based on the ANN model.(1)where(2)and T indicates the number of time points in the microarray dataset, R is the number of microarray replicates at each time point, r is the current replicate, n is the number of genes in the pathway, K is the activation function, is the weight of the interaction between gene j and gene i, and is the gene expression value in the microarray set for gene i at time point t. Since the difference of the expression value of a gene over each time course is considered to be a direct function of only the joint effect of other genes in the pathway and the gene expression values used have been normalized between −1 and 1, we use a linear function as the activation function, effectively dropping the term. A visual representation of the ANN model is presented in Figure 6.
MEDLINE contains many more citations related to human genes than yeast genes. A simple PubMed search for “yeast and gene” returns a little more than 86,000 citations while the query “human and gene” returns over a million. When combining either species term with “cell cycle” (a concept with a heavy amount of research done in yeast), there are over 96,000 citations related to human and just over 15,000 for yeast. In addition to this limitation, SemRep was designed to identify human genes and uses Entrez Gene database entries specific to humans for genes and proteins. Although this led to an initial study with human genes, we made modifications to SemRep to support a study with yeast. This required changing the data source to the Entrez Gene fungi dataset that contains Saccharomyces species and processing relevant citations (resulting from the PubMed query “Saccharomyces[mh] AND cell cycle[major]”) to extract predications. All subsequent procedures were consistent with the breast cancer study except that the predications were not filtered by argument distance or document frequency due to the lower initial number of predications. The number of citations and predications at each step are given in Table 2.
Evaluating a gene regulatory network inference model for human data is challenging, since there is no gold standard providing a complete reference of gene connectivity. Therefore, we evaluated the implementation of our approach using breast cancer in two ways: (1) comparing the accuracy of the model against the highest-performing model reported in the literature, (2) comparing the resulting networks to KEGG pathways and literature that formed the basis for their extraction, paying particular attention to the p53 pathway. Comparison to a KEGG pathway provides an assessment of the contribution of our natural language processing techniques (i.e., SemRep).
As a measure of accuracy of the model, we determine how well the model fits the data, in line with previous research [21]. For this purpose, we use microarray data from a human breast cancer experimental set (GEO: GSE29917), which contains expression values for 7 time points and 6 replicates (two microarrays are missing from the set for a total of 40 microarrays). We compare the accuracy of our model with that of the Repeated Genetic Algorithm with Neural Network model described by Keedwell et al. [21], the best performing model in a recent comparison of evolutionary algorithms [10] and the basis for the interaction quantification component of our algorithm. Since we use the same algorithm for interaction quantification, the comparison helps isolate the effect of literature-derived knowledge on GRN inference. We downloaded their source code to be able to run their algorithm on our data, with only minimal modification for data format differences. Additionally we included a second comparison model, with a different type of algorithm (time-delayed Spearman Rank-correlation or TDSRC) to help identify any biases in our methodology [63]. This model was included by Gupta et al. as a part of a composite model and available for download and use within MatLab. For this comparison, we limit the microarray data to 78 genes included in the KEGG P53 pathway (listed in Table 3). We use a leave-one-out approach based on time points, sequentially using gene expression values for each time point as test data against models trained with values from all of the other time points. Fitness is defined as the standard deviation of predicted values from the average microarray expression of all 6 replicates in the test set. Because the microarrays for 2 different replicates at two different timepoints were missing, the total number of microarrays was 40. The root mean square error (RMSE) combines fitness for all genes in the given test set. Figure 7 shows the root mean square errors over the 78 genes for each time point for each model. We used a paired t-test to assess the statistical significance of the differences between models. The p-values for each comparison are included in Figure 7. Our model shows improvement at every test data set/time point over the Keedwell et al. model and is significantly different (p<0.05) in 4 out of 6 time points and for all time points combined (overall p = 8.26×10−8). The RSP model performed significantly worse than both of the other models. This demonstrates that our model is significantly better in terms of fitting the microarray data.
We assessed the value of gene regulatory networks generated by our approach by comparing them to KEGG pathways, both the kgml format downloaded from the KEGG website and manually using the search function in the online pathway map. Nodes, representing genes, and edges, representing interactions, were independently compared and the results are shown in Table 4. The ‘Predication’ column represents the number of nodes and edges in the network generated by our approach after removing any genes that were not included in the microarray set, the ‘KEGG’ column provides the equivalent information for the kgml network, ‘Common’ indicates the intersection of our results and the corresponding kgml network, and ‘New’ provides interactions exclusive to our results. As the numbers indicate, new nodes and edges significantly outnumber those in common between the two networks. The p53 pathway had the highest ratio of common edges compared to the kgml network (19∶65, 29.2%), so was chosen for further validation at the predication level.
A comparison was made for each of the new interactions with the online KEGG p53 signaling pathway map. 49 of the 92 interactions contained a gene not included in the map, 7 had both genes present but no interaction, and 9 existed in the map but were not included in the kgml version.
We also validated each new interaction in the resulting p53 network against the literature in three ways: a) whether the literature asserts an interaction between the two at either a gene or protein level (i.e., the precision of our natural language processing techniques), b) whether we capture correctly the direction of the effect, i.e. stimulatory or inhibitory, as compared to the weight generated by GRN analysis of the microarray data, and c) combining the two above, whether we capture correctly both the presence of the interaction and the direction. In addition to assessing the accuracy of SemRep in capturing relevant interactions, this validation allows us to establish the specific contribution of using semantic predications over term co-occurrence. The comparison was limited to citations from which the predications were extracted, ranging from 1 to 87 citations for each interaction. Although there may exist another citation that would validate the resulting interaction, this approach was taken to limit the man-hours required to a reasonable amount while still allowing a reasonable possibility for validation. As seen in Table 5, 78.3% of the new interactions in the resulting p53 network (Figure 8) were asserted in at least one of the source citations. When comparing the sign of the interaction weight to the direction of the effect provided in the source literature, 76.4% were consistent. Those interactions that were both consistent with the literature and were weighted in the appropriate direction numbered 55 out of 92 (59.8%).
We additionally focused on those interactions having a weight with absolute value >0.1 (Table 5, bottom), exploring the hypothesis that stronger interactions should be less affected by noise. Within these interactions a total of 32 interactions (84.2%) were stated in the literature and 34 (87.5%) were correct in the direction of effect, yielding 28 (73.7%) correct on both counts.
Finally, we investigated whether argument-predicate distance filtering had a detrimental effect on the results, since long distance syntactic dependencies are common in biomedical literature [64]. For this purpose, we focused on the Jak-STAT pathway and checked whether any of the 27 KEGG interactions, none of which appeared in our predication network, could be derived from the initial set of unfiltered, non-normalized predications. We found three such predications for the Jak-STAT pathway; one (Jak1 INTERACTS_WITH PTPN11) was eliminated due to argument-distance filtering, while the other two (Jak1 INTERACTS_WITH STAT1 and GRB2 INTERACTS_WITH PTPN11) passed the argument-distance filter but not the document frequency filter. Note that the former (Jak1 INTERACTS_WITH PTPN11) occurred in a single document and, therefore, would also be eliminated in the subsequent document frequency filtering step, had it passed the predicate-argument distance filter.
We compared the performance of our model against that of Keedwell et al. in the same manner as mentioned in our results section Comparison of model accuracy, but now on the yeast cell-cycle dataset, which contains 24 time points. Figure 9 shows the root mean-square errors for each time interval for the two models. We assessed statistical significance of the differences between models by calculating the p-values (included in Figure 9) using a paired t-test. Using the yeast data, our model had increased fitness at every interval over the Keedwell et al. model, but this time every difference was significantly different (p<0.05).
In this work, we augmented microarray data with literature knowledge to infer gene regulatory networks, replacing the random selection of component genes generally used in similar modeling efforts. We use SemRep to extract information from the literature, which forms the backbone of the network. State-of-the-art genetic algorithm-based analysis of the microarray data was used to determine the strengths of the effects between gene-gene interactions in the network. Our model provides better fitness than the state-of-the-art model used as the basis for the genetic algorithm and fitness function components of our model and the difference between error rate of the models is statistically significant (p<0.05), demonstrating that literature-derived knowledge provides a significant advantage over random selection of genes in this task, as suggested in [10]. Another advantage of our approach is that it maximizes the best possible networks without being limited to relationships that are already well established and considered important by a curator, as would be the case if a standard interaction database were used. These pathways provide targets for novel therapeutic interventions and a mechanistic understanding of current therapeutic approaches with poorly understood mechanisms.
Although some of the genes in the KEGG pathways did not appear in the results due to the strict filtering process, the presence of essential member genes in the resulting predication networks (for example, TP53, MDMD2, BAX, CDKN1A, GADD45A, and CDK2 in the p53 network) and known interactions (p53 STIMULATES Gadd45a, E2F1 STIMULATES cdkn2a, and RCHY1 INHIBITS tp53) demonstrate the potential of this technique to replicate “known” pathways. Perhaps more importantly, we were able to identify new interactions that were not included in the KEGG maps, which is not surprising since these maps are curated and therefore provide only the most thoroughly established and important interactions as determined by the curators.
Two of the highest-weighted new interactions included in the resulting p53 pathway but not in the kgml file are p53 STIMULATES BIK and MDM2 INHIBITS CDKN1A. The weight of the interaction between p53 and BIK was very strongly stimulatory at 0.977. BIK (BCL-2 interacting killer) is not included in the KEGG P53 pathway map. It is a pro-apoptotic protein discovered in 1995 and interacts strongly with BCL-2 and BCL-xL [65]. P53 has been shown to induce expression of BIK under certain conditions, especially in breast cancer cells, as in our microarray dataset [66]. Our p53 pathway result also included an interaction between MDM2 an CDKN1A with a weight of −0.987, i.e., very strongly inhibitory. MDM2 is shown to inhibit p53 in the KEGG p53 map but there is only an indirect inhibitory action of MDM2 on CDKN1A through P53 (by diminishing p53's activation of CDKN1A). However, a direct interaction as a negative regulator of CDKN1A is present in the literature [67], supporting our result.
A major limitation of this method is the accumulation of system errors. As shown in Table 5, the accuracy of interactions identified using SemRep with the filtering used in our method is 78.3% overall and 84.2% for significantly weighted interactions (>0.1). The reproducibility of the microarray data between platforms and even with different algorithms on the same platform has been determined to be as low as 50–60% [68]. At a result, the accuracy of both the structure and the weights of the resulting GRNs is limited if viewed at the most precise level of granularity. The training of the weights is also limited by an insufficient time course in the microarray data, especially for large numbers of arguments (i.e. genes in a pathway). Although it is beyond our means to improve microarray reproducibility or experimental design of publicly available datasets, improvements to SemRep with regard to gene and protein interactions can potentially improve results and the incorporation of multiple microarray datasets as training data may be able to overcome some of the inherent limitations. Additionally, although these limitations affect the use of these techniques for determining the precise quantitative nature of interactions, this is true generally of such studies, and does not prevent their use for hypothesis generation.
Currently our network backbone is limited to genes included in the predications, but this set could be expanded by incorporating genes from the microarrays that are similar to the genes provided by the predications, using an ensemble of machine learning techniques such as support vector machines, random forests, and prediction analysis of microarrays. This approach of classifying genes from the microarray into each pathway would additionally reduce bias in the technique by not limiting interactions to what is already known. It would be valuable to maintain a distinction between the literature-based genes and the expansion set because a significant expansion in the final resulting networks would suggest that what is known in these pathways is only a fraction of what is waiting to be discovered.
Another technique to extend the set of genes would be to adopt more sophisticated predication filtering mechanisms. The current mechanisms were effective in significantly reducing the computational complexity; however, our limited analysis of the effect of argument-distance filtering showed that some relevant predications were also eliminated due to filtering. A potentially useful approach would be to train a classifier that can identify ‘good’ predications, based on predication features, such as the predicate type, indicator types, etc. Predicate-argument distance could also serve as a feature for such a classifier.
Additionally, although we used interactions from KEGG to evaluate our GRN, such interactions from this and similar databases could be incorporated into our interaction network to augment the proposed interactions from literature with established interactions, expanding the utility of prior knowledge.
Although our current method of selecting source citations for predication extraction yielded usable results, there are many possible permutations in the method and a systematic comparison of source citations and their resulting networks should be explored. Our current search method facilitates the generation of hypothetical member genes for established pathways, but this approach can also be used to generate novel pathways by specifying a gene set and/or biological functions in the predication citation search, thereby providing an appropriate predication network.
SemRep representation of biomolecular interactions (subject-predicate-object triples) is simple, intuitively accessible and lends itself easily to downstream applications. On the other hand, more complex representations, particularly event representation (as discussed in Related work section), have been gaining in popularity in the BioNLP community, mostly due to available corpora [69] and shared task competitions [27]. An obvious question is whether such complex representations could provide an advantage over or could complement SemRep representation in the task of gene regulatory network inference. It seems very likely that generating a seed network from more complex representations would require some non-trivial post-processing along the lines of the algorithm described by Bossy et al. [36]. Since that algorithm essentially breaks down complex relations to simpler SemRep-style triples, it seems safe to assume that SemRep representation can adequately capture the complexity of biomolecular interactions. However, this needs further testing and validation.
To assess the precision of interactions for the yeast study, we performed an automated comparison using Cytoscape (http://www.cytoscape.org) against a Biogrid (http://thebiogrid.org) yeast interaction dataset. We used the Biogrid Saccharomyces_cerevisiae-3.2.109 tab file, which contains 339,921 interactions and experimental information from multiple database sources including Saccharomyces Genome Database, MINT, IntAct, and Pathway Commons. As seen in Table 6, although 346 out of 349 genes are found in the reference set, only 147 out of 520 interactions are included. That is a relatively low precision of 28%, which only increases to 31% for interactions weighted 0.1 or higher. This is not particularly meaningful without appropriate context. Since this is a relatively straightforward and easily undertaken evaluation approach, it would be worthwhile to conduct a study applying the same evaluation across various published models to see how precision compares among them.
We present a methodology of gene regulatory network inference that combines literature knowledge and microarray data. Using SemRep, we extract gene and protein interactions from citations related to the pathways included in the KEGG Pathways of Cancer. These predications are linked together to form a network and a genetic algorithm is used on a breast cancer time sequence microarray dataset to determine the weight of contribution of each stimulating or inhibiting gene on its target, thereby providing a weighted gene regulatory network. Our model performs significantly better than comparable models in terms of fitness of predictive output to microarray results. The resulting networks contain both interactions included in the appropriate KEGG pathways and interactions not included but validated through a literature search. The accuracy of these interactions increases from 60% overall to 74% when minimally-weighted interactions are excluded. Our model also performed better in terms of fitness against a comparison model when modified for yeast predications and microarray data. This approach offers significant potential in elucidating new interactions in existing pathways as well as the possibility of identifying novel pathways. In a broader sense, it also provides a blueprint for incorporating automatically extracted knowledge from literature with large-scale biological analysis. Incorporating such knowledge underpins more accurate understanding of both normal and disturbed molecular physiology, leading to advances in diagnosis and treatment of disease.
|
10.1371/journal.pgen.1002181 | Pervasive Sign Epistasis between Conjugative Plasmids and Drug-Resistance Chromosomal Mutations | Multidrug-resistant bacteria arise mostly by the accumulation of plasmids and chromosomal mutations. Typically, these resistant determinants are costly to the bacterial cell. Yet, recently, it has been found that, in Escherichia coli bacterial cells, a mutation conferring resistance to an antibiotic can be advantageous to the bacterial cell if another antibiotic-resistance mutation is already present, a phenomenon called sign epistasis. Here we study the interaction between antibiotic-resistance chromosomal mutations and conjugative (i.e., self-transmissible) plasmids and find many cases of sign epistasis (40%)—including one of reciprocal sign epistasis where the strain carrying both resistance determinants is fitter than the two strains carrying only one of the determinants. This implies that the acquisition of an additional resistance plasmid or of a resistance mutation often increases the fitness of a bacterial strain already resistant to antibiotics. We further show that there is an overall antagonistic interaction between mutations and plasmids (52%). These results further complicate expectations of resistance reversal by interdiction of antibiotic use.
| Bacteria can become resistant to antibiotics by spontaneous mutation of chromosomal genes or through the acquisition of horizontally mobile genetic elements, mainly conjugative plasmids. Plasmid-borne resistance is widespread among bacterial pathogens. Plasmids generally entail a cost to the host, associated with the replication and maintenance of the genetic element and with the expression of its genes. Therefore, in the absence of antibiotic, both plasmids and resistance mutations are often deleterious and confer a fitness cost to the cell. Here we studied epistatic interactions between five natural conjugative plasmids and ten chromosomal mutations conferring resistance to three types of antibiotics, making a total of 50 different combinations of chromosomal mutations and conjugative plasmids. We show that sometimes plasmids confer an advantage to bacterial strains carrying resistance mutations in their chromosome. This occurs in 32% (16 out of 50) of tested combinations. Furthermore, in 5 out of 50 plasmid-mutations combinations studied (10%), we observed an increased fitness when a plasmid-bearing bacterial cell acquires a drug-resistant mutation. These examples of sign epistasis are highly unexpected. This work explains, at least in part, how multidrug resistance evolved so rapidly.
| Multidrug resistance is a major hurdle for modern medicine, putting at risk commonplace medical practices [1] and the treatment of infection by bacterial pathogens [2]. Bacteria can become resistant by spontaneous mutation of chromosomal genes or through the acquisition of horizontally mobile genetic elements [1]. In the absence of antibiotics, resistance mutations are often deleterious and confer a fitness cost to the cell [3], [4], [5], [6]. It is logical to expect that, in the absence of antibiotic selective pressure, resistant strains will be outcompeted by the susceptible ones. Thus, a possible procedure to eliminate resistance is to ban the use of an antibiotic. This policy has been applied in different countries with varying results. For example, a deliberate reduction in the prescription of macrolides in Finland, resulted in a 50% decrease in the frequency of macrolide-resistant group A streptococci [7]. However, in the UK, a 98% decrease in the consumption of sulfonamides was accompanied by an increase of 6.2% in the frequency of sulfonamide resistance in Escherichia coli [8]. Clearly, there are other factors affecting the reversal to susceptibility. For example, resistant-bacteria often gain second-site mutations that ameliorate the fitness cost of resistance [3], [4], [9], [10], [11], [12], [13]. Sometimes, compensatory mutations even increase the level of resistance itself [14], [15].
The exchange of accessory genetic elements, in particular of conjugative plasmids, is frequent [16], [17], [18] and can disseminate genes among related and phylogenetically distant bacteria [16], [18], [19]. In addition, conjugative plasmids are able to mobilize other plasmids from a donor to a recipient cell [20]. Thus, resistance genes can quickly spread among bacterial communities. Plasmid-encoded resistance is generally the result of the activity of efflux pumps, agent-modifying enzymes [3], or protection of the antibiotic target [21].
Harboring mobile genetic elements generally creates a cost to the host, associated with the replication and maintenance of the genetic element and with the expression of its genes. Such cost has been experimentally demonstrated in a number of resistance-encoding plasmids [22], [23], [24], [25], [26].
A recent study of the interaction between resistance-determining chromosomal mutations, responsible for resistance to nalidixic acid, rifampicin and streptomycin in E. coli, found that, in the majority of the cases, the combined fitness cost of double resistance is smaller than one would expect if they were independent [27]. Gene interaction, or epistasis, is generally accepted as being relevant for the understanding of the evolution and dynamics of complex genetic systems [28]. Epistasis can vary in strength and form. When epistasis affects fitness, one can expect two possible outcomes. A positive epistatic interaction has an antagonistic effect on deleterious mutations. Thus, the double mutant has a higher fitness than the expected sum of costs. Negative epistasis between deleterious mutations creates a synergistic effect. Here, the double mutant is less fit than the expected sum of costs. Different studies of epistasis gathered evidence for both antagonistic and synergistic gene interaction. Positive epistasis between random deleterious mutations has been experimentally detected in phage ΦX174 [29], HIV-1 [30], RNA virus Φ6 [31], Salmonella typhimurium [32] and in the yeast Saccharomyces cerevisiae [33]. Other studies have found no evidence of epistatic interactions within the HIV-1 transcriptional promoter [34], in RNA virus [35], [36] and in S. cerevisiae [33], or evidence that positive epistasis occurs as often as negative epistasis in RNA virus [37] and in E. coli [38].
Here we focus on the interplay between conjugative plasmids and chromosomal mutations in E. coli. In particular, we look at how bacterial fitness is affected by genetic interactions between plasmids and resistance mutations. First, we quantify the degree of epistasis between five conjugative resistance plasmids (R124, R831, R16, R702 and RP4, carrying between one and four resistance genes and belonging to four different incompatibility groups) and 10 mutant alleles of the housekeeping genes gyrA, rpoB and rpsL, conferring resistance to nalidixic acid, rifampicin and streptomycin, respectively. The plasmids were isolated from nature and the resistance mutations are polymorphic in natural populations of different species of bacteria [13]. These genes are involved in different steps of the cell's essential flow of information from DNA to protein. Specifically, gyrA codes for DNA gyrase, an enzyme involved in DNA replication. Nalidixic acid and other quinolones inhibit DNA replication by binding to DNA gyrase and resistance to this class of drugs arises from the prevention of this binding. Rifampicin belongs to the rifamycin class of antibiotics which bind to the β-subunit of RNA polymerase, coded by rpoB, thereby inhibiting transcription. Mutations in rpsL, which codes for ribosomal protein S12, interfere with translation and can produce resistance to streptomycin by blocking the binding of this drug to the ribosome 30S subunit. Secondly, using the same plasmids, we estimate epistasis between pairwise combinations of conjugative plasmids inside the same cell.
We find pervasive sign epistasis in the interaction between resistance mutations and conjugative plasmids. This implies that the acquisition of an additional resistance plasmid to the existing chromosomal resistance or the appearance of a chromosomal drug-resistance mutation in a bacterial cell already containing a plasmid may ameliorate the initial fitness cost of resistance and therefore complicate resistance reversal. We also observed an overall positive level of epistasis between mutations and plasmids. Both the chromosomal allele and the plasmid seem to contribute to determine the nature of the epistatic interaction, although the host genotype appears to have a more determinant effect. In contrast, the interaction between plasmids exhibit sign epistasis only once, and, despite the occurrence of several cases of somewhat strong epistasis, on average it appears to be null.
Pairwise epistasis, , between loci A and B can be measured as follows. Suppose that the wild-type strain contains alleles A and B. If and are the fitnesses of each of the single mutants relative to the wild-type strain, and the relative fitness of the double mutant, then multiplicative epistasis is given by: . To estimate epistasis between plasmids and mutations, we defined these quantities in a similar way. If is the relative fitness of the strain with the wild-type allele (A) and containing a plasmid, is the relative fitness of the mutant strain (with A allele replaced by the a allele), and is the relative fitness of the strain containing both the mutation (a allele) and the plasmid, then epistasis between a plasmid and a chromosomal mutation becomes: .
Each conjugative plasmid was introduced in E. coli K12 MG1655 cells by conjugation. Then, we determined the fitness cost due to the presence of each plasmid relative to plasmid-free E. coli K12 MG1655 cells. This was performed using a competition assay, in the absence of antibiotics (see Materials and Methods). Fitness costs of plasmids span from 2.8% to 8% (Table S1).
The fitness cost imposed by ten different spontaneous antibiotic-resistance mutations was previously determined in ref. [27]. Table S2 presents the clones chosen from ref. [27] and the fitness costs of these mutations. Fitness costs of mutations vary between 0.5% and 27.5% (Table S2) [27].
To screen for epistatic interactions between chromosomal mutations and conjugative plasmids, we further constructed, by conjugation, all possible 50 combinations between these ten mutations and the five plasmids. Then we determined the fitness for each of these 50 combinations.
We found that 52% of the interactions present positive epistasis and only 8% present negative epistasis (Figure 1A). Figure 1A additionally shows that the nature of the epistatic interaction is not gene but allele specific. In fact, the conjugative plasmid influences how a specific allele interacts with plasmid-borne resistance determinants. This means that, depending on the plasmid, an allele can display no epistasis, positive epistasis or negative epistasis. For example, allele gyrA D87G exhibits no epistatic interaction with plasmid R124, however the same allele displays negative epistasis with plasmid R831 and positive epistasis with plasmids R16, R702 and RP4 (Figure 1A). The same pattern (allele specific nature of epistasis) had been observed for epistasis between resistance chromosomal mutations [27]. Supporting the pervasive nature of antagonistic interactions between mutations and plasmids, the distribution of the ε values (Figure 1B) has a significant positive median (median = 0.037, bootstrap 95% CI [0.021; 0.065]). Figure S1 plots the observed fitness against the fitness expected in the absence of epistasis (ε = 0).
To rule out the existence of compensatory mutations, we reconstructed five (double) combinations independently and in the opposite direction from what we did before: gyrA S83L(R16), rpoB I572F(R831), rpoB H526N(R16), rpoB R529H(R702), and rpsL K43N(RP4). We constructed these five clones by transducting [27] the antibiotic resistance mutation from our mutant E. coli strains into the wild-type strain (E. coli K12 MG1655) already containing the plasmid. Using this method we decreased the number of generations involving the antibiotic-resistance mutation by a half (because the plasmid was already there). In this way, we decrease the probability of occurrence of compensatory mutations. We measured the fitness of two independent clones corresponding to each of the five (double) combinations. For all five combinations, fitness values are not significantly different from the ones obtained using the previous method (Kruskal-Wallis, p>0.05). Four (gyrA S83L(R16), rpoB I572F(R831), rpoB H526N(R16) and rpoB R529H(R702)) of these five combinations correspond therefore to cases of sign epistasis (Figure 1A). One combination (rpsL K43N(RP4)) shows no interaction with these new independent clones as observed before (Figure 1A). The fact that independent clones exhibit the same fitness (and the same type of epistasis) shows that our results are robust. To further strengthen this point, we constructed two new independent clones of rpoB R529H(R702), this time using yet a different method: by simultaneous conjugation and transduction. The fitnesses of these clones were again not significantly different from before (Kruskal-Wallis, p>0.05).
Focusing on the resistance mutations, we notice that the mean epistatic value significantly varies among them (Kruskal-Wallis p = 0.0016). Figure 2A shows that mutations rpoB R529H, rpsL K88E and rpsL K43N exhibit positive and large ε values. The other mutations show lower mean ε values (Mann-Whitney U-Test, p = 0.000002).
Figure 2B shows that there is a significant correlation between the fitness cost created by a mutation and its mean epistatic value (deviation from zero in absolute value) (Spearman p = 0.006). In other words, mutations with a more deleterious effect on the cell tend to be more epistatic. This relationship had been initially proposed after in silico studies of digital organisms and theoretical modeling of RNA secondary structures [39]. Our results are in accordance with previous experimental data from studies of epistasis amongst antibiotic resistance alleles in E. coli [27], and from a study of enzymes involved in gene expression and protein synthesis in Pseudomonas aeruginosa [40].
Focusing on the conjugative plasmids, Figure 3A shows their mean ε values. There are no significant differences in the mean ε values between plasmids (Kruskal-Wallis p = 0.676). This means that, on average, all studied conjugative plasmids tend to interact in the same way with chromosomal mutations. Moreover, comparison between Figure 2A and Figure 3A seems to suggest that the mutation (and not the plasmid) may be the major factor determining the type and the strength of the epistatic interactions we observed. In contrast to what was observed with the effect of mutations on epistasis [27], there is no significant correlation between the fitness cost created by a plasmid and its mean epistatic value (Figure 3B – Spearman p = 0.188). This may simply be due to lack of power, as variation in plasmids cost is much smaller than for mutations.
A specific mutation can be deleterious on a particular genetic background and beneficial on others – a phenomenon known as sign epistasis. Strikingly, we report that 40% of the combinations between resistance chromosomal mutations and conjugative plasmids present sign epistasis (Figure 1A, where “++” indicates sign epistasis). These are cases where the strain carrying both resistant determinants was fitter than the strain carrying only the mutation or only the plasmid (Figure 4). One of the genotypes (D87G(R16)) presents reciprocal sign epistasis [41], meaning that the mutant D87G and harboring the R16 plasmid is fitter than both the plasmid-free mutant and the plasmid-bearing strain without the mutation (hence sensitive to nalidixic acid).
We found examples of sign epistasis in all resistance alleles and all plasmids of this study, i.e. there was no plasmid nor mutation where we did not find, at least one case of sign epistasis. This high prevalence of positive epistasis is not a consequence of plasmid transfer to the reference strain. The proportion of transconjugants was monitored at stationary phase, when bacterial density is higher than 109 cells per ml, and was found to be less than 3% (Table S3). Also, computer simulations show that these plasmid transfer events imply an error in the calculation of epistasis that is less than 1%, hence less than the experimental error.
Another interesting aspect of our data is that, three out of the 50 combinations (plasmid+mutation) presented fitness costs not significantly different from zero (t-test, p>0.05); these strains are the following: rpsL K43R(R16), rpsL K43R(R831) and gyrA D87G(R16).
Finally, we measured epistasis between conjugative plasmids. This is relevant because there have been several reports of bacterial pathogens harboring multiple resistance plasmids [42], [43]. We constructed nine out of the 10 possible pairwise combinations of the five plasmids (plasmids R702 and RP4 belong to the same incompatibility group, thereby preventing the construction of this double transconjugant). In this context, epistasis was estimated as: .
In this mathematical expression, and are the fitnesses of single-plasmid carrying strains relative to the wild-type plasmid-free strain, and is the fitness of the strain carrying both plasmids, relative to the same wild-type plasmid-free strain.
Two different plasmids inside the same bacterial cell can interact either antagonistically or synergistically (Figure 5A). Epistatic interaction was found in 7 out of 9 (78%) strains. Positive epistasis (antagonistic interaction) is nearly as frequent (4/9) as negative epistasis (3/9). One out of nine pairwise combinations of plasmids presented sign epistasis (“++” in Figure 5A). Figure 5B shows the distribution of ε values for all pairwise combinations of plasmids. On average, plasmid pairwise epistasis is close to 0 (median = −0.000830, bootstrap 95% CI [−0.034690; 0.061360]). It is interesting to note that one of the plasmids, R16, interacts antagonistically with all other plasmids, and also with most mutations. This plasmid is also the most costly (Table S1).
Our results show that 52% (26/50) of the combinations between antibiotic resistance mutations and resistance conjugative plasmids interact antagonistically. This is a remarkable result because the fitness cost of these strains that carry both resistance determinants is lower than the independent sum of the cost of each determinant. Moreover, 20 out of these 26 antagonistic interactions (77%) exhibit sign epistasis or reciprocal sign epistasis, also an outstanding finding because it means that the fitness cost of harboring both resistance determinants is lower than the fitness cost of bearing one of them. In other words, an initially deleterious antibiotic resistance mutation can become beneficial through the acquisition of a transferable antibiotic resistant plasmid (16 cases); likewise, an initially costly antibiotic resistant plasmid may become beneficial through the acquisition of a mutation conferring resistance to an additional antibiotic (5 cases). This adds up to 20 cases of sign epistasis because of one instance of reciprocal sign epistasis. Last, but not least, three of the plasmid+mutation combinations presented fitnesses not significantly different from the fitness of the wild-type strain.
Positive epistasis has been shown to occur between resistance alleles in multidrug resistant E. coli. [27], P. aeruginosa [44] and Streptococcus pneumoniae [45]. Such phenomena reduce the fitness cost associated with multidrug resistance and may drive its spread. Our study aimed to detect the putative occurrence of epistatic interactions involving conjugative resistance plasmids. Such knowledge may help predict how a bacterial population will evolve after the introduction of plasmid-borne resistance determinants through horizontal gene transfer.
Our data strikingly suggests the pervasive occurrence of sign epistasis in the interaction between chromosomal antibiotic resistance mutations and conjugative plasmids. Sign epistasis has been shown to have the power to constrain protein adaptation by limiting the number of possible mutational paths and is therefore relevant to the understanding of multidrug resistance emergence [46]. Moreover, bacterial adaptation to the cost of mutation-determined resistance involves the acquisition of second-site mutations that compensate the fitness cost of the original mutation [10]. Thus, compensatory mutations are an example of sign epistasis [3], [10], [11], [46]. Our finding of pervasive sign epistasis with conjugative plasmids is one of the worst possible scenarios for the current efforts to eradicate resistance through antibiotic bans. Sign epistasis allows strains carrying a resistance mutation and a plasmid to exhibit higher fitness, thus being able of outcompeting strains carrying only the mutation or the plasmid (depending on the specific case). These results pinpoint the need for future studies involving other plasmids and other resistances.
Also important in the context of antibiotic resistance is our finding of the ubiquitous occurrence of positive epistasis between resistance plasmids and chromosomal resistance mutations. If such antagonistic interaction is a common phenomenon, then multidrug resistance determined by the simultaneous presence of plasmid-borne and chromosomal determinants will not create such a high fitness cost as one could predict based on the individual costs. Hence, such multiresistant strains may be able to persist at significant frequencies in populations where the antibiotic selective pressure has been removed.
Our findings are in accordance with the results of a large-scale survey for genes of the E. coli chromosome that are affected by the presence of the conjugative F-plasmid [47]. Such study found 107 genes exhibiting epistatic effects with the F-plasmid. Although such effect was not found for gyrA, rpoB and rpsL, other host genes involved in information transfer were reported to be affected by the presence of the F-plasmid [47]. Under the framework of the complexity hypothesis, these interactions between plasmids and informational genes (rpoB [48], [49], [50] and rpsL [51]) and a topoisomerase (gyrA [52]) are expected, given their pleiotropic interactions with other genes. For example, Schmitt et al. have shown that certain rpoB, rpsL and gyrA alleles affect F-exclusion of bacteriophage T7 [53]. In addition, Ozawa et al. [54] showed that rpoB mutations interact with a plasmidic gene (in Enterococcus faecalis). Similarly, gyrB may also interact with plasmids, eventually leading to their elimination from cells [55]. In conclusion, resistance genes present on plasmids are not necessarily responsible for the epistatic interactions observed.
We also report here the occurrence of significant epistasis between two types of conjugative plasmids within the same host. This finding has relevance for clinical isolates exhibiting multidrug resistance afforded by the co-existence of several plasmids, a situation which appears to be relatively common [43]. Our data indicates that, on average, epistasis between the conjugative plasmids is close to zero. However, we do not believe that our results suggest a tendency for no epistatic interactions between conjugative plasmids. In fact, our near-zero median level of epistasis between conjugative plasmids is the consequence of having a similar frequency of somewhat strong positive and negative epistatic interaction pairs. Our results may indicate that plasmid interaction follows an all-or-nothing type of response where the net epistatic effect is either strongly negative or strongly positive. However, further studies should use a larger sample of plasmids. Recently, in silico studies of E. coli and S. cerevisiae metabolic networks have suggested that genes involved in essential reactions tend to interact antagonistically, while negative epistasis was mainly limited to non-essential gene pairs [56]. The accessory nature of plasmids versus the essential role of gyrA, rpoB and rpsL in information flow may explain why positive epistasis appears to be more frequent in the interaction between chromosomal mutations and a plasmid than between two types of plasmids.
Our finding of pervasive positive epistasis and, in particular, of sign epistasis, between mutations and conjugative plasmids raises serious concerns to the reversal of antimicrobial-drug resistance. Plasmid-borne multidrug resistance is widespread in microbial clinical, animal and environmental isolates. Dissemination is facilitated by the conjugative plasmids' ability to mobilize their own transfer (and of other plasmids) from the original host to a new cell. Many plasmids are even able to move between phylogenetically distant organisms. Furthermore, it is known that plasmids act as recruiting platforms for resistance genetic determinants, many of them able to transpose between the plasmid and the host chromosome (and vice-versa). Thus, and given the widespread nature of horizontal gene transfer in prokaryotes it has been suggested that microbes share a common gene pool [57]. Therefore, we predict that plasmid-borne resistance dissemination control through antibiotic bans is not likely to be successful. We suggest that resistance reversal policies must target plasmids vulnerabilities. Three approaches have been suggested [58]: inhibition of plasmid conjugation, inhibition of plasmid replication, and exploitation of plasmid-encoded toxin-antitoxin systems.
We used five natural conjugative plasmids, R124, R702, R16, R831, and RP4, kindly provided by the Institute for Health, Environment and Safety of the Belgian Nuclear Research Centre. Plasmid characteristics are listed in Table S1. We introduced these plasmids in wild-type E. coli K12 MG1655 and in a set of 10 spontaneous antibiotic-resistant clones derived from the wild-type strain (Table S2). These mutations have been previously mapped to gyrA, rpoB and rpsL resulting in resistance to nalidixic acid, rifampicin and streptomycin (ref. 27).
For the construction of bacterial strains with conjugative plasmids, donors and recipients (either wild-type E. coli K12 MG1655 or strains shown in Table S2) were put together for 24 hours. All donor strains are auxotrophic for specific amino-acids and/or unable to use maltose, due to deletions in essential genes/operons, as indicated in chromosomal markers: Mal−: maltose; Trp−: tryptophan; Met−: methionine and Pro−: proline. Selection of transconjugants was performed in M9 minimal medium (56.4 g/L M9 minimal salts, 2 mM magnesium sulfate, 4 g/l sugar (see bellow), 15 g/l agar), supplemented with the appropriate antibiotics. If donors are auxotrophic (Table S4) for two amino-acids (E. coli CM140 and E. coli CM597), transconjugants were selected on minimal medium plates containing glucose and no amino-acids. Otherwise, we used maltose and tryptophan (E. coli CM317, E. coli CM319, E. coli CM312). As a control we confirmed that neither donors (due to auxotrophies or inability to use maltose as carbon source) nor recipient (due to antibiotics selecting for plasmidic resistance genes) grow on these plates.
Transduction was done with P1 bacteriophage, according to the methods described by Trindade et al. [27].
In competition assays, we used E. coli K12 MG1655 Δara as “reference strain”. Due to a deletion in the arabinose operon this strain produces red colonies when grown in tetrazolium arabinose (TA) indicator agar, allowing it to be distinguished from its competitor, which produces white colonies. TA medium contains 1% peptone, 0.1% yeast extract, 0.5% sodium chloride, 1.5% agar, 1% arabinose and 0.005% tetrazolium chloride.
All bacterial strains were grown in liquid Luria-Bertani (LB) medium at 37°C with agitation. Solid media was obtained by the addition of agar (15 g/l). For growth and transconjugant selection, antibiotics were added as follows: 40 µg/ml of nalidixic acid, 100 µg/ml of rifampicin, 100 µg/ml of streptomycin, 20 µg/ml of tetracycline, 100 µg/ml of kanamycin and 100 µg/ml of ampicillin.
Dilutions of cultures were done in MgSO4 0.01 M. All strains were kept frozen in 15% glycerol stocks.
Competition assays were performed to determine the fitness cost of the resistance determinants, either the plasmid carriage alone, the coexistence of both plasmid and mutation or the carriage of two plasmids. The method used has been previously described by ref. [27]. The strains carrying resistance determinants were competed against a susceptible reference strain, E. coli K12 MG1655 Δara, in an approximate proportion of 1∶1 and in the absence of antibiotic selective pressure. (i) Both strains were grown in 10 ml of liquid LB medium for 24 hours at 37°C with aeration. (ii) 50 µl of the dilution 10−4 of each strain was added to 50 ml screw-cap tubes containing 10 ml of liquid LB medium. (iii) Values of both strain's initial ratio were estimated by plating a dilution of the mixture in TA agar medium. (iv) Competitions proceeded by a period of 24 hours at 37°C with aeration. (v) At the end of the competition, appropriate dilutions were plated onto TA agar plates to obtain the final ratios of both competitors. These competitions spanned about 19 to 22 bacterial generations. If a high fitness cost precluded the resistant strain of being recovered in the TA plates, a smaller dilution was plated onto minimal medium supplemented with arabinose, which does not allow the growth of the reference strain. The fitness cost of each strain – i.e. the selection coefficient, s, – was estimated as the per generation difference in Malthusian parameters between the mutant and the wild-type (rm and rw respectively): specifically, , where T is the final time and g is the total number of generations from t = 0 until t = T. Then, we discounted the cost of the Δara marker [59]. The fitness cost was estimated as an average of three independent competition assays.
As explained in the main text, epistasis between a mutation and plasmid can be calculated as, where is the relative fitness of the strain with the wild-type allele (A) and carrying the plasmid, is the relative fitness of the mutant strain (with A allele replaced by the a allele), and is the relative fitness of the strain containing both the mutation (a allele) and the plasmid. Similarly, we defined epistasis between plasmids as , where and are the fitnesses of single-plasmid strains relative to the wild-type plasmid-free strain, and is the fitness of the strain carrying two types of plasmids, relative to the same wild-type plasmid-free strain. Then, the error (σε) of the value of ε is estimated by the method of error propagation;
for pairwise combinations of mutation and plasmid:
for pairwise combinations of plasmids:
If the value of ε was within the calculated error, we considered that the two resistance determinants (mutation and plasmid or plasmid and plasmid) did not show significant epistasis (indicated as white boxes labeled “no epistasis” in Figure 1A and Figure 5A). From the distribution of values of ε, provided in Figure 1B and Figure 5B, we calculated the median value of ε and its 95% CI by bootstrap where we took 10 000 samples.
To test the presence of sign epistasis, we compared the fitness of each strain carrying two resistance determinants (mutation and plasmid or plasmid and plasmid) and its corresponding single resistance-determinant strains. We used a Student t-test to assess if the fitness of the double-resistance-determinants strain was higher than the fitness of any of the single resistance-determinant strains.
Statistical analyses performed using software Statistica 9.0 and MatLab R2009b. Computer simulations performed with Mathematica 7.
|
10.1371/journal.pntd.0000729 | Cyclosporin A Treatment of Leishmania donovani Reveals Stage-Specific Functions of Cyclophilins in Parasite Proliferation and Viability | Cyclosporin A (CsA) has important anti-microbial activity against parasites of the genus Leishmania, suggesting CsA-binding cyclophilins (CyPs) as potential drug targets. However, no information is available on the genetic diversity of this important protein family, and the mechanisms underlying the cytotoxic effects of CsA on intracellular amastigotes are only poorly understood. Here, we performed a first genome-wide analysis of Leishmania CyPs and investigated the effects of CsA on host-free L. donovani amastigotes in order to elucidate the relevance of these parasite proteins for drug development.
Multiple sequence alignment and cluster analysis identified 17 Leishmania CyPs with significant sequence differences to human CyPs, but with highly conserved functional residues implicated in PPIase function and CsA binding. CsA treatment of promastigotes resulted in a dose-dependent inhibition of cell growth with an IC50 between 15 and 20 µM as demonstrated by proliferation assay and cell cycle analysis. Scanning electron microscopy revealed striking morphological changes in CsA treated promastigotes reminiscent to developing amastigotes, suggesting a role for parasite CyPs in Leishmania differentiation. In contrast to promastigotes, CsA was highly toxic to amastigotes with an IC50 between 5 and 10 µM, revealing for the first time a direct lethal effect of CsA on the pathogenic mammalian stage linked to parasite thermotolerance, independent from host CyPs. Structural modeling, enrichment of CsA-binding proteins from parasite extracts by FPLC, and PPIase activity assays revealed direct interaction of the inhibitor with LmaCyP40, a bifunctional cyclophilin with potential co-chaperone function.
The evolutionary expansion of the Leishmania CyP protein family and the toxicity of CsA on host-free amastigotes suggest important roles of PPIases in parasite biology and implicate Leishmania CyPs in key processes relevant for parasite proliferation and viability. The requirement of Leishmania CyP functions for intracellular parasite survival and their substantial divergence form host CyPs defines these proteins as prime drug targets.
| Visceral leishmanisasis, also known as Kala Azar, is caused by the protozoan parasite Leishmania donovani. The L. donovani infectious cycle comprises two developmental stages, a motile promastigote stage that proliferates inside the digestive tract of the phlebotomine insect host, and a non-motile amastigote stage that differentiates inside the macrophages of mammalian hosts. Intracellular parasite survival in mouse and macrophage infection assays has been shown to be strongly compromised in the presence of the inhibitor cyclosporin A (CsA), which binds to members of the cyclophilin (CyP) protein family. It has been suggested that the toxic effects of CsA on amastigotes occurs indirectly via host cyclophilins, which may be required for intracellular parasite development and growth. Using a host-free L. donovani culture system we revealed for the first time a direct and stage-specific effect of CsA on promastigote growth and amastigote viability. We provided evidence that parasite killing occurs through a heat sensitivity mechanism likely due to direct inhibition of the co-chaperone cyclophilin 40. Our data allow important new insights into the function of the Leishmania CyP protein family in differentiation, growth, and intracellular survival, and define this class of molecules as important drug targets.
| The cyclophilin (CyP) protein family consists of highly conserved proteins that share a common signature region of approximately 109 amino acids, the cyclophilin-like domain (CLD, Prosite access number: PS50072). The CLD carries the peptidylprolyl isomerase (PPIase) activity characteristic of CyPs [1], which has been implicated in protein folding, assembly of multi-protein complexes, and signal transduction [2]–[4]. CyPs are characterized by the binding of the cyclic peptide inhibitor cyclosporin A (CsA), which inhibits the protein phosphatase calcineurin and finds application for example as immune-suppressive drug in organ transplantation [5]. In addition to its inhibitory effect on T cell-mediated immunity [6]–[8], CsA displays anti-microbial activity against a variety of protozoan pathogens [9]–[11], including Leishmania [12]–[15].
Parasites of the genus Leishmania cause important human diseases collectively termed leishmaniasis, which range from mild, self-healing cutaneous lesions generated by L. major to fatal visceral infection of liver and spleen caused by L. donovani [16], [17]. Leishmania is transmitted by infected sand flies, which harbor the proliferating flagellate promastigote form of the parasite. Highly infectious metacyclic promastigotes are inoculated into the mammalian host during sand fly blood feeding, where they are engulfed by phagocytes of the endo-reticular system and develop inside the phagolysosome into amastigotes, which subvert the host immune response and cause the immunopathologies characteristic of the various forms of leishmaniasis [18], [19].
CsA has been shown to exert a leishmanicidal effect on intracellular L. tropica [12] and L. major in mouse and macrophage infection [13]–[15]. Although these findings define members of the Leishmania CyP protein family as potential important drug targets, only little is known on this protein family in trypanosomatids and the mechanisms of the anti-parasitic effects of CsA on intracellular Leishmania remain elusive. A potential role of Leishmania CyPs in amastigote differentiation and virulence can be postulated based on the role of Leishmania donovani LdCyP in disaggregation of adenosine kinase aggregates [20], an important enzyme in the Leishmania purine salvage pathway, whose activity substantially increases during the pro- to amastigote differentiation [21]. Furthermore, the amastigote-specific phosphorylation of cyclophilin 40 [22], [23] suggests that activity, localization, and interaction of this protein may be regulated in a stage-specific manner by post-translational modification.
The use of CsA for anti-leishmanial chemotherapy is limited by its suppressive action on host immunity, which leads to aggravation of experimental visceral leishmaniasis [24]. A better understanding on the biology of Leishmania CyPs, and their structural and functional differences to human CyPs is required to pave the way for the identification of new inhibitors with increased specificity for parasite CyPs. Here we initiated a first genome-wide analysis of the Leishmania CyP protein family and used the L. donovani axenic culture system [25], [26] to investigate the effects of CsA on promastigotes and amastigotes in culture. Our data indicate substantial evolutionary divergence between parasite and host CyPs, which may be exploitable for drug development. We provide evidence for stage-specific functions of Leishmania CyPs in regulation of promastigote cell shape and proliferation, and amastigote thermotolerance. We demonstrate for the first time a stage-specific and direct toxic effect of CsA on host-free amastigotes, validating Leishmania CyPs as drug targets.
Leishmania donovani strain 1S2D (MHOM/SD/62/1S-CL2D) clone LdB [27] was maintained at 26°C, pH 7.4 in M199 medium supplemented with 10% FCS, 20 mM HEPES pH 6.9, 12 mM NaHCO3, 2 mM glutamine, 1× RPMI 1640 vitamin mix, 10 µg/ml folic acid, 100 µM adenine, 30 µM hemin, 8 µM biopterin, 100 U/ml penicillin and 100 µg/ml streptomycin. Axenic amastigotes were differentiated at 37°C with 5% CO2, in RPMI 1640 medium pH 5.5 supplemented with 20% FCS, 2 mM glutamine, 28 mM MES, 1× RPMI 1640 vitamin mix, 10 µg/ml folic acid, 100 µM adenine, 1× RPMI 1640 amino acid mix, 100 U/ml of penicillin and 100 µg/ml of streptomycin.
Both cyclosporin A (CsA) isolated from Tolypocladium inflatum (Calbiochem) and FK506 isolated from Streptomyces tsukubaensis (A.G. Scientific) were dissolved in absolute ethanol at a final concentration of 10 mM and the stock was stored at −20°C. Log-phase promastiogtes (2×106/ml) were cultured in medium containing either solvent, CsA or FK506 at concentrations ranging from 5 to 25 µM and incubated at 26°C, pH 7.4 for 48 hours unless otherwise specified. Axenic amastigotes were differentiated at 37°C for 72 hours and were incubated at a density of 2×106 parasites/ml at 37°C with 5% CO2, pH 5.5 for 48 hours in medium containing either solvent, CsA or FK506 unless otherwise specified.
The growth of solvent treated cells controls and CsA treated parasites was determined using a CASY cell counter (Schärfe System) or determined microscopically by cell counting using 2% glutaraldeyhde fixed cells. Cell proliferation was determined by CellTiter-Blue assay (Promega) according to the manufacturer's instructions. Briefly, 20 µl of CellTiter-Blue was added to the cells in 96-well plate and incubated at 37°C for 4 hours. Fluorescence was measured (exλ = 560 nm; emλ = 590 nm) using a spectrometer SP-2000 (Safas). Results were expressed in % of fluorescence intensity compared to solvent treated cells control. The tests were performed in quadruplicate.
The sequences of human and Leishmania cyclophilins were retrieved using the UniProt (www.uniprot.org) and GeneDB (www.genedb.org) databases, respectively, and conserved protein domains were identified by ScanProsite (www.expasy.ch/tools/scanprosite). In order to determine the level of conservation of CLD domains across human and trypanosomatid parasites, all putative CLD containing proteins of the sequenced genomes of L. major, L. infantum, L. barsiliensis, T. brucei, and T. cruzi [28]–[30] were retrieved from the TriTrypDB database (http://tritrypdb.org/tritrypdb/) using HUMAN_PPIA as an initial query for PSI-BLAST. After three cycles, all hits with a significant E-value (<10E-5) and more than 70% coverage of the CLD domain were selected, and their putative CLD domain was then extracted using the alignment to HUMAN_PPIA as a guide. Given the high level of conservation of the CLD domains, it is realistic to consider this dataset as a complete set of the CLD proteins present in the species covered by the current release of TriTrypDB (Release 1.1). These sequences were aligned with T-Coffee (default mode) [31], and a Neighbor-Joining tree was computed with 500 bootstrap replicates. Positions in contact between CsA and cyclophilin A were identified on the multiple sequence alignment and the corresponding columns were extracted. The resulting functional residues were compared and clustered for similarity using UPGMA.
We first identified Leishmania CyPs that are predicted to bind CsA using multiple sequence alignment of human and Leishmania major cyclophilins, and assessing the conservation of the residues known to be involved in cyclosporin A binding in known complexes. Based on these criteria, six Leishmania major cyclophilins shared the CsA binding residues with human PPIA or PPID (LmaCyP1, LmaCyP2, LmaCyP4, LmaCyP5, LmaCyP11 and LmaCyP40) and were selected for further analysis. The leishmanial cyclophilins were modelled with the automated mode of the Swiss-Model tool [32] using the following PDB structures as templates: 2 bit [33] for LmaCyp1; 3eov [34] for LmaCyP2, 2hqj (Arakaki and Merritt, unpublished), corresponding to LmaCyP11, for LmaCyP4 and LmaCyP5; 1ihg [35] for LmaCyP40. For each model or structure, the corresponding putative model complex with cyclosporin A was built based on the complex of L. donovani cyclophilin with CsA (3eov) as a template using the program insightII. Each model complex was subjected to a very limited energy refinement (100 cycles with the insightII Discover Module, steepest descent algorithm).The 3eov CsA binding residues (R78, I80, F83, M84, Q86, G95, T96, A123, N124, A125, G126, Q133, F135, W143, L144, H148) at less than 4 Å from CsA, were used for the superposition. The subsequent analyses of the 3D model complexes and evaluations of the putative interaction with the CsA were performed with the program insightII.
Cell death was assessed by propidium iodide exclusion assay [36]. Briefly, 107 promastigotes or axenic amastigotes from control or CsA treated cultures were washed and resuspended in PBS containing 2 µg/ml of propidium iodide and incubated at room temperature for 15 min in the dark. The stained cells were subjected to FACS analysis (exλ = 488 nm; emλ = 617 nm). 10,000 events were analyzed. For cell cycle analysis, 107 late-log phase promastigotes were washed once with cold PBS and resuspended in pre-chilled 90% methanol in PBS and kept at −20°C overnight. The fixed cells were washed once with cold PBS and then resuspended in propidium iodide staining solution (10 µg/ml PI, 100 µg/ml RNase A in PBS) and incubated at 37°C for 30 min in the dark. The stained cells were subjected to FACS analysis as described above. Cell cycle distribution was calculated by FlowJo (Tree Star, Inc.) using the Dean-Jett-Fox model.
For Giemsa staining, 107 promastigotes or axenic amastigotes were immobilized on poly-L-lysine coated cover slips, fixed with methanol and stained with Giemsa reagent (Sigma) according to the manufacturer's instructions. The stained cells were mounted with Mowiol 4-88 (Sigma) [37] and observed with a Zeiss Axioplan 2 wide field light microscope.
Cells were prepared for scanning electron microscopy as described [38]. Briefly, parasites were washed in PBS, fixed with 2.5% glutaraldehyde in PBS, and treated with 1% OsO4. The cells were then dehydrated and critical-point dried (Emitech K850 or Balzers Union CPD30) and coated with gold (Joel JFC-1200 or Gatan Ion Beam Coater 681). Samples were visualized with scanning microscope Joel JM6700 F.
Indirect immunofluorescence staining was performed with 107 promastigotes or axenic amastigotes that were settled on poly-L-lysine coated coverslips and fixed in methanol at -20°C for 5 min. The fixed cells were rehydrated with PBS, and sequentially incubated with a mouse anti-α-tubulin antibody (Sigma, 1∶2500 dilution) and an anti-mouse IgG-rhodamine antibody (Molecular Probes, 1∶250 dilution). Nuclei and kinetoplasts were stained with DAPI and the slides were mounted with Prolong (Molecular Probes).
Modified CsA with a primary amine side chain was provided by the Texas A&M Natural Products LINCHPIN Laboratory, Assistant Director Dr. Jing Li [39]. The CsA-amine was coupled to the Affi-Gel®10 resin (Bio-Rad) by reaction with the N-hydroxysuccinimide ester groups of the resin. Briefly, 7.5 µmol of CsA-amine was mixed with 500 µl of Affi-Gel® 10 and incubated at room temperature for 6 hours. The coupling reaction was quenched by removing the CsA-amine and blocking the unreacted Affi-Gel® 10 with 0.2 M ethanolamine. Logarithmic promastigotes were lysed with lysis buffer (50 mM HEPES, 100 mM NaCl, 10% glycerol, 0.5% NP-40 and 1 mM PMSF) followed by sonication on ice (30 s sonication with 15 s pause for 5 min). Insoluble debris was removed by centrifugation. The cleared cell lysate (1 mg protein/ml) was mixed with the CsA-Affi-Gel and incubated at 4°C for 3 hours. Bound proteins were eluted using hot Laemmli buffer. The elution was subjected to 10% SDS-PAGE, stained with SyproRuby®protein gel stain (Invitrogen), and CsA-binding proteins were identified by MS analysis as described [22] and Western blotting.
Leishmania major CyP40 was amplified from L. major Friedlin V1 (MHOM/JL/80/Friedlin) genomic DNA using the primers 5′-CTCGAGGGAGGAATGCCGAACACATACTGC-3′ (XhoI site and 2 glycine residues are underlined) and 5′-GCGGCCGCAACCCTCACGAGAACATC-3′ (NotI site is underlined) and ligated to pGEM-T (Promega). The insert was then released by XhoI and NotI and ligated into pGEM-HAstrep. The intermediate construct was digested with BamHI and NotI to release the strep::CyP40 and ligated into pGEX-5X-1 (Amersham Biosciences). The resulting plasmid pGEX-5X-Strep::CYP40 was replicated in E.coli BL21. Recombinant GST::strep::CyP40 was induced with 0.2 mM IPTG overnight at room temperature and then purified with GSH-sepharose and strep-tactin sepharose (Fig. S1) using an Äkta Purifier FPLC system (Amersham Biosciences).
Measurements were performed according to [40]. Briefly, the peptidyl prolyl cis/trans isomerization reaction was initiated by diluting the peptide Abz-Ala-Ala-Pro-Phe-pNA in an anhydrous 0.5 M LiCl/TFE mixture with 35 mM HEPES pH 7.8. Inhibition of PPIase activity was measured by pre-incubating CsA with the enzyme (29.5 nM) for 5 min at 10°C before starting the reaction by the addition of the substrate. Data analysis was performed by single exponential non-linear regression using Sigma Plot Scientific Graphing System.
Parasites (108 cells) were lysed in 1× Laemmli buffer (1×109 cells/ml) and vortexed vigorously for 30 seconds. The lysates were sonicated for 1 min on ice and boiled for 5 min. Soluble fractions were collected as protein extracts after brief centrifugation. Twenty microliters of samples (equivalent to 2×107 cells) were separated by 10% SDS-PAGE and then transferred to PVDF membrane. Mouse anti-LPG antibody (clone CA74E, 1∶5000 dilution) [41], mouse anti-A2 antibody (clone C9, 1∶200 dilution) [42], rabbit polyclonal anti-CyP40 (established using recombinant strep::CyP40 as antigen, 1∶5000 dilution, Eurogentec), and appropriate HRP-conjugated secondary antibodies were used to probe the membrane using dilutions of 1∶10000 and 1∶50000, respectively, and signals were revealed by SuperSignal ECL from ThermoFisher.
PPIases are classified according to the binding of the inhibitors cyclosporin A (CsA) and FK506 in two major protein families, cyclophilins (CyPs) and FK506 binding proteins (FKBPs), respectively [43]-[45]. A third PPIase family is represented by PpiC/parvulin-like proteins implicated in proline-directed phosphorylation [46], [47]. Based on the presence of a conserved CyP-type PPIase signature sequence, termed cyclophilin-like domain, CLD, (Prosite accession number: PS50072, FY-xx-STCNLVA-x-FV-H-RH-LIVMNS-LIVM-xx-F-LIVM-x-Q-AGFT), the Leishmania major genome encodes for 17 cyclophilin-like proteins (LmaCyPs), five FKBP-like LmaFKBPs, and two PpiC/parvulin-like LmaPPICs (Fig. 1 and Table 1), all of which are conserved in the L. infantum and L. braziliensis genomes (Fig. 2). According to the current nomenclature [2], the LmaCyPs were distinguished by numbering from the smallest to the highest predicted molecular weight (Table 1).
Based on length and domain structure, three types of L. major cyclophilins (LmaCyPs) can be distinguished. A first group of four proteins (LmaCyP1–3, 6) is characterized by a single CsA-binding domain without any significant N- or C-terminal sequence extensions (Fig. 1 and Table 1). A second group of 11 proteins shows significant (50 or more amino acids) N-terminal (LmaCyP4, 5, 8, 10, 12, 16), C-terminal (LmaCyP7, 11), or both N- and C-terminal extensions (LmaCyP9, 13–15). These extensions are unique and not conserved in human CyPs, but are mostly conserved across other trypanosomatids, including L. infantum, T. brucei and T. cruzi. Exceptions are the C-terminus of LmaCyP13 and the N-termini of LmaCyP8, 10, and 14, which are unique to Leishmania suggesting highly parasite specific functions absent in Trypanosoma. Finally, two LmaCYPs are characterized by the presence of additional functional domains, including LmaCyP5 containing a conserved prokaryotic lipid attachment domain (PLD, prosite access number PS5125), and LmaCyP40, the cyclophilin-40 homolog containing two tetratricopeptide repeat domains (TPR, prosite accession number PS50005) known to interact with HSP90 [48]–[51].
We investigated the relationship between human and trypanosomatid CyPs by multiple alignment and cluster analysis using the sequence of the conserved CLD or the functional residues implicated in PPIase function and CsA binding. The clustering tree obtained for the CLD demonstrates that all LmaCyPs have conserved homologs in L. infantum, L. braziliensis, T. brucei, and T. cruzi, which cluster together with highly significant bootstrap values (Fig. 2A). All LmaCyPs have one homologue in the other Leishmania or Trypanosoma species, with the exception of LmaCyP5, which underwent expansion in the T. brucei genome with five sequentially arranged copies of the gene. It is interesting to speculate that the expansion of the only cyclophilin family member that contains a conserved lipid binding domain may be a reflection of the T. brucei biology, with a potential role for example for the expression of abundant gpi-anchored VSG proteins [52].
Many of the nodes support the existence of CyP subclasses across the trypanosomatids with a significant bootstrap value. In contrast, the nodes that cluster these subclasses together with their human homologues have only poor bootstrap support. This observation suggests that the various classes of CLDs encountered in trypanosomatid cyclophilins are quite distinct from one subclass to another and to their human counterparts. Substantial conservation however was observed in the cluster analysis performed with the functional CyP residues implicated in PPIase function and CsA binding (Fig. 2B). For instance, eight human CyPs and five LmaCyPs are clustering together showing a complete conservation of the canonical signature sequence characteristic for the human CsA-binding protein PPIA (Fig. 2B and Table 2). This represents a significant conservation when considering that the overall CLD domain is only 64% conserved between the Leishmania and Human CyPs. These results indicate that a subset of Leishmania CyPs are likely functionally conserved and implicated in PPIase function, while other, less conserved LmaCyPs may carry different enzymatic activities.
In conclusion, our analysis reveals a large Leishmania CyP protein family suggesting an important role of PPIases in parasite biology, and identifies unique sequence elements in the LmaCyP CsA-binding domains that may be exploitable for drug development. Identification of 5 out of 17 LmaCyPs with a highly conserved CsA binding motif strongly suggests inhibitor-binding to multiple LmaCyPs with potentially important consequences on the biological functions of these proteins and Leishmania infectivity. In the following we investigate this possibility studying the effects of CsA on L. donovani promastigotes and amastigotes in culture.
CsA has been previously shown to reduce the intracellular growth of L. major amastigotes [13], [14]. To further elucidate the mechanisms underlying this inhibition, we investigated the effects of CsA treatment on cultured L. donovani promastigotes and axenic amastigotes. Log-phase parasites from both stages (2×106/ml) were cultured in medium containing either ethanol (vehicle) or CsA at concentrations ranging from 5 to 25 µM, and incubated at 26°C, pH 7.4 (promastigote) or 37°C, pH 5.5 (amastigote) for 48 hours. At the time points indicated, the cells were fixed and counted microscopically or processed for CellTiter-Blue assay to test for proliferation. CsA-treated promastigotes showed a dose-dependent, progressive reduction of growth with an IC50 at 48 hours between 15 and 20 µM and a more than 5-fold decrease in growth at the highest inhibitor concentration compared to the control (Fig. 3A and B, left panels). Growth reduction was associated with a strong inhibition of resazurin reduction indicating reduced cell proliferation or cell viability (Fig. 3B, right panel). CsA-mediated growth reduction was reversible, as parasite growth resumed normally after removal of the drug after 48 hours of treatment (data not shown). Likewise, CsA had a striking effect on the growth of L. donovani axenic amastigotes. The parasites showed substantially higher susceptibility to CsA at this stage with an IC50 between 5 and 10 µM (Fig. 3A, right panel, and Fig. 3B, left panel), and strongly reduced resazurin reduction (Fig. 3B, right panel). Together, our data demonstrate that CsA interferes with the in vitro growth of both L. donovani promastigotes and axenic amastigotes. In the following we used FACS-based approaches to investigate the mechanisms underlying this growth defect.
To elucidate the mechanisms of CsA-mediated growth inhibition, we first investigated the effects of CsA on the viability of treated promastigotes and axenic amastigotes using a propidium iodide (PI) exclusion assay [36]. The percentages of PI positive, dead promastigotes and axenic amastigotes after 48 hours of CsA treatment was revealed by FACS analysis. Promastigotes did not show any significant increase in PI positive cells after incubation with CsA ranging from 5 to 15 µM (Fig. 4A), and more than 80% of cells were viable even at 25 µM CsA. In contrast, the proportion of PI positive axenic amastigotes increased dramatically with increasing CsA concentration, with a 4-fold decrease in cell viability at only 10 µM CsA (Fig. 4A). Thus, the decrease in cell number of CsA-treated promastigotes results from a slow-down in proliferation rather than parasite killing.
This result was further confirmed by cell cycle analysis. Late-log phase promastigotes were fixed with 90% methanol in PBS, stained with PI, and cell cycle phase distribution was determined by FACS analysis. Treatment of the parasites with 15 µM and 20 µM CsA did not affect the cell cycle distribution (Fig. 4B), suggesting that inhibition of parasite proliferation results from a non-synchronous slow-down in cell cycle progression.
CsA-treatment of promastigote cultures induced a striking effect on parasite morphology. We documented these alterations by microscopic analysis using Giemsa staining and scanning electron microscopy. Treatment of promastigotes with 10 to 20 µM CsA induced morphological changes reminiscent of axenic amastigotes, including increased aggregate formation (Fig. 5A), oval cell shape (Fig. 5B), and shortened and retracted flagella (Fig. 5C).
The CsA effects on L. donovani promastigotes are reminiscent to parasites treated with the HSP90 inhibitor geldanamycin, which results in amastigote differentiation [53]. We evaluated the effect of CsA on the differentiation state by following the expression of two markers, the promastigote specific surface glycoconjugates lipophosphoglycan (LPG) [54], which is lost during amastigote differentiation, and the A2 protein, which is induced during the pro- to amastigote conversion [42], [55]. Logarithmic promastigotes were incubated with vehicle alone (0.15% ethanol) or 15 µM CsA at 26°C, pH 7.4 for 72 hours, and the expression profile was compared to axenic amastigotes by Western blotting using monoclonal anti-lipophosphoglycan antibody CA7AE [41] and anti-A2 antibody C9 [42]. Despite the amastigote-like morphology, CsA-treated promastigotes maintain expression of LPG, comparable to the level of solvent treated cells promastigotes, and do not show induction of the amastigote marker protein A2 (Fig. 5D). CsA treatment of promastigotes at pH 5.5 did not result in further differentiation as judged by morphology and expression of LPG, nor did it have an effect on parasite viability (data not shown). These results demonstrate that unlike geldanamycin, CsA induces morphological features similar to amastigotes without inducing the appropriate expression profile.
CsA exerts its inhibitory action through binding of CyPs and inactivation of the cellular phosphatase calcineurin by CsA/CyP complexes [8], [56]. In the following, we used the unrelated calcineurin inhibitor FK506 to analyze if the CsA effects on the parasite are mediated through inhibition of this phosphatase, a test that has been previously applied on Leishmania [57]. Log-phase promastigotes and axenic amastigotes (2×106/ml) were cultured for 48 hours in medium containing either ethanol (vehicle) or FK506 at concentrations ranging from 5 to 25 µM, and incubated at 26°C, pH 7.4 (promastigote) or 37°C, pH 5.5 (amastigote). FK506 treatment of promastigotes induced morphological changes similar to CsA treated parasites, and strongly reduced in vitro growth and cell proliferation in a dose-dependent manner (Fig. 6A and B, left panels, and data not shown). Like CsA, FK506 did not significantly affect promastigote cell viability at the lower drug concentrations (Fig. 6B, left panel). In contrast, FK506 treatment of axenic amastigotes did not reproduce the CsA effects. First, as judged by proliferation and viability assay, amastigotes were more resistant to FK506, with an IC50 between 15 and 20 µM, compared to ca. 7 µM for CsA (Fig. 6B, right panel). Second, unlike CsA, FK506 did not induce massive cell death in amastigotes even at the highest concentration (Fig. 6C, right panel). These data show that CsA and FK506 have different effects on L. donovani axenic amastigotes, which may be due to either stage-specific differences in inhibitor uptake or distinct intracellular cellular targets.
Based on previously published observations, Leishmania CyPs may have important amastigote-specific chaperone functions and participate in protein disaggregation [20]. We tested if CsA treatment affects thermotolerance of promastigotes and amastigotes following the number of propidium iodide stained, dead parasites as a read out. Log-phase promastigotes or amastigotes were treated with 15 µM CsA and parasites were simultaneously incubated for various time periods at either 26°C or 37°C. As expected, CsA treated amastigotes showed increased cell death in the presence of CsA during the 20 hours time course experiment (Fig. 7, right panel). Significantly, CsA-treatment of amastigotes at 26°C completely abrogated the toxic effect of the inhibitor. This data shows that CsA-mediated amastigote killing is temperature dependent. We confirmed this result using the complementary set up, incubating CsA-treated promastigotes at high temperature. Just like amastigotes, CsA-treated promastigotes underwent cell death as soon as 10 hours after temperature shift (Fig. 7, left panel). CsA alone or heat shock alone had no significant effect on promastigote viability. Thus, CsA affects thermotolerance of both the promastigote and amastigote stages.
The effect of CsA on parasite thermotolerance primed us to investigate the potential interaction between this inhibitor and LmaCyP40, a bifunctional cyclophilin that has both PPIase and co-chaperone function and interacts with members of the HSP protein family through TPR domains [58]. We first used a structural approach applied on six leishmanial cyclophilins selected for their similarity to the cyclosporin A binding pocket of human orthologs. We built the corresponding model complexes with CsA and evaluated their geometric fit and ability to establish inter-molecular hydrogen bonds with the ligand. The experimentally identified CsA binding residues of the L. donovani cyclophilin (3eov) and the putative binding residues of the L. major 3D model complexes, including the one for LmaCyP40, are highly conserved (Fig. 8A). All models, even if built on different templates, display a root mean square deviation of less than 0.6 angstrom on the CsA binding residues of the experimentally determined complex structure. Consequently, all models can accommodate the CsA ligand with no molecular clash and the hydrogen-bonding pattern is conserved with respect to the experimental structure (Fig. 8A, lower panel). Furthermore, manual inspection of the model complexes revealed a good geometric complementarity between the protein and the ligand. All these evidences support the hypothesis that these L. major cyclophilins, including LmaCyP40, are good candidates for CsA binding.
We confirmed binding of the CsA ligand to LmaCyP40 by studying the proposed interaction by affinity chromatography using CsA-loaded resin. L. donovani promastigote extracts were incubated with the resin and bound proteins were separated by SDS-PAGE. One major band, specifically retained on the CsA-resin, was revealed by fluorescent protein gel staining, and identified as CyP2 by MS analysis (Fig. 8B, left panel, and Dataset S1). Western blot analysis of the gel revealed cyclophilin 40 (Fig. 8B, right panel), thus confirming the CsA-CyP40 interaction suggested by the structural modelling.
We next analyzed the biochemical characteristics of the LmaCyP40-CsA interaction using GST::Strep::CyP40 purified from recombinant bacteria (Fig. S1). We first determined the kcat/Km of Leishmania major GST::Strep::CyP40 PPIase activity by evaluating the linear dependency between kenz and enzyme concentration ranging from 14.7 to 59 nM. The catalytic efficiency of Leishmania major GST::Strep::CyP40 for Abz-Ala-Ala-Pro-Phe-pNa was found to be kcat/KM = (3.725±0.16)×105 M−1 s−1 (Fig. 8C, upper panel). We then tested direct inhibition of the LmaCyP enzymatic activity by CsA using the substrate Abz-Ala-Ala-Pro-Phe-pNA and increasing amounts of inhibitor. The IC50 value of CsA was determined to be 162±46 nM CsA (Fig. 8C, lower panel) and thus similar to human CyP40 with an IC50 value of 195 nM [59].
The leishmanicidal activity of CsA has been first demonstrated in L. tropica infected BALB/c mice, which showed a dose-dependent inhibition of parasite burden and reduction in lesion formation [12]. This anti-parasitic activity has been subsequently confirmed for L. major in mouse and macrophage infection assays, and various modes of CsA action have been proposed [13], [14], [57]. The observation that CsA has no overt anti-microbial activity against L. major promastigotes in culture, but efficiently kills amastigotes in infected macrophages, provided support to the idea that the toxic effect of CsA on intracellular parasites depends on inhibition of host rather than Leishmania CyPs. This hypothesis was further supported by findings showing that the phosphatase calcineurin, the prime target of the inhibitory CsA/CyP complex, is expressed at very low levels and is not recognized by Leishmania LmaCyP19 (corresponding to LmaCyP1 according to our nomenclature), although this protein efficiently bound CsA [60], [61]. In contrast to these previous reports, our data provide several lines of evidence for a direct action of CsA on Leishmania CyPs.
A first line of evidence resulted from the bio-informatics analysis and structural modeling of Leishmania CyPs. Blast search of the L. major and L. infantum genome databases (www.genedb.org) identified a surprisingly large family of 17 CyP-like proteins in these protozoan, compared to yeast, Drosophila, and human with 8, 14 and 19 CyPs, respectively (Table 1, Fig. 2) [62]–[64]. Multiple sequence alignment of trypanosomatid and human CyPs, cluster analysis of the functional residues implicated in PPIase catalytic activity and CsA binding of the CLD, and structural modelling revealed the presence of six Leishmania CyPs that showed conservation of the functional residues (Table 2, Figs. 2 and 8A) and were predicted to form a complex with CsA. This remarkable conservation indicates that multiple Leishmania CyPs are likely binding to CsA, a fact that we subsequently confirmed by affinity chromatography and Western blotting, revealing direct interaction of the inhibitor with Leishmana CyP2 and CyP40 (Fig. 8 B).
The effects of CsA on L. donovani promastigotes and axenic amastigotes further support this possibility and provided a second line of evidence for a direct action of CsA on Leishmania CyPs in vitro. We showed that inhibitor treatment of L. donovani promastigotes leads to dose-dependent, reversible inhibition of proliferation (Figs. 3A and B), without significant effects on cell viability (Fig. 4A) and cell cycle distribution (Fig. 4B). These results confirmed previous observations that CsA does not exert a toxic effect on Leishmania promastigotes, but revealed a strong effect on promastigote in vitro growth that escaped previous analysis, likely due to the lower CsA concentration (4 µM) used in these studies [13], [14]. In contrast to promastigotes, CsA showed a direct toxic effect on L. donovani axenic amastigotes with more than 50% of parasite death in the presence of 10 µM inhibitor (Fig. 4A). This result demonstrates for the first time that the observed anti-leishmanial effect on intracellular amastigotes in mouse and macrophage infection [13], [14], [57] may rely mainly on direct inhibition of parasite CyPs by CsA, although a participation of host CyPs can not be excluded. We further investigated the mechanisms underlying the stage-specific effects of CsA using the unrelated antifungal macrolide inhibitor FK506. FK506 binds to FKBPs, a second class of PPIases (Table 1), which similar to the CsA/CyP complexes inhibit calcineurin [8]. FK506 treatment reproduced the effects observed in CsA-treated promastigotes, suggesting inhibition of calcineurin as one of the mechanisms underlying the observed growth defect of this parasite stage (Fig. 6). To our surprise, unlike CsA, FK506 did not exert a toxic effect on axenic amastigotes at concentrations between 5 and 15 µM (Fig. 6B), a fact previously observed in intracellular L. major amastigotes [57]. These data indicate that the toxic effect of CsA on amastigotes occurs likely through calcineurin-independent mechanisms, which may be directly linked to inhibition of stage-specific enzymatic functions of Leishmania CyPs.
Cyclophilins are protein chaperones with PPIase activity, which catalyzes the cis-trans isomerization of peptidyl-prolyl bonds, affecting stability, activity, and localization of client proteins [2], [65]. Thus, inhibition of CyP functions by CsA may provoke pleiotropic downstream effects that may lead to the observed growth inhibition and loss of viability. In the context of the current literature, two pathways may be singled out with potential relevance for the CsA-dependent toxicity. First, L. donovani adenosine kinase aggregates have been identified as clients for CyP2, which disaggregates complexes of this protein [20], [66], thereby playing an important function in the purine salvage pathway [67]. Inhibition of this important CyP2 chaperone function may limit the intracellular concentration of adenosine and affect DNA synthesis with consequences for promastigote growth and amastigote viability. Second, cyclophilins have been reported to participate in the response to heat stress in other microbial pathogens. In the human pathogenic fungi Cryptococcus neoformans, CsA treatment prevents growth at elevated temperatures [68], [69] and the CyP-related protein Cp1a is required for full expression of fungal virulence [70]. Our data indeed established a direct link between the sensitivity of Leishmania to CsA and the parasite thermotolerance. We demonstrated that CsA-treated amastigotes are insensitive to the drug when incubated at 26°C, while CsA-resistant promastigotes are efficiently killed by the inhibitor at 37°C (Fig. 7). A second observation linked Leishmania CyPs with the response to increased temperature. We observed a striking effect of CsA on promastigote morphology, which acquired an oval cell shape and shortened their flagella, thus showing some (but not all) features characteristic for amastigote differentiation (Fig. 5). A similar morphogenic effect has been previously observed on promastigotes treated with the HSP90 inhibitor geldanamycin [53]. It is possible that both CsA and geldanamycin target different proteins are part of the same heat shock complex implicated in Leishmania differentiation and thermotolerance, such as cyclophilin 40, a multifunctional protein that interacts with various members of the HSP family through conserved TPR domains [58]. Indeed, our data identified LmaCyP40 as a direct target for CsA as judged from the direct interaction between the enzyme and the inhibitor (Fig. 8B) and CsA-dependent inhibition of LmaCyP40 PPIase activity (Fig. 8C). It is interesting to speculate that the temperature-dependent CsA effect on Leishmania viability is the result of CyP40 inhibition. Future studies employing LmaCyP40 conditional null mutants with the aim to dissociate the PPIase and chaperone functions of this enzyme may allow testing this hypothesis and shed important new light on the function of LmaCyP40 in parasite thermotolerance and infectivity.
In conclusion, our data revealed for the first time a direct cytostatic and cytotoxic effect of CsA on L. donovani in culture. We provided evidence that the stage-specific effects of CsA are governed by independent mechanisms linked to inhibition of calcineurin phosphatase activity in promastigotes, and inhibition of CyP functions relevant for thermotolerance in amastigote. We identified unique sequence elements in Leishmania CyPs and documented a considerable evolutionary expansion of this protein family, compared to other organisms, emphasizing the importance of this class of molecules for trypanosomatid-specific biology. The requirement of Leishmania CyP functions for intracellular parasite survival and their substantial divergence from host CyPs defines these proteins as prime drug targets. The suppressive action of CsA on host immunity and its exacerbating effects on murine toxoplasmosis, trypanosomiasis, and visceral leishmaniasis [24], [71], [72] obviously eliminates this drug for anti-parasitic intervention. Hence, the focus of future research should lie on the identification of novel CyP inhibitors that specifically target parasite CyPs without altering the host immune status.
|
10.1371/journal.pcbi.1006423 | The physiological variability of channel density in hippocampal CA1 pyramidal cells and interneurons explored using a unified data-driven modeling workflow | Every neuron is part of a network, exerting its function by transforming multiple spatiotemporal synaptic input patterns into a single spiking output. This function is specified by the particular shape and passive electrical properties of the neuronal membrane, and the composition and spatial distribution of ion channels across its processes. For a variety of physiological or pathological reasons, the intrinsic input/output function may change during a neuron’s lifetime. This process results in high variability in the peak specific conductance of ion channels in individual neurons. The mechanisms responsible for this variability are not well understood, although there are clear indications from experiments and modeling that degeneracy and correlation among multiple channels may be involved. Here, we studied this issue in biophysical models of hippocampal CA1 pyramidal neurons and interneurons. Using a unified data-driven simulation workflow and starting from a set of experimental recordings and morphological reconstructions obtained from rats, we built and analyzed several ensembles of morphologically and biophysically accurate single cell models with intrinsic electrophysiological properties consistent with experimental findings. The results suggest that the set of conductances expressed in any given hippocampal neuron may be considered as belonging to two groups: one subset is responsible for the major characteristics of the firing behavior in each population and the other is responsible for a robust degeneracy. Analysis of the model neurons suggests several experimentally testable predictions related to the combination and relative proportion of the different conductances that should be expressed on the membrane of different types of neurons for them to fulfill their role in the hippocampus circuitry.
| The peak conductance of many ion channel types measured in any given animal is highly variable across neurons, both within and between neuronal populations. The current view is that this occurs because a neuron needs to adapt its intrinsic electrophysiological properties either to maintain the same operative range in the presence of abnormal inputs or to compensate for the effects of pathological conditions. Limited experimental and modeling evidence suggests this might be implemented via the correlation and/or degeneracy in the function of multiple types of conductances. To study this mechanism in hippocampal CA1 neurons and interneurons, we systematically generated a set of morphologically and biophysically accurate models. We then analyzed the ensembles of peak conductance obtained for each model neuron. The results suggest that the set of conductances expressed in the various neuron types may be divided into two groups: one group is responsible for the major characteristics of the firing behavior in each population and the other is more involved with degeneracy. These models provide experimentally testable predictions on the combination and relative proportion of the different conductance types that should be present in hippocampal CA1 pyramidal cells and interneurons.
| Any given neuron in the brain is part of a network, in which it exerts its action by transforming the input it receives into an output. This function is specified by the particular shape and passive electrical properties of the neuronal membrane, the composition and spatial distribution of ion channels across its processes, and the functional properties of the synaptic inputs themselves. During development and during the entire lifetime of a neuron, its input/output function is adapted to realize ongoing refinement of the function of the neuron and circuit, or maintain functional robustness in the face of constant protein turnover or an evolving pathological condition. Such adaptability of individual neurons can be achieved through a myriad of dynamic mechanisms, including structural, intrinsic, and synaptic plasticity. A direct experimental evidence for these mechanisms is the high variability observed for the current generated by specific types of ion channels measured across individual neurons, from either a homogeneous population or different cell populations (e.g. [1]). The mechanisms responsible for this variability are not well understood, although there are clear experimental and modeling indications that correlation and degeneracy among a variety of conductances can be involved [2,3]. The phenomenon of degeneracy allows the possibility, for a complex biological system, to perform the same function using structurally different elements [4]. In the context considered in this paper, it refers to the robust and tunable adjustment of a neuron’s firing properties [5]. For example, a neuron can be tuned to perform a given function by expressing in the membrane a specific set of conductances with a specific dendritic distribution (Migliore (2003)); degeneracy can result in this tuning being robust, by implementing the same function with many different configurations of the same set of conductances. This property has been systematically studied in crab stomatogastric ganglion neurons [2, 6] and in Globus Pallidus neurons of the rat [7]. In the present study, we investigate this issue for neurons of the hippocampal CA1 region. These neurons are important because they have a critical position as the main output stage of the hippocampal circuitry [8]. The hippocampal CA1 pyramidal neurons, in particular, exhibit a peculiar ensemble and distribution of conductances (reviewed in [9]), subject to significant changes following activity-dependent biochemical processes, such as activation of protein kinase A and C, or Ca/calmodulin dependent kinase II [10, 11, 12], pathological conditions (e.g. [13, 14]), or traumatic brain injuries [15, 16]. There must then be an extremely robust compensatory mechanism in these neurons, or in the network, which maintains or re-establishes the physiological activity within an operation range, in spite of a potentially large change in its intrinsic properties or synaptic input. Here we study the mechanisms of robustness of intrinsic properties by using a unified data-driven workflow and open source analysis and simulation tools. From a set of experimental recordings and morphological reconstructions, we implemented many morphologically and biophysically accurate models for CA1 pyramidal neurons and interneurons, with intrinsic electrophysiological properties constrained by and consistent with the experimental findings. The results indicate that a few currents need to be expressed at a relatively stable level, whereas others can be expressed within a much wider range. The analysis of the model neurons suggests many specific experimentally testable predictions on the combination and relative proportion of the different ionic conductances, and their relationship to robustness of intrinsic properties.
To implement a set of data-driven neuron models, we start from a set of morphological reconstructions of neurons and somatic voltage traces obtained from in vitro slice preparations of rat hippocampal tissue to use as constraints (see Methods). In Fig 1 we show several examples of the 34 morphologies used in this work (19 pyramidal cells and 15 interneurons), superimposed on a rat hippocampal slice stained for parvalbumin for illustrative purposes.
A total number of 1456 experimentally obtained somatic voltage traces for a range of stimulation protocols were used in the optimization pipeline to constrain the models (see Methods). Collections of traces for individual neurons were manually assigned to four electrical types (e-type), according to the firing pattern exhibited during increasing somatic current injections [18], and using the classification proposed in the Petilla convention [19]. The 832 traces from pyramidal neurons, with an increasing inter-spike-interval (ISI), were all classified as continuous accommodating cells (cAC). For interneurons, 240 traces were classified as cAC, 160 traces as bursting accommodating cells (bAC), and 224 traces, whose firing rate is constant, as continuous non-accommodating cells (cNAC). Typical examples illustrating the physiological variability observed for these e-types are shown in Fig 2. A more quantitative analysis and comparison of their features will be presented elsewhere (Bologna et al., manuscript in preparation). Different pyramidal neurons (Fig 2, pyr cAC) exhibited significantly different responses to the same input. For example, a near-threshold 0.4 nA somatic current injection may or may not generate a few action potentials, whereas a 0.8nA input can result in a 2-fold range for the number of elicited action potentials (APs) (Fig 2, pyr cAC, blue traces). Interneurons classified as cAC also exhibited a large inter-cell variability, with different cells responding to the same stimulus with a wide range of spike patterns, such as tonic firing (Fig 2, int cAC plots, cell 970428A1), stuttering (cell 970509HP2), and depolarization block (cell 980205FHP). The other two interneuron e-types, bAC and cNAC, also exhibited a large variability among different cells (Fig 2, bottom plots). This variability can be the result of different morphologies and/or a different density and distribution of the conductances expressed on the membrane of the different neurons. In the following sections, we will explore in more detail this issue by implementing and analyzing cellular level models that are able to reproduce these results.
For each e-type (see S1–S4 Tables and Methods), a set of electrophysiological features were extracted from all voltage traces belonging to that e-type. All the pyramidal cell morphologies were used to implement cAC models, whereas interneuron morphologies were used to obtain cAC, cNAC, and bAC models following the known firing behavior of each type of morphology (see legend of Fig 1 and S5 Table). Features and morphologies were then used to obtain a set of optimized models for each e-type, using a heuristic parameter optimization process that employed multi-objective genetic algorithms. Each optimization run (see Methods for details) returned a number of viable “individuals”, each one with a specific ensemble of peak ion channel conductance and passive properties consistent with the chosen “objectives” (i.e. a set of experimental features). As a cost function for the optimization process we used a score defined by the total error associated with each individual, calculated as the sum of the absolute deviations of model features from the experimental mean, in units of the experimental standard deviation (sd) obtained for the value of each objective. A score = 0 would correspond to an individual with all parameters equal to the average value of the corresponding experimental electrophysiological feature. The total error thus gave an idea of how good the individual was in representing the neuron’s overall expected behavior under a series of 400 ms long somatic current injection steps. The final choice to accept an individual as a plausible representation of a given e-type was based on the error obtained for each objective. An individual with a sd<2 for all objectives was considered acceptable.
Typical optimization results for pyramidal and interneuron cAC e-types are shown in Fig 3. Traces obtained for different somatic current injections from three individuals (Fig 3, traces on top left graph of each panel), showed that the optimization process was able to take into account the experimental variability. Different individuals exhibited significantly different responses to the same stimulus, as in the experiments. The evolution of the total score as a function of the number of generations in the optimization process (bottom graph in each panel), showed that the optimization converged nearly monotonically in relatively few iterations, having reached a relatively stable minimum within approximately 60 generations. The list of objective scores for the best individual in each case (Fig 3 right graph in each panel) showed that for most features (n = 60 for pyramidal cells and n = 47 for cAC interneurons, see S1–S4 Tables) the associated error was below 2 sd. Similar results were obtained for the optimizations of bAC and cNAC interneurons (see individual optimization files at https://collab.humanbrainproject.eu/#/collab/18565). Taken together, these results show that the overall optimization process is a robust way to obtain a number of biophysically accurate neuron models of hippocampal CA1 pyramidal cells and interneurons, which are able to reproduce many of the properties observed experimentally in different types of neurons.
A more direct comparison between experimental and modeling traces for the different e-types is shown in Fig 4A, revealing a very good qualitative agreement between the modeling results and experimental traces. The optimization enabled the production of models that correctly reproduced many characteristics of the firing patterns, such as the strong accommodation observed in cAC interneurons (Fig 4A, cAC int @0.4nA), the high firing frequency of bAC interneurons at the beginning of a current injection (Fig 4A, bAC @0.6nA), and the progressive reduction in the AP amplitude during the first part of stronger stimuli (Fig 4A, bAC @1nA). The pyramidal cell models also exhibited a typical property often observed experimentally in this type of cells, i.e. the decrease in the peak amplitude of an AP backpropagating into the apical dendrites [20]. This effect has been shown to depend on the high density of A-type potassium channel in the apical dendrites [21], but not all CA1 pyramidal neurons exhibit this effect [22, 23]. It is important to note that this feature was not used to constrain the optimization but, interestingly, the optimized models were able to reproduce it, as shown in Fig 4B, for a few cases using morphologies from both young adult (cells 050921AM2, and 990803) and P14-23 animals. The dichotomy in AP backpropagation observed in the experiments [22] was also reproduced by the model neurons, with the AP amplitude either strongly decreasing beyond ~150 μm from the soma or limited to ~50% of the maximum, with very few cases in between. Taken together, this comparison between experiments and models at the individual trace level, suggests that the optimization process was able to correctly capture and explain both intra- and inter-cell variability in firing behavior in terms of different combinations of active and passive membrane properties.
An indication of how the optimized models may capture the variety of experimental input/output properties can be drawn from Fig 5, where the number of spikes for each e-type was plotted against the somatic current injection, for experimental (blue lines) and modeling traces (red lines). In all cases, experimental traces exhibited a rather large inter-cell variability in the number of spikes elicited by any given input current. It is quite common to see up to a ~5-fold difference in the number of spikes elicited in different cells under the same current injection. In most cases, the models were in quantitative agreement with the average number of spikes generated as a function of the input current (Fig 5, insets, Mann Whitney Rank Sum test p>0.05 in all cases except for 1nA injection in pyramidal neurons).
With the set of data-driven neuron models obtained for each e-type, we can now analyze how different combinations of peak conductances can result in models able to reproduce equally well the firing properties observed experimentally under different current injection steps. The optimization process generates many of these models (termed “individuals”) because of ion channel degeneracy [5]. As discussed in the Introduction, this phenomenon is thought to allow a neuron to adjust its firing properties in a robust and tunable manner.
To obtain further insight into on how degeneracy is achieved in hippocampal CA1 neurons, we analyzed all the individuals obtained from the optimization runs. For each optimization run, the 10 best individuals were considered based on their total score (see Methods). Note that these individuals were obtained from the same morphology with different channel densities. In Fig 6, the value of the optimized parameters, normalized to the maximum value chosen for each conductance, were plotted for each optimization run (10 individuals for each run, opt id). For clarity, in each graph the values obtained for any given parameter were placed on the Y-axis according to the corresponding average value calculated from all optimizations. In this way, the bottom rows in each graph correspond to parameters with an average low value whereas top rows correspond to parameters with higher values. Furthermore, parameters that were relatively stable across all optimizations (i.e. with a sd<0.2) for any given e-type are highlighted using a red label in the y axis. For pyramidal cells (Fig 6, pyr cAC) the most stable parameters were some of the passive properties, Ih, KM, Calcium, and Ca-dependent K currents. Interestingly, we noted that whereas passive properties were consistently optimized with a stable value across the optimizations for all e-types (Fig 6, see top rows in all graphs), conductances were shown to be somewhat different depending on e-types. For example, for interneurons, Ih, somatic KM and dendritic KDR were the most stable for all e-types, whereas dendritic KA was stable for cAC and Cagk for cNAC. These results suggested that each e-type has specific active properties that may be particularly important to obtain the appropriate firing pattern in response to a given input. While these properties need to be well constrained for each e-type, degeneracy can be achieved by combining the other conductances in a relatively large number of ways. The functional consequences of this situation will be discussed below.
To explore whether a cell’s morphology can also be related to degeneracy, we fixed the peak conductance values to those found for the best overall individual (obtained for morphology oh140521_B0_Rat_idA) and calculated the total error by using different morphologies. The results are shown in Fig 7A. We found that the total error using the same set of conductances on different morphologies was within the range obtained for each cell’s optimization for 10 out of 16 morphologies. For these cases, there was no correlation between the total error and the main morphological properties, such as soma area, total cell volume, or number of sections (Fig 7B, Spearman correlation, p>0.05 in all cases). These results suggest that degeneracy can also be obtained using different morphologies equipped with identical peak channels conductance. A deeper analysis of this issue however was not further considered in this work.
For a more detailed analysis of the configuration of peak conductance values for all models, we first considered the results for pyramidal neurons. In Fig 8A we show a typical distribution of normalized values obtained for membrane properties where optimizations yielded a relatively narrow range (somatic KM, Ih, and Ra), or a wider range of values across individuals (dendritic Na). Note that two of the conductances with a narrow distribution are, in pyramidal CA1 neurons, the dominant factors in controlling major properties such as excitability and accommodation (KM, reviewed in [24]), and synaptic integration (Ih, [25]). The paramount importance of these two types of conductance for reproducing the experimental traces, suggested by their value lying in a narrow range across individuals, emerged from the optimization process without any specific constraint.
An insight on degeneracy in these neurons can be obtained by considering correlation between parameter pairs. In most cases, we found no statistically significant correlation (see S6 Table for the Spearman correlation coefficients). However, for several cases a significant correlation between selected parameters was found (S6 Table, grey cells). The conductance which was most correlated with others was Cagk, a Ca- and voltage-dependent K+ conductance that is one of the major determinants for accommodation in these neurons. The inverse correlation with the KM is particularly interesting, since it supports the experimental finding that these channels operate in combination to control intrinsic hyperexcitability [26], and modeling results suggesting how they must both be involved to obtain a strong accommodation [27, 28].
To explore the configuration of the conductances in a more qualitative and intuitive way, we arranged a radar plot of the conductances most correlated with Cagk (Fig 8B), and one of those showing little variability (in this case the reversal potential of the leakage current in the dendrites, e_pas d). The different individuals were sorted with respect to Cagk (Fig 8B, thick black line) and, for clarity, we plotted only 40 of the 160 individuals. The highly jagged and intermixed lines represent the different peak conductance type and value for different individuals giving equally good representations of 60 electrophysiological features experimentally observed in these neurons (see S1 Table). Examples of model traces from a few individuals (all obtained with a 0.4nA somatic current injection) displayed the same number of spikes obtained with very different channel configurations. The number of spikes elicited for each individual is plotted in Fig 8C.
These results give a clear indication that degeneracy in CA1 pyramidal cells can easily emerge from many different combinations of many, but not all, channels. The reason for the lack of pairwise correlation between most parameters does not exclude that the parameter space may be shaped by higher order correlations that can be ultimately responsible for degeneracy. However, a full quantitative study of higher order correlations was outside the scope of this study.
The results obtained for interneurons are shown in Fig 9. In this case, to allow an easier comparison of the parameters among the different e-types, individuals were sorted according to the somatic Na conductance (Fig 9, thick black lines), which was among the most correlated with all the others (see S7–S9 Tables). The models suggest a few distinct differences among the different e-types. Note, for example, the distribution of values obtained for the peak conductance of dendritic KDR or KA in the various e-types (Fig 9, dark red and blue lines, respectively), or the difference in the overall values of dendritic Na (Fig 9, orange lines) between cAC and cNAC. In general, however, the distribution of values were analogous to those obtained for pyramidal cells, with each individual characterized by a highly variable combination of values for many conductances.
Finally, one important factor in determining the firing characteristics of different neurons, in addition to a substantial change in morphology [29] and/or gene expression profile [30], is the relative proportion with which specific channels are expressed on the membrane. For this reason, from the optimized models we calculated the relative contribution of each channel in each e-type, by considering the average value of each peak conductance calculated across all individuals. The results are presented in Fig 10A. In all cases, we found that Na, KA and KDR could account for most of the channels expressed on the membrane. Interestingly, each e-type showed a distinct proportion of these channels, with axonal Na channels playing a relatively large role in all e-types, axonal KA being relatively more important in pyramidal neurons than in interneurons, and dendritic KDR being significantly higher in cNAC e-types. An analysis of the relative level of each conductance in the various e-types (Fig 10B) also showed significant differences in several cases (Pairwise Multiple Comparison Procedure, p<0.05). From the results it is clear, for example, that dendritic Na should be higher in pyramidal cells than in any type of interneuron, cAC interneurons should have a higher dendritic Na among interneurons (Fig 10B, dark blue squares for Na d), and that the axonal KM is essentially independent from cell type. In summary, these results suggest the experimentally testable prediction that different e-types can be characterized by a different combination of the same set of conductances.
It has been shown that any individual neuron can express a distinct combination of many channel types [30] determining its electrical properties [31]. Furthermore, several seminal papers demonstrated that each cell type could exhibit specific correlation between channels expression [32], which may emerge from a homeostatic rule [2]. The overall picture is one in which many different conductances coincide to produce the electrophysiological patterns that characterize the operating range of any given population of neurons, and they do so in such a way to compensate for relatively large changes in individual channel density or synaptic connectivity [33]. The robustness of this mechanism relies on degeneracy [4], which can be practically implemented through a large and flat parameter space for channel conductance. This issue has been studied in the crab pyloric neurons [3], stomatogastric ganglion neurons (e.g. [2, 6]), in the Globus Pallidus neurons of the rat [7]. The presence of degeneracy had yet to be studied in hippocampal neurons. Two recent modeling studies, in the mouse corticospinal neurons and motor cortex, have explicitly shown how degeneracy in cortical neurons can work to implement some electrophysiological features but not others [34], and that degeneracy can also generate multitarget routes from pathological to physiological network dynamics [35]. The first finding was particularly relevant for our study, and it was among the reasons why we choose not to include the voltage between spikes among the optimized features. Its accurate reproduction would have required us to additionally optimize channel kinetics, which was not within the scope of this work.
The analysis of the modeling results presented in this paper provides many experimentally testable predictions on the possible co-regulation of ion currents in hippocampal CA1 neurons. Correlation between pairs of specific conductances has been found for cells in the stomatogastric ganglion of the crab (STG, [32]) and in the pyloric network of the spiny lobster [36]. These experiments found that several pairwise correlations between the same conductances can be present in different type of cells, but no cell type showed conductances with the same set of pairwise correlations. Our optimized models confirmed this result also for the hippocampal CA1 neurons. The models also confirmed pairwise correlations already observed in STG, such as that between KA and Ih, Na, KDR, and Cagk, and between Na and Cagk. Like in the STG, these correlations were observed in different combinations among different cell types. It is important to stress that the optimization process did not bias the parameter values against each other. Correlations thus emerged naturally from the optimization process, and reflected a better reproduction of the experimental features. The models predict several additional pairwise correlations between conductances (see S6–S9 Tables), which are specific for each e-type. All predictions can be tested experimentally, by directly measuring and comparing peak ion currents or (better) channel densities in different neurons or by a genetic perturbation of channel expression [36, 37].
A limitation of this work is that the optimization process was not able to generate a population of models reproducing the very large experimental variability. The reason for this effect is that, in this work, we choose to optimize the different e-types using for each feature the average and standard deviation calculated from all traces, rather than independently optimizing models constrained by traces from an individual cell. A partial explanation for this choice was the limited availability of experimental data on individual cells. Nevertheless, we think that these results offer a significant improvement on the current state of the art, and a necessary step towards building a full-scale cellular model of the rat hippocampus CA1 circuit (Romani et al., in preparation).
Another experimentally testable prediction of the models is that each type of cell should have a small number of channel types that would be expressed at the same density in the same neuronal population. There is already some experimental indication that this is the case for STG cells in the crab [1], where it has been found that KDR is relatively constant among the lateral pyloric neurons of different animals, whereas KA and Cagk varied more than threefold. In this study, we found that passive properties, KM, and Ih were among the most stable intrinsic membrane properties in any given neuron population, together with dendritic KDR for interneurons.
The models also predict that a different combination of axosomatic Na, KA, and KDR channels may dominate the distribution of channels on the membrane of a neuron belonging to a given e-type. This is also experimentally testable, by directly measuring the density of the different channels expressed on the membrane of different type of neurons.
Our analysis suggests a physiological plausible explanation for why single channel mutations can have more or less pathological consequences. A clear example stands out for KM and Ih channels in pyramidal cells. We found that these channels must be expressed with a relatively stable density; they do not appear to contribute to degeneracy. This may explain why specific mutations of KM channels can result in neonatal epilepsies in humans [38], or why the decrease in Ih caused by experimental models for temporal lobe epilepsy can result in major changes in the electrophysiological mechanisms related to cognitive functions [39].
Finally, the modeling effort presented and discussed in this work is part of a larger modeling workflow currently underway in the framework of the EU Human Brain Project (https://www.humanbrainproject.eu/en/), with the main goal to implement a cellular data-driven model of the entire hippocampus. The Hippocampus is a complex brain structure, deeply embedded into the temporal lobe, with a paramount importance for higher brain functions such as learning and memory, and spatial navigation, and is involved in several major brain diseases. In spite of intensive experimental and computational studies, the mechanisms underlying these functions (and dysfunctions) are still poorly understood. A model implementation and analysis at the cellular level may pave the way for a deeper understanding of the diverse and complex functions of this brain region, and of its levels of organization. One of the major steps towards this goal is the implementation of morphologically and biophysically accurate single cell models for the main neuronal populations, equipped with a set of axonal, somatic, and dendritic currents consistent with many experimentally measured electrophysiological features, in such a way as to be able to capture the main I/O properties observed experimentally. Here we have used a general, robust, and flexible tool able to produce, using reasonable computational resources, ensembles of this type of models for CA1 pyramidal cells and interneurons.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.